kelseye commited on
Commit
48cf674
·
verified ·
1 Parent(s): 4375e1b

Upload folder using huggingface_hub (part 23)

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. parse/dev/QDdJhACYrlX/images/9becef61539f1118df436f5341192c61d1cb7f755f868cd727b3bf797788ee40.jpg +3 -0
  2. parse/dev/QDdJhACYrlX/images/a3cb1246375e83911e1c3dd3a3d07c8fd688315b5853000c0482ebd0667a70c6.jpg +3 -0
  3. parse/dev/QDdJhACYrlX/images/a3ef5b43ccb2e661888ac50707a19658f8b2699c2c0f09ad2fd9022345745223.jpg +3 -0
  4. parse/dev/QDdJhACYrlX/images/af8722ba66d345311be1ae680a9516837a89a87fe3cd2b4f0a2fd97d4e38d10f.jpg +3 -0
  5. parse/dev/QDdJhACYrlX/images/b9989602e5f081dc7ac4ea8a3040de0d3256ce416ffb3534d6e4f937f7ccba15.jpg +3 -0
  6. parse/dev/QDdJhACYrlX/images/cb2812d32407c26900f1d01873383bc5ca02a8bbe3edf9f8734a6647821212af.jpg +3 -0
  7. parse/dev/QDdJhACYrlX/images/e4ef50aaf114a39c69ff6f5ee7abc0787937d0c7df0978812d6ad66c180576a1.jpg +3 -0
  8. parse/dev/QDdJhACYrlX/images/e85b0ac8caee07f38ee0a088bcfdbb93f03571dff59a21651d8a8b1ff2345225.jpg +3 -0
  9. parse/dev/QDdJhACYrlX/images/f2aa394fced168b31ab355c9972d010e64ac3fc2ef104a28f4c3825000c2ecf9.jpg +3 -0
  10. parse/dev/QFQoxCFYEkA/QFQoxCFYEkA.md +380 -0
  11. parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_content_list.json +1866 -0
  12. parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_layout.pdf +3 -0
  13. parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_middle.json +0 -0
  14. parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_model.json +0 -0
  15. parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_origin.pdf +3 -0
  16. parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_span.pdf +3 -0
  17. parse/dev/QFQoxCFYEkA/images/0908ac74d76efb399be6b4e6ed74bf3d3fca08f06344eae5ac521211e38a743a.jpg +3 -0
  18. parse/dev/QFQoxCFYEkA/images/1a988e48c53129ddc9758f443bfe72aa03843d05f07004aaa83c1cf5cc583af6.jpg +3 -0
  19. parse/dev/QFQoxCFYEkA/images/27c99bde5f56a13daec53f95a059c921d6106a82bd4cb6ae2766b67a7e1d46cf.jpg +3 -0
  20. parse/dev/QFQoxCFYEkA/images/2b2e84d45592acc7d50816d7c7dbfcba9119ec8fa3c317113edc2160d6ff6eef.jpg +3 -0
  21. parse/dev/QFQoxCFYEkA/images/2ce9636880a058cd0713d603581fa7f3e6acfe162a440e0ad227547b6544a41c.jpg +3 -0
  22. parse/dev/QFQoxCFYEkA/images/47710160045685a73f2907778920be78282559b5f39d434ffbd97b2a67637d92.jpg +3 -0
  23. parse/dev/QFQoxCFYEkA/images/520285b0f14a26c736e532c26652dfd9fc8275befa8c250aad7dda976e95bca6.jpg +3 -0
  24. parse/dev/QFQoxCFYEkA/images/6c71d9f5a4178d1b48f2d19bbb98de16ffa57a60378aaed0df0bb58c218884ba.jpg +3 -0
  25. parse/dev/QFQoxCFYEkA/images/6cb24af2fdb71e7a2fae8cfbeb64dcd6225000df4d23a9f2c3fa4e998eb09be9.jpg +3 -0
  26. parse/dev/QFQoxCFYEkA/images/76d9a9a4dd165c97d9c6b81648bf69bc02c6b45b06c64e217312f1914f3b33a5.jpg +3 -0
  27. parse/dev/QFQoxCFYEkA/images/7a8fc861426d5b4f5057042fc82a4f484bcbc4cdbccc40842aeea7e1202f7fbe.jpg +3 -0
  28. parse/dev/QFQoxCFYEkA/images/7afe915a39613e67bcc59c3ee4f8eec8f1486e67c0d910606b0ebda3e3ab2535.jpg +3 -0
  29. parse/dev/QFQoxCFYEkA/images/86839b766d8522061366b4faa54518774846b2b1c44f4ba7188558d299c00bfc.jpg +3 -0
  30. parse/dev/QFQoxCFYEkA/images/997f84347ee2f7b0394beb308bb0fc8e563a958887577e7c3d644b5d494bb0e8.jpg +3 -0
  31. parse/dev/QFQoxCFYEkA/images/b9a7f312faf7e4a780e52e674aa461846742d3a0e70f090e75b0124546d4c1e1.jpg +3 -0
  32. parse/dev/QFQoxCFYEkA/images/bc06aa78c13f13ab8794fd3beec2575b3d720ca22c180a49252fd14c96d37d5f.jpg +3 -0
  33. parse/dev/QFQoxCFYEkA/images/c560833d9a5c700338ba1d8898f10c61b70db8bc9e23feb190e95096a4188af9.jpg +3 -0
  34. parse/dev/QFQoxCFYEkA/images/c963de924077ad4d47b94f877b9169c34bb7f0e549bd8ea7f5d0b06ddaa6a9a1.jpg +3 -0
  35. parse/dev/QFQoxCFYEkA/images/de28f63560d2cb651e1066d0a90842918e0e047cf416964e25375636f1fc64fa.jpg +3 -0
  36. parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux.md +351 -0
  37. parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_content_list.json +1839 -0
  38. parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_layout.pdf +3 -0
  39. parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_middle.json +0 -0
  40. parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_model.json +0 -0
  41. parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_origin.pdf +3 -0
  42. parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_span.pdf +3 -0
  43. parse/dev/QJb1-8NH2Ux/images/1703fa013d6a1ee6b00378a016c2a93fc0e3073c9d6734ac06a5bcc5e5e8f741.jpg +3 -0
  44. parse/dev/QJb1-8NH2Ux/images/280c3197d6ce35502a31b2d1bc26122b4c360c6fa70eabe111b6e26a291e0bec.jpg +3 -0
  45. parse/dev/QJb1-8NH2Ux/images/308cd424cf793a3ba0100fdea38ed9dfb95bea708f1ec26aa902357976d5420b.jpg +3 -0
  46. parse/dev/QJb1-8NH2Ux/images/35afe76e5c882d9d74a70d86ea496028c6c8d6414cd130e3ebb49471d266dd3d.jpg +3 -0
  47. parse/dev/QJb1-8NH2Ux/images/44c657694cd74a56c1870443ece67a7d806cd61222c843330a77a072ca61eca5.jpg +3 -0
  48. parse/dev/QJb1-8NH2Ux/images/8e622d645b3f35c117ff52694ec9773f02e8ea8238d5cfacd531d2db3bd601ae.jpg +3 -0
  49. parse/dev/QJb1-8NH2Ux/images/bf3313c2eebf9b088d68940d3fcdf78503af5aea78a5151d35b998f8ac2a42af.jpg +3 -0
  50. parse/dev/QNBzcgY0f4e/QNBzcgY0f4e.md +526 -0
parse/dev/QDdJhACYrlX/images/9becef61539f1118df436f5341192c61d1cb7f755f868cd727b3bf797788ee40.jpg ADDED

Git LFS Details

  • SHA256: 5db4665db24dcac3519b1770c461bc1ea995c7164aa1705e98394c9f11c49631
  • Pointer size: 131 Bytes
  • Size of remote file: 126 kB
parse/dev/QDdJhACYrlX/images/a3cb1246375e83911e1c3dd3a3d07c8fd688315b5853000c0482ebd0667a70c6.jpg ADDED

Git LFS Details

  • SHA256: 64faa1df406abda56bc8ba4f699e7e553ce0aeec2f499f88c5517ac7628282e6
  • Pointer size: 129 Bytes
  • Size of remote file: 4.22 kB
parse/dev/QDdJhACYrlX/images/a3ef5b43ccb2e661888ac50707a19658f8b2699c2c0f09ad2fd9022345745223.jpg ADDED

Git LFS Details

  • SHA256: 38fbfbd6d2f94fadcf619c6e58d56e88e3f98412afb5a88f7145367b884ca3e2
  • Pointer size: 129 Bytes
  • Size of remote file: 9.93 kB
parse/dev/QDdJhACYrlX/images/af8722ba66d345311be1ae680a9516837a89a87fe3cd2b4f0a2fd97d4e38d10f.jpg ADDED

Git LFS Details

  • SHA256: cc0aa924c845e1c2d1ea4516275807897ab1e65e78a86c526def77aaee248bbc
  • Pointer size: 130 Bytes
  • Size of remote file: 13 kB
parse/dev/QDdJhACYrlX/images/b9989602e5f081dc7ac4ea8a3040de0d3256ce416ffb3534d6e4f937f7ccba15.jpg ADDED

Git LFS Details

  • SHA256: 116c19d00ff4ab8d98ce7b83d406e9fcce0f728bfd91538596551c696d050093
  • Pointer size: 129 Bytes
  • Size of remote file: 9.27 kB
parse/dev/QDdJhACYrlX/images/cb2812d32407c26900f1d01873383bc5ca02a8bbe3edf9f8734a6647821212af.jpg ADDED

Git LFS Details

  • SHA256: b6736f243bdac1a6a74eac24d1f2311487b980b26c65cd3ca70849b4fe8a1e9e
  • Pointer size: 130 Bytes
  • Size of remote file: 35.1 kB
parse/dev/QDdJhACYrlX/images/e4ef50aaf114a39c69ff6f5ee7abc0787937d0c7df0978812d6ad66c180576a1.jpg ADDED

Git LFS Details

  • SHA256: 0db16db3eccf642b303f61290e85f66107ff7473d782cd2629170b542aabd500
  • Pointer size: 130 Bytes
  • Size of remote file: 32.9 kB
parse/dev/QDdJhACYrlX/images/e85b0ac8caee07f38ee0a088bcfdbb93f03571dff59a21651d8a8b1ff2345225.jpg ADDED

Git LFS Details

  • SHA256: 7c39cee8feb7e8cf41be00f4c1c87f3cdae4b8ef600147dee15c79fab6a45f43
  • Pointer size: 130 Bytes
  • Size of remote file: 29.6 kB
parse/dev/QDdJhACYrlX/images/f2aa394fced168b31ab355c9972d010e64ac3fc2ef104a28f4c3825000c2ecf9.jpg ADDED

Git LFS Details

  • SHA256: d45b9a69f98ec921eedc3e115ce5597ce3f23148260514cac912b68fbf127b8a
  • Pointer size: 131 Bytes
  • Size of remote file: 107 kB
parse/dev/QFQoxCFYEkA/QFQoxCFYEkA.md ADDED
@@ -0,0 +1,380 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # DENSE: Data-Free One-Shot Federated Learning
2
+
3
+ Jie Zhang1∗ Chen Chen1∗ Bo Li2‡ Lingjuan Lyu3‡ Shuang Wu2 Shouhong Ding2 Chunhua Shen1 Chao Wu1‡
4
+
5
+ 1Zhejiang University 2Youtu Lab, Tencent 3 Sony AI
6
+
7
+ {zj_zhangjie, chao.wu, cc33} $@$ zju.edu.cn, Lingjuan.Lv $@$ sony.com {libraboli, calvinwu, ericshding}@tencent.com, chunhua $@$ me.com
8
+
9
+ # Abstract
10
+
11
+ One-shot Federated Learning (FL) has recently emerged as a promising approach, which allows the central server to learn a model in a single communication round. Despite the low communication cost, existing one-shot FL methods are mostly impractical or face inherent limitations, e.g., a public dataset is required, clients’ models are homogeneous, and additional data/model information need to be uploaded. To overcome these issues, we propose a novel two-stage Data-freE oNeShot federated lEarning (DENSE) framework, which trains the global model by a data generation stage and a model distillation stage. DENSE is a practical one-shot FL method that can be applied in reality due to the following advantages: (1) DENSE requires no additional information compared with other methods (except the model parameters) to be transferred between clients and the server; (2) DENSE does not require any auxiliary dataset for training; (3) DENSE considers model heterogeneity in FL, i.e., different clients can have different model architectures. Experiments on a variety of real-world datasets demonstrate the superiority of our method. For example, DENSE outperforms the best baseline method Fed-ADI by $5 . 0 8 \%$ on CIFAR10 dataset.
12
+
13
+ # 1 Introduction
14
+
15
+ Deep neural networks (DNNs) have recently gained popularity as a powerful tool for advancing artificial intelligence in both established and emerging fields [25, 22, 13, 9, 8, 51, 52, 58, 5, 4, 16]. Federated learning (FL) [42] has emerged as a promising learning paradigm which allows multiple clients to collaboratively train a global model without exposing their private training data. In FL, each client trains a local model on its own data and is required to periodically share its high-dimensional model parameters with a central server. Recent years, FL has shown its potential to facilitate real-world applications in many fields, including medical image analysis [36, 6], recommender systems [34, 38], natural language processing [63, 46] and computer vision [27, 26].
16
+
17
+ The original FL framework requires participants to communicate frequently with the central server in order to exchange models. In real-world FL, such high communication cost may be intolerable and impractical. Reducing the communication cost between clients and the server is desired both for system efficiency and to support the privacy goals of federated learning [40, 41]. Recent research proposed some common methods to reduce communication costs, e.g., utilize multiple local updates [18], employ compression techniques [48], and one-shot FL [10]. Among them, one-shot FL is a promising solution which only allows one communication round. There are several motivations behind one-shot FL: 1) First of all, multi-round training is not practical in some scenarios such as model markets [45], in which users can only buy the pre-trained models from the market without any real data. 2) Furthermore, frequent communication poses a high risk of being attacked. For instance, frequent communication can be easily intercepted by attackers, who can launch man-in-the-middle attacks [47] or even reconstruct the training data from gradients [54]. In this way, one-shot FL can reduce the probability of being intercepted by malicious attackers due to the one-round property. Thus, in this paper, we mainly focus on one-shot FL.
18
+
19
+ However, existing one-shot FL studies [10, 30, 62, 7] are still hard to apply in real-world applications, due to impractical settings. For example, Guha et al. [10] and Li et al. [30] involved a public dataset for training, which may be impractical in very sensitive scenarios such as the biomedical domains. Zhu et al. [62] adopted dataset distillation [50] in one-shot FL, but they need to send distilled data to the central server, which causes additional communication cost and potential privacy leakage. Dennis et al. [7] utilized cluster-based method in one-shot FL, which requires to upload the cluster means to the server, causing additional communication cost. Additionally, none of these methods consider model heterogeneity, i.e., different clients have different model architectures [31], which is very common in practical scenarios. For instance, in model market, models sold by different sellers are likely to be heterogeneous. Besides, when several medical institutions participate in FL, they may need to design their own model to meet distinct specifications. Therefore, developing a practical one-shot FL method is in urgent need.
20
+
21
+ In this work, we propose a novel two-stage Data-freE oNe-Shot federated lEarning (DENSE) framework, which trains the global model by a data generation stage and a model distillation stage. In the first stage, we utilize the ensemble models (i.e., ensemble of local models uploaded by clients) to train a generator, which can generate synthetic data for training in the second stage. In the second stage, we distill the knowledge of the ensemble models to the global model. In contrast to traditional FL methods based on FedAvg [42], our method does not require averaging of model parameters, thus it can support heterogeneous models, i.e., clients can have different model architectures. In summary, our main contributions are summarized as follows:
22
+
23
+ • We propose a novel data-free one-shot FL framework named DENSE, which consists of two stages. In the first stage, we train a generator that considers similarity, stability, and transferability at the same time. In the second stage, we use the ensemble models and the data generated by the generator to train a global model.
24
+ The setting of DENSE is practical in the following aspects. First, DENSE requires no additional information (except the model parameters) to be transferred between clients and the server; Second, DENSE does not require any auxiliary dataset for training; Third, DENSE considers model heterogeneity, i.e., different clients can have different model architectures.
25
+ DENSE is a compatible approach, which can be combined with any local training techniques to further improve the performance of the global model. For instance, we can adopt LDAM [1] to train the clients’ local models, and improve the accuracy of the global model (refer to Section 2.3 and Section 3.2). Extensive experiments on various datasets verify the effectiveness of our proposed DENSE. For example, DENSE outperforms the best baseline method Fed-ADI [55] by $5 . 0 8 \%$ on CIFAR10 dataset.
26
+
27
+ # 2 Data-Free One-Shot Federated Learning
28
+
29
+ # 2.1 Framework Overview
30
+
31
+ To tackle the problems in recent one-shot FL methods as mentioned in Sec. 1, we propose a novel method named DENSE, which conducts one-shot FL without the need to share additional information or rely on any auxiliary dataset, while considering model heterogeneity. To simulate real-world applications, we consider a more challenging yet practical setting where the data on each client are not independent and identically distributed (non-IID).
32
+
33
+ The illustration of the learning procedure is demonstrated in Figure 1, and the whole training process of DENSE is shown in Algorithm 1. After clients upload their local models to the server, the server trains a global model with DENSE in two stages. In the data generation stage (first stage), we train an auxiliary generator that can generate synthetic data by the ensemble models, i.e., ensemble of local models uploaded by clients. In the model distillation stage (second stage), we use the ensemble models and the synthetic data (generated by the generator) to train the global model.
34
+
35
+ ![](images/2ce9636880a058cd0713d603581fa7f3e6acfe162a440e0ad227547b6544a41c.jpg)
36
+ Figure 1: An illustration of training process of DENSE on the server, which consists of two stages: (1) In data generation stage, we train an auxiliary generator that considers similarity, stability, and transferability at the same time; (2) In model distillation stage, we distill the knowledge of the ensemble models and transfer to the global model. Note that the fixed global model is used as an additional discriminator in the divergence loss $\mathcal { L } _ { d i v }$ .
37
+
38
+ # 2.2 Data Generation
39
+
40
+ In the first stage, we aim to train a generator to generate synthetic data. Specifically, given the ensemble of well-trained models uploaded by clients, our goal is to train a generator that can generate data that have similar distribution to the training data of clients. In addition, we aim not to leak private information from our generated data, i.e., attackers are not able to predict any sensitive information of clients from the generated data. Recent work [35] generated data by utilizing a pre-trained generative adversarial network (GAN). However, such a method is unable to generate data as the pre-trained GAN is trained on public datasets, which is likely to have different data distribution from the training data of clients. Moreover, we need to consider model heterogeneity, which makes the problem more complicated.
41
+
42
+ To solve these issues, we propose to train a generator that considers similarity\*, stability, and transferability. The data generation process is shown in line 8 to 11 in Algorithm 1. In particular, given a random noise $\mathbf { z }$ (generated from a standard Gaussian distribution) and a random one-hot label y (generated from a uniform distribution), the generator $G ( \cdot )$ aims to generate a synthetic data $\hat { \mathbf { x } } = G ( \mathbf { z } )$ such that $\hat { \bf x }$ is similar to the training data (with label $\mathbf { y }$ ) of clients.
43
+
44
+ Similarity. First, we need to consider the similarity between synthetic data $\hat { \bf x }$ and the training data. Since we are unable to access the training data of clients, we cannot compute the similarity between the synthetic data and the training data directly. Instead, we first compute the average logits (i.e., outputs of the last fully connected layer) of $\hat { \bf x }$ computed by the ensemble models.
45
+
46
+ $$
47
+ D ( \hat { \mathbf { x } } ; \{ \pmb { \theta } ^ { k } \} _ { k = 1 } ^ { m } ) = \frac { 1 } { m } \sum _ { k \in \mathcal { C } } f ^ { k } \left( \hat { \mathbf { x } } ; \pmb { \theta } ^ { k } \right) ,
48
+ $$
49
+
50
+ where $m = | \mathcal { C } |$ , and $D ( \hat { \mathbf { x } } ; \{ \pmb { \theta } ^ { k } \} _ { k = 1 } ^ { m } )$ is the average logits of $\hat { \bf x }$ , $\pmb { \theta } ^ { k }$ is the parameter of the $k$ -th client.
51
+ And $f ^ { k } \left( { \hat { \mathbf { x } } } ; \pmb { \theta } ^ { k } \right)$ is the prediction function of client $k$ that outputs the logits of $\hat { \bf x }$ given parameter $\theta ^ { k }$ .
52
+ For simplicity, we use $D ( \hat { \mathbf { x } } )$ to denote $D ( \hat { \mathbf { x } } ; \{ \pmb { \theta } ^ { k } \} _ { k = 1 } ^ { m } )$ in the rest of the paper.
53
+
54
+ Then, we minimize the average logits and the random label $y$ with the following cross-entropy (CE) loss.
55
+
56
+ $$
57
+ \mathcal { L } _ { C E } ( \hat { \mathbf { x } } , \mathbf { y } ; \pmb { \theta } _ { G } ) = C E ( D ( \hat { \mathbf { x } } ) , \mathbf { y } ) ,
58
+ $$
59
+
60
+ It is expected that the synthetic images can be classified into one particular class with a high probability by the ensemble models. In fact, during the training phase, the loss between $D ( \hat { \mathbf { x } } )$ and $\mathbf { y }$ can easily reduce to almost 0, which indicates the synthetic data matches the ensemble models perfectly. Moreover, we do not directly compute the similarity between the synthetic data and the training data, which can reduce the probability of leaking sensitive information of the clients.
61
+
62
+ However, by utilizing only the CE loss, we cannot achieve a high performance (please refer to Section 3.2 for detail). We conjecture this is because the ensemble models are trained on non-IID data, the generator may be unstable and trapped into sub-optimal local minima or overfit to the synthetic data [49, 32].
63
+
64
+ Stability. Second, to improve the stability of the generator, we propose to add an additional regularization to stabilize the training. In particular, we utilize the Batch Normalization (BN) loss to make the synthetic data conform with the batch normalization statistics [55].
65
+
66
+ $$
67
+ \mathcal { L } _ { B N } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { G } ) = \frac { 1 } { m } \sum _ { k \in \mathcal { C } } \sum _ { l } \left( \lVert \mu _ { l } ( \hat { \mathbf { x } } ) - \mu _ { k , l } \rVert + \lVert \sigma _ { l } ^ { 2 } ( \hat { \mathbf { x } } ) - \sigma _ { k , l } ^ { 2 } \rVert \right) ,
68
+ $$
69
+
70
+ where $\mu _ { l } ( \hat { \mathbf { x } } )$ and $\sigma _ { l } ^ { 2 } ( \hat { \mathbf { x } } )$ are the batch-wise mean and variance estimates of feature maps corresponding to the $l$ -th BN layer of the generator $G ( \cdot ) ^ { \dagger }$ , $\mu _ { k , l }$ and $\sigma _ { k , l } ^ { 2 }$ are the mean and variance of the $l$ -th BN layer [17] of $f ^ { k } ( \cdot )$ . The BN loss minimizes the distance between the feature map statistics of the synthetic data and the training data of clients. As a result, the synthetic data can have a similar distribution to the training data of clients, no matter if the data is non-IID or IID.
71
+
72
+ Transferability. By utilizing the CE loss and BN loss, we can train a generator that can generate synthetic data, but we observed that the synthetic data are likely to be far away from the decision boundary (of the ensemble models), which makes the ensemble models (teachers) hard to transfer their knowledge to the global model (student). We illustrate the observation in the left panel of Figure 2. S and T are the decision boundaries of the global model (the detail of the global model is introduced in Section 2.3) and ensemble models respectively. The essence of knowledge distillation is transferring the information of decision boundary from the teacher model to the student model [12]. We aim to learn the decision boundary of global model and have a high classification accuracy on the real test data (blue diamonds). However, the generated synthetic data (red circles) are likely to be on the same side of the two decision boundaries and unhelpful to the transfer of knowledge [12]. To solve this problem, we
73
+
74
+ ![](images/76d9a9a4dd165c97d9c6b81648bf69bc02c6b45b06c64e217312f1914f3b33a5.jpg)
75
+ Figure 2: The illustration of generated data and decision boundary of ensemble models (teachers) and global model (student). Left panel: Synthetic data (red circles) are far away from the decision boundary, which is less helpful to the transfer of knowledge. Right panel: By utilizing our boundary support loss, we can generate more synthetic data near the decision boundaries (black circles), which helps the student better learn the decision boundary of the teacher.
76
+
77
+ argue to generate more synthetic data that fall between the decision boundaries of the ensemble models and the global model. We illustrate our idea in the right panel of Figure 2. Red circles are synthetic data on the same side of the decision boundary, which are less helpful in learning the global model. Black circles are synthetic data between the decision boundaries, i.e., the global model and the ensemble models have different predictions on these data. Black circles can help the global model better learn the decision boundary of the ensemble models.
78
+
79
+ Motivated by the above observations, we introduce a new boundary support loss, which urges the generator to generate more synthetic data between the decision boundaries of the ensemble models and the global model. We divide the synthetic data into 2 sets: (1) the global model and the ensemble models have the same predictions on data in the first set $\begin{array} { r } { \arg \operatorname* { m a x } _ { c } D ^ { ( c ) } ( \hat { \mathbf { x } } ) = \arg \operatorname* { m a x } _ { c } f _ { S } ^ { ( c ) } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { S } ) ) . } \end{array}$ ; (2) different predictions on data in the second set ${ ( \arg \operatorname* { m a x } _ { c } D ^ { ( c ) } ( \hat { \bf x } ) \neq \arg \operatorname* { m a x } _ { c } f _ { S } ^ { ( c ) } ( \hat { \bf x } ; \pmb { \theta } _ { S } ) ) }$ , where $D ^ { ( c ) } ( \hat { \mathbf { x } } )$ and $f _ { S } ^ { ( c ) } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { S } )$ are the logits for the $c$ -th label of the ensemble models and the global model respectively. The data in the first set are on the same side of those two decision boundaries (red circles in Figure 2) while the data in the second set (black circles in Figure 2) are between the decision boundaries of the ensemble models and the global model. We maximize the differences of predictions of the global model and the ensemble models on data in the second set with Kullback-Leibler divergence loss as follows.
80
+
81
+ $$
82
+ \mathcal { L } _ { d i v } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { G } ) = - \omega K L \left( D ( \hat { \mathbf { x } } ) , f _ { S } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { S } ) \right) ,
83
+ $$
84
+
85
+ where $K L ( \cdot , \cdot )$ denotes the Kullback-Leibler (KL) divergence loss, $\omega = \mathbb { 1 } ( \arg \operatorname* { m a x } _ { c } D ^ { ( c ) } ( \hat { \mathbf { x } } ) \neq$ arg $\operatorname* { m a x } _ { c } f _ { S } ^ { ( c ) } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { S } ) )$ outputs 0 for data in the first set and 1 for data in the second set, and $\mathbb { 1 } ( a )$ is the indicator function that outputs 1 if $a$ is true and outputs 0 if $a$ is false. By maximizing the KL divergence loss, the generator can generate more synthetic data that are more helpful to the model distillation stage (refer to Section 2.3 for detail) and further improve the transferability of the ensemble models.
86
+
87
+ By combining the above losses, we can obtain the generator loss as follows,
88
+
89
+ $$
90
+ \begin{array} { r } { \mathcal { L } _ { g e n } ( \hat { \mathbf { x } } , \mathbf { y } ; \pmb { \theta } _ { G } ) = \mathcal { L } _ { C E } ( \hat { \mathbf { x } } , \mathbf { y } ; \pmb { \theta } _ { G } ) + \lambda _ { 1 } \mathcal { L } _ { B N } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { G } ) + \lambda _ { 2 } \mathcal { L } _ { d i v } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { G } ) , } \end{array}
91
+ $$
92
+
93
+ where $\lambda _ { 1 }$ and $\lambda _ { 2 }$ are scaling factors for the losses.
94
+
95
+ # Algorithm 1 Training process of DENSE
96
+
97
+ Input: Number of client $m$ , clients’ local models $\{ f ^ { 1 } ( ) , \cdots , f ^ { m } ( ) \}$ , generator $G ( \cdot )$ with parameter $\theta _ { G }$ , learning rate of the generator $\eta _ { G }$ , number of training rounds $T _ { G }$ for generator in each epoch, global model $f _ { S } ( \ u )$ with parameter $\theta _ { S }$ , learning rate of the global model $\eta _ { S }$ , global model training epochs $T$ , and batch size $b$ .
98
+
99
+ for each client $k \in { \mathcal { C } }$ in parallel do $\pmb { \theta } ^ { k } \mathbf { L o c a l U p d a t e } ( k )$
100
+ end for
101
+ Initialize parameter $\theta _ { G }$ and $\theta _ { S }$
102
+ for epoch $\scriptstyle 1 = 1 , \cdots , T$ do Sample a batch of noises and labels $\{ \mathbf { z } _ { i } , \mathbf { y } _ { i } \} _ { i = 1 } ^ { b }$ // data generation stage for $j = 1 , \cdots , T _ { G }$ do Generate $\{ \hat { \mathbf { x } } _ { i } \} _ { i = 1 } ^ { b }$ with $\{ { \mathbf { z } } _ { i } \} _ { i = 1 } ^ { b }$ and $G ( \cdot )$ enc $\begin{array} { r } { \pmb { \theta } _ { G } \pmb { \theta } _ { G } - \eta _ { G } \frac { 1 } { b } \sum _ { i = 1 } ^ { b } \nabla _ { \pmb { \theta } _ { G } } \ell _ { g e n } ( \hat { \mathbf { x } } _ { i } , \mathbf { y } _ { i } ; \pmb { \theta } _ { G } ) } \end{array}$ $/ /$ model distillation stage Generate $\{ \hat { \mathbf { x } } _ { i } \} _ { i = 1 } ^ { b }$ with $\mathbf { \bar { \{ z } } _ { i } \} _ { i = 1 } ^ { b }$ and $G ( \cdot )$ $\begin{array} { r } { \pmb { \theta } _ { S } = \pmb { \theta } _ { S } - \eta _ { S } \frac { 1 } { b } \sum _ { i = 1 } ^ { b } \nabla _ { \pmb { \theta } _ { S } } \ell _ { d i s } ( \hat { \bf x } _ { i } ; \pmb { \theta } _ { S } ) } \end{array}$
103
+ end for
104
+
105
+ Note that the generated synthetic data have similar features but different from the training data (of clients), which reduces the probability of leaking sensitive information of clients. More discussions of the privacy issues are in Section 3.3.3.
106
+
107
+ # 2.3 Model Distillation
108
+
109
+ In the second stage, we train the global model with the generator (discussed in the previous section) and the ensemble models. Previous research [60, 35] showed that model ensemble provides a general method for improving the accuracy and stability of learning models. Motivated by [60], we propose to use the ensemble models as a teacher to train a student (global) model. A straightforward method is to obtain the global model by aggregating the parameters of all client models (e.g., by FedAvg [42]). However, in real-world applications, clients are likely to have different model architectures [44], making FedAvg useless. Moreover, since the data in different clients are non-IID, FedAvg cannot deliver a good performance or even diverge [59, 32].
110
+
111
+ To this end, we follow [35] to distill the knowledge of the ensemble models to the global model by minimizing the predictions between the ensemble models (teacher) and the global model (student) on the same synthetic data. The model distillation process is shown in line 13 to 14 in Algorithm 1. First, we compute the average logits of the synthetic data according to Eq. (1), i.e., $D ( \hat { \mathbf { x } } ) =$ $\begin{array} { r } { \frac { 1 } { m } \sum _ { k \in \mathcal { C } } f ^ { k } \left( \hat { \mathbf { x } } ; \pmb { \theta } ^ { k } \right) } \end{array}$ . In contrast to traditional aggregation methods (e.g., FedAvg) that are unable to aggregate heterogeneous models, averaging logits can be easily applied to both heterogeneous and homogeneous FL systems.
112
+
113
+ Then, we use the average logits to distill the knowledge of the ensemble models by minimizing the following objective function.
114
+
115
+ $$
116
+ \mathcal { L } _ { d i s } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { S } ) = K L \left( D ( \hat { \mathbf { x } } ) , f _ { S } ( \hat { \mathbf { x } } ; \pmb { \theta } _ { S } ) \right) .
117
+ $$
118
+
119
+ By minimizing the KL loss, we can train a global model with the knowledge of the ensemble models and the synthetic data regardless of data and model heterogeneity.
120
+
121
+ Note that DENSE has no restriction on the clients’ local models, i.e., clients can train models with arbitrary techniques. Thus, DENSE is a compatible approach, which can be combined with any local training techniques to further improve the performance of the global model. We further discuss the combination of local training techniques in Section 3.2.
122
+
123
+ Discussions on privacy-preserving Research [15] has shown that it is possible to launch an attack where a malicious user uses GANs to recreate samples of another participant’s private datasets. Besides, in FL, exchanging models between the server and clients can result in potential privacy leakage. Note that our method prohibits the generator from seeing the real data directly, and there is only one communication round, which reduces the risk of privacy leakage. In addition, we display our generated images in Section 3.3.3, which does not directly reveal the information of real data. Several existing privacy-preserving methods can be incorporated into our framework to better protect clients from adversaries [37, 20]. We leave this as our future work.
124
+
125
+ Discussions on Knowledge Distillation in FL In traditional FL frameworks, all users have to agree on the specific architecture of the global model. To support model heterogeneity, Li et al. [28] proposed a new federated learning framework that enables participants to independently design their models by knowledge distillation [14]. With the use of a proxy dataset, knowledge distillation alleviates the model drift issue induced by non-IID data. However, the requirement of proxy data renders such a method impractical for many applications, since a carefully designed dataset is not always available on the server. Data-free knowledge distillation is a promising approach, which can transfer knowledge of a teacher model to a student model without any real data [2, 55]. Lin et al. [35] proposed data-free ensemble distillation for model fusion through synthetic data in each communication round, which requires high communication costs and computational costs. However, in this paper, we are more concerned with obtaining a good global model through only one round of communication in cases of heterogeneous models, which is more challenging and practical. Zhu et al. [64] also proposed a data-free knowledge distillation approach for FL, which learns a generator derived from the prediction of local models. However, the learned generator is later broadcasted to all clients, and then clients need to send their generators to the server, which increases the communication burden. More seriously, the generator has direct access to the local data (the generator can easily remember the training samples [39]), which can cause privacy concerns. As the generator used in our method is always stored in the central server, it never sees any real local data.
126
+
127
+ # 3 Experiments
128
+
129
+ # 3.1 Experimental Setup
130
+
131
+ # 3.1.1 Datasets
132
+
133
+ Our experiments are conducted on the following 6 real-world datasets: MNIST [24], FMNIST [53], SVHN [43], CIFAR10 [21], CIFAR100 [21], and Tiny-ImageNet [23].MNIST dataset contains binary images of handwritten digits. There are 60,000 training images and 10,000 testing images in MNIST dataset. CIFAR10 dataset consists of $6 0 , 0 0 0 \ 3 2 \mathrm { x } \ 3 2$ color images in 10 classes, with 6,000 images per class. There are 50,000 training images and 10,000 test images in CIFAR10 dataset. CIFAR100 dataset is similar to CIFAR10 dataset, except it has 100 classes containing 600 images each. There are 500 training images and 100 testing images per class. Tiny-ImageNet contains 100000 images of 200 classes (500 for each class) downsized to $6 4 { \times } 6 4$ colored images. Each class has 500 training images, 50 validation images and 50 test images.
134
+
135
+ Table 1: Accuracy of different methods across $\alpha = \{ 0 . 1 , 0 . 3 , 0 . 5 \}$ on different datasets.
136
+
137
+ <table><tr><td>Dataset</td><td colspan="3">MNIST</td><td colspan="3">FMNIST</td><td colspan="3">CIFAR10</td><td colspan="3">SVHN</td><td colspan="3">CIFAR100</td><td colspan="3">Tiny-ImageNet</td></tr><tr><td>Method</td><td>a=0.1</td><td>Q=0.3</td><td>a=0.5</td><td>α=0.1</td><td>Q=0.3</td><td>Q=0.5</td><td>α=0.1</td><td>a=0.3</td><td>α=0.5</td><td>α=0.1</td><td>Q=0.3</td><td>α=0.5</td><td>Q=0.1</td><td>a=0.3</td><td>α=0.5</td><td></td><td>a=0.1 a=0.3</td><td>a=0.5</td></tr><tr><td>FedAvg</td><td>48.24</td><td>72.94</td><td>90.55</td><td>41.69</td><td>82.96</td><td>83.72</td><td>23.93</td><td>27.72</td><td>43.67</td><td>31.65</td><td>61.51</td><td>56.09</td><td>4.58</td><td>11.61</td><td>12.11</td><td>3.12</td><td>10.46</td><td>11.89</td></tr><tr><td>FedDF</td><td>60.15</td><td>74.01</td><td>92.18</td><td>43.58</td><td>80.67</td><td>84.67</td><td>40.58</td><td>46.78</td><td>53.56</td><td>49.13</td><td>73.34</td><td>73.98</td><td>28.17</td><td>30.28</td><td>36.35</td><td>15.34</td><td>18.22</td><td>27.43</td></tr><tr><td>Fed-DAFL</td><td>64.38</td><td>74.18</td><td>93.01</td><td>47.14</td><td>80.59</td><td>84.02</td><td>47.34</td><td>53.89</td><td>58.59</td><td>53.23</td><td>76.56</td><td>78.03</td><td>28.89</td><td>34.89</td><td>38.19</td><td>18.38</td><td>22.18</td><td>28.22</td></tr><tr><td>Fed-ADI</td><td>64.13</td><td>75.03</td><td>93.49</td><td>48.49</td><td>81.15</td><td>84.19</td><td>48.59</td><td>54.68</td><td>59.34</td><td>53.45</td><td>77.45</td><td>78.85</td><td>30.13</td><td>35.18</td><td>40.28</td><td>19.59</td><td>25.34</td><td>30.21</td></tr><tr><td>DENSE (ours)</td><td>66.61</td><td>76.48</td><td>95.82</td><td>50.29</td><td>83.96</td><td>85.94</td><td>50.26</td><td>59.76</td><td>62.19</td><td>55.34</td><td>79.59</td><td>80.03</td><td>32.03</td><td>37.32</td><td>42.07</td><td>22.44</td><td>28.14</td><td>32.34</td></tr></table>
138
+
139
+ # 3.1.2 Data partition
140
+
141
+ To simulate real-world applications, we use Dirichlet distribution to generate non-IID data partition among clients [56, 29]. In particular, we sample $p _ { k } \sim D i r ( \alpha )$ and allocate a $p _ { k } ^ { i }$ proportion of the data of class $k$ to client $i$ . By varying the parameter $\alpha$ , we can change the degree of imbalance. A small $\alpha$ generates highly skewed data. We set $\alpha = 0 . 5$ as default.
142
+
143
+ # 3.1.3 Baselines
144
+
145
+ To ensure fair comparisons, we neglect the comparison with methods that require to download auxiliary models or datasets, such as FedBE [3] and FedGen [64]. Moreover, since there is only one communication round, aggregation methods that are based on regularization have no effect. Thus, we also omit the comparison with these regularization-based methods, e.g., FedProx [31], FedNova [49], and Scaffold [18]. Instead, we compare our proposed DENSE with FedAvg [42] and FedDF [35]. Furthermore, since DENSE is a data-free method, we derive some baselines from prevailing data-free knowledge distillation methods, including: 1) DAFL [2], a novel data-free learning framework based on generative adversarial networks; 2) ADI [55], an image synthesizing method that utilizes the image distribution to train a deep neural network without real data. We apply these methods to one-shot FL, and name these two baselines as Fed-DAFL and Fed-ADI.
146
+
147
+ # 3.1.4 Settings
148
+
149
+ For clients’ local training, we use the SGD optimizer with momentum $= 0 . 9$ and learning rate ${ \mathrm { \Omega } } { = } 0 . 0 1$ . We set the batch size $b = 1 2 8$ , the number of local epochs $E = 2 0 0$ , and the client number $m = 5$ . Following the setting of [2], we train the auxiliary generator $G ( \cdot )$ with a deep convolutional network. We use Adam optimizer with learning rate $\eta _ { G } = 0 . 0 0 1$ . We set the number of training rounds in each epoch as $T _ { G } = 3 0$ , and set the scaling factor $\lambda _ { 1 } = 1$ and $\lambda _ { 2 } = 0 . 5$ . For the training of the server model $f _ { S } ( \ v r )$ , we use the SGD optimizer with learning rate $\eta _ { S } = 0 . 0 1$ and momentum $\scriptstyle 1 = 0 . 9$ . The number of epochs for distillation $T = 2 0 0$ . All baseline methods use the same setting as ours.
150
+
151
+ # 3.2 Results
152
+
153
+ # 3.2.1 Evaluation on real-world datasets
154
+
155
+ To evaluate the effectiveness of our method, we conduct experiments under different non-IID settings by varying $\alpha = \{ 0 . 1 , 0 . 3 , 0 . 5 \}$ and report the performance on different datasets and different methods in Table 1. The results show that: (1) Our DENSE achieves the highest accuracy across all datasets. In particular, DENSE outperforms the best baseline method Fed-ADI [55] by $5 . 0 8 \%$ when $\alpha = 0 . 3$ on CIFAR10 dataset. (2) FedAvg has the worst performance, which implies that directly averaging the model parameters cannot achieve a good performance under nonIID setting in one-shot FL. (3) As $\alpha$ becomes smaller (i.e., data become more imbalanced), the performance of all methods decrease significantly, which shows that all methods suffer
156
+
157
+ ![](images/7afe915a39613e67bcc59c3ee4f8eec8f1486e67c0d910606b0ebda3e3ab2535.jpg)
158
+ Figure 3: Left panel: Accuracy of FedAvg and clients’ local models across different local training epochs $E = \{ 2 0 , 4 0 , 6 0 , \cdots , 4 0 0 \}$ . Right panel: The accuracy curve for local training. The dotted lines represent the best results of two one-shot FL methods (FedAvg and DENSE). Our DENSE outperforms FedAvg and local models consistently.
159
+
160
+ Table 2: Accuracy comparisons across heterogeneous client models on CIFAR10. There are five clients in total, and each client has a personalized model.
161
+
162
+ <table><tr><td rowspan="2">Model</td><td colspan="5">Client</td><td colspan="4">Server (ResNet-18)</td></tr><tr><td>ResNet-18</td><td>CNN1</td><td>CNN2</td><td>WRN-16-1</td><td>WRN-40-1</td><td>FedDF</td><td>Fed-DAFL</td><td>Fed-ADI</td><td>DENSE (ours)</td></tr><tr><td>Q=0.1</td><td>40.83</td><td>33.67</td><td>35.21</td><td>27.73</td><td>32.93</td><td>42.35</td><td>43.12</td><td>44.63</td><td>49.76</td></tr><tr><td>Q=0.3</td><td>51.49</td><td>52.78</td><td>44.96</td><td>47.35</td><td>37.24</td><td>52.72</td><td>57.72</td><td>58.96</td><td>63.25</td></tr><tr><td>α=0.5</td><td>59.96</td><td>58.67</td><td>54.28</td><td>53.39</td><td>58.14</td><td>60.05</td><td>61.56</td><td>63.24</td><td>67.42</td></tr></table>
163
+
164
+ ![](images/997f84347ee2f7b0394beb308bb0fc8e563a958887577e7c3d644b5d494bb0e8.jpg)
165
+ Figure 4: Visualization of the test accuracy and data distribution for CIFAR10 with $\alpha = \{ 0 . 3 , 0 . 5 \}$
166
+
167
+ from highly skewed data. Even under highly skewed setting, DENSE still significantly outperforms other methods, which further demonstrates the superiority of our proposed method.
168
+
169
+ # 3.2.2 Impact of model distillation
170
+
171
+ We show the impact of model distillation by comparing with FedAvg. We first conduct one-shot FL and use FedAvg to aggregate the local models. We show the results of the global model and clients’ local models across different local training epochs $E = \{ 2 0 , 4 0 , 6 0 , \cdots , 4 0 0 \}$ in the left panel of Figure 3. The global model achieves the best performance (test accuracy $= 3 4 \%$ ) when $E = 4 0$ , while a larger value of $E$ can cause the model to degrade even collapse. This result can be attributed to the inconsistent optimization objectives with non-IID data [49], which leads to weight divergence [61]. Then, we show the results of one-shot FL when $E = 4 0 0$ and report the performance of FedAvg and DENSE in the right panel of Figure 3. We also plot the performance of clients’ local models. DENSE outperforms each client’s local model while FedAvg underperforms each client’s local model. This validates that model distillation can enhance training while directly aggregating is harmful to the training under non-IID setting in one-shot FL.
172
+
173
+ # 3.2.3 Results in heterogeneous FL
174
+
175
+ Note that our proposed DENSE can support heterogeneous models. We apply five different CNN models on CIFAR10 dataset with Dirichlet distribution $\alpha = \{ 0 . 1 , 0 . 3 , 0 . 5 \}$ . The heterogeneous models include: 1) one ResNet-18 [11], 2) two small CNNs: CNN1 and CNN2; 3) two Wide-ResNets (WRN) [57]: WRN-16-1 and WRN-40-1. For knowledge distillation, we use ResNet-18 as the server’s global model. Detailed architecture information of the given deep networks can be found in Appendix. Table 2 evaluates all methods in heterogeneous one-shot FL under practical non-IID data settings. We omit the results for FedAvg as FedAvg does not support heterogeneous models. We remark that FL under both the non-IID data distribution and different model architecture setting is a quite challenging task. Even under this setting, our DENSE still significantly outperforms other baselines. In addition, we report the accuracy curve of global distillation. As shown in Figure 4, our method outperforms other baselines by a large margin.
176
+
177
+ # 3.3 Analysis of Our Method
178
+
179
+ # 3.3.1 Impact of the number of clients
180
+
181
+ Furthermore, we evaluate the performance of these methods on CIFAR10 and SVHN datasets by varying the number of clients $m = \{ 5 , 1 0 , 2 0 , 5 0 , 1 0 0 \}$ . According to [33], the server can become a bottleneck when the number of clients is very large, we are also concerned with the model performance when $m$ increases. Table 3 shows the results of different methods across different $m$ . The accuracy of all methods decreases as the number of clients $m$ increases, which is consistent with observations in [33, 42]. Even though the number of clients can affect the performance of one-shot FL, our method still outperforms other baselines. The increasing number of clients poses new challenges for ensemble distillation, which we leave for future investigation.
182
+
183
+ Table 3: Accuracy across different number of clients $m = \{ 5 , 1 0 , 2 0 , 5 0 , 1 0 0 \}$ on CIFAR10 and SVHN datasets.
184
+
185
+ <table><tr><td>Dataset</td><td colspan="5">CIFAR10</td><td colspan="5">SVHN</td></tr><tr><td>m</td><td>FedAvg</td><td>FedDF</td><td>Fed-DAFL</td><td>Fed-ADI</td><td>DENSE (ours)</td><td>FedAvg</td><td>FedDF</td><td>Fed-DAFL</td><td>Fed-ADI</td><td>DENSE (ours)</td></tr><tr><td>5</td><td>1 43.67</td><td>53.56</td><td>55.46</td><td>58.59</td><td>62.19</td><td>56.09</td><td>73.98</td><td>78.03</td><td>78.85</td><td>80.03</td></tr><tr><td>10</td><td>38.29</td><td>54.44</td><td>56.34</td><td>57.13</td><td>61.42</td><td>45.34</td><td>62.12</td><td>63.34</td><td>65.45</td><td>67.57</td></tr><tr><td>20</td><td>36.03</td><td>43.15</td><td>45.98</td><td>46.45</td><td>52.71</td><td>47.79</td><td>60.45</td><td>62.19</td><td>63.98</td><td>66.42</td></tr><tr><td>50</td><td>37.03</td><td>40.89</td><td>43.02</td><td>44.47</td><td>48.47</td><td>36.53</td><td>51.44</td><td>54.23</td><td>57.35</td><td>59.27</td></tr><tr><td>100</td><td>33.54</td><td>36.89</td><td>37.55</td><td>36.98</td><td>43.28</td><td>30.18</td><td>46.58</td><td>47.19</td><td>48.33</td><td>52.48</td></tr></table>
186
+
187
+ Table 4: Performance analysis of DENSE+LDAM.
188
+
189
+ <table><tr><td>Dataset</td><td colspan="3">CIFAR10</td><td colspan="3">SVHN</td></tr><tr><td>Method</td><td>α=0.1</td><td>a=0.3</td><td>a=0.5</td><td>a=0.1</td><td>α=0.3</td><td>Q=0.5</td></tr><tr><td>DENSE</td><td>50.26</td><td>59.76</td><td>62.19</td><td>55.34</td><td>79.59</td><td>80.03</td></tr><tr><td>DENSE+LDAM</td><td>57.24</td><td>63.13</td><td>64.76</td><td>58.04</td><td>81.28</td><td>81.77</td></tr></table>
190
+
191
+ # 3.3.2 Combination with imbalanced learning
192
+
193
+ The accuracy of federated learning reduces significantly with non-IID data, which has been broadly discussed in recent studies [29, 49]. Additionally, previous studies [1, 19] have demonstrated their superiority on imbalanced data. The combination of our method with these techniques to address imbalanced local data can lead to a more effective FL system. For example, by using LDAM [1] in clients’ local training, we can mitigate the impact of data imbalance, and thereby build a more powerful ensemble model. We compare the performance of the original DENSE and DENSE combined with LDAM (DENSE+LDAM) across $\alpha = \{ 0 . 1 , 0 . 3 , 0 . 5 \}$ on CIFAR10 and SVHN datasets.
194
+
195
+ As demonstrated in Table 4, DENSE+LDAM can significantly improve the performance, especially for highly skewed non-IID data (i.e. $\alpha = 0 . 1$ ). To help understand the performance gap and data skewness, in Figure 5, we visualize the accuracy curve and data distribution of CIFAR10 $( \alpha { = } 0 . 1 )$ in the left panel and right panel respectively. The number in the right panel stands for the number of examples associated with the corresponding label in one particular client. These figures imply that significant improvement can be achieved by combining DENSE with LDAM on highly skewed data.
196
+
197
+ ![](images/86839b766d8522061366b4faa54518774846b2b1c44f4ba7188558d299c00bfc.jpg)
198
+ Figure 5: Left panel: Accuracy curves of DENSE and DENSE $+$ LDAM. Right panel: Data distribution of different clients for CIFAR10 dataset $( \alpha { = } 0 . 1 )$ .
199
+
200
+ ![](images/c560833d9a5c700338ba1d8898f10c61b70db8bc9e23feb190e95096a4188af9.jpg)
201
+ Figure 6: Visualization of synthetic data on CIFAR10 and SVHN datasets.
202
+
203
+ # 3.3.3 Visualization of synthetic data
204
+
205
+ To compare the synthetic data with the training data, we visualize the synthetic data on CIFAR10 and SVHN datasets in Figure 6. As shown in the figure, the first / third row is the original data of CIFAR10 / SVHN dataset, and the second / last row is the synthetic data generated by the model trained on CIFAR10 / SVHN dataset. The synthetic data are not similar to the original data, which can effectively reduce the probability of leaking sensitive information of clients. Note that although the synthetic data look much different from the original data, our method still achieves a higher performance than other baseline methods by training with these synthetic data (as shown in Table 1). Note that the ideal synthetic data should be visually distinct from the real data.
206
+
207
+ # 3.3.4 Extend to multiple rounds
208
+
209
+ We also extend DENSE to multi-round FL to test its effectiveness, i.e., there are multiple communication rounds between clients and server. Table 5 demonstrates the results of DENSE across different communication rounds $T _ { c } =$ $\{ 1 , 2 , 3 , 4 , 5 \}$ on CIFAR10 and SVHN datasets. The local training epoch is fixed as $E = 1 0$ . The performance of
210
+
211
+ Table 5: Accuracy for multiple communication rounds.
212
+
213
+ <table><tr><td>Dataset</td><td colspan="3">CIFAR10</td><td colspan="3">SVHN</td></tr><tr><td>Communication rounds</td><td>α=0.1</td><td>α=0.3</td><td>α=0.5</td><td>a=0.1</td><td>α=0.3</td><td>α=0.5</td></tr><tr><td>Tc=1</td><td>50.72</td><td>59.41</td><td>63.89</td><td>54.344</td><td>79.87</td><td>80.14</td></tr><tr><td>T=2</td><td>63.08</td><td>65.90</td><td>71.16</td><td>56.13</td><td>79.75</td><td>85.18</td></tr><tr><td>Te=3</td><td>61.61</td><td>69.73</td><td>73.91</td><td>74.41</td><td>86.42</td><td>86.18</td></tr><tr><td>T=4</td><td>66.26</td><td>69.40</td><td>74.39</td><td>78.67</td><td>86.36</td><td>86.43</td></tr><tr><td>T=5</td><td>67.65</td><td>71.42</td><td>76.01</td><td>80.28</td><td>86.25</td><td>86.55</td></tr></table>
214
+
215
+ DENSE improves as $T _ { c }$ increases, and DENSE achieves the best performance when $T _ { c } = 5$ . This shows that DENSE can be extended to multi-round FL and the performance can be further enhanced by increasing the communication rounds.
216
+
217
+ # 3.3.5 Contribution of $\mathcal { L } _ { B N }$ and $\mathcal { L } _ { d i v }$
218
+
219
+ We investigate the contributions of different loss functions used in data generation. We conduct leave-one-out testing and show the results by removing $\mathcal { L } _ { d i v }$ (w/o $\mathcal { L } _ { d i v , }$ ), and removing $\mathcal { L } _ { B N }$ (w/o $\mathcal { L } _ { B N } \mathrm { . }$ ). Additionally, we report the result by removing both $\mathcal { L } _ { d i v }$ and $\mathcal { L } _ { B N }$ , i.e., using only $\mathcal { L } _ { C E }$ (w/ $\mathcal { L } _ { C E } )$ ). As illustrated in Table 6, using only $\mathcal { L } _ { C E }$ to train the generator leads to poor performance. Be
220
+
221
+ Table 6: Impact of loss functions in data generation.
222
+
223
+ <table><tr><td>Dataset</td><td>CIFAR10</td><td>SVHN</td><td>CIFAR100</td></tr><tr><td>DENSE</td><td>62.19</td><td>80.03</td><td>42.07</td></tr><tr><td>wlLcE</td><td>53.12</td><td>73.11</td><td>36.47</td></tr><tr><td>W/o LBN</td><td>61.05</td><td>78.36</td><td>39.89</td></tr><tr><td>w/o Ldiv</td><td>59.18</td><td>77.59</td><td>39.14</td></tr></table>
224
+
225
+ sides, removing either the $\mathcal { L } _ { B N }$ loss or $\mathcal { L } _ { d i v }$ loss also affects the accuracy of the global model. A combination of these loss functions leads to a high performance of global model, which shows that each part of the loss function plays an important role in enhancing the generator.
226
+
227
+ # 4 Conclusion
228
+
229
+ In this paper, we propose an effective one-shot federated learning method called DENSE, which trains the global model by a data generation stage and a model distillation stage. Extensive experiments across various settings validate the efficacy of our method. Overall, DENSE is by far the most practical framework that can conduct data-free one-shot FL with model heterogeneity. A promising future direction is to consider the potential privacy attacks in one-shot FL.
230
+
231
+ # 5 Acknowledgement
232
+
233
+ This work was supported by Sony AI, the National Key Research and Development Project of China (2021ZD0110400 No. 2018AAA0101900), National Natural Science Foundation of China (U19B2042), Zhejiang Lab (2021KE0AC02), Academy Of Social Governance Zhejiang University, Fundamental Research Funds for the Central Universities.
234
+
235
+ References
236
+ [1] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss, 2019.
237
+ [2] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Dafl: Data-free learning of student networks. In ICCV, 2019.
238
+ [3] Hong-You Chen and Wei-Lun Chao. Fedbe: Making bayesian model ensemble applicable to federated learning, 2021.
239
+ [4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: Deformable patch attack. In European conference on computer vision, 2022.
240
+ [5] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15148–15158, June 2022.
241
+ [6] Zhen Chen, Meilu Zhu, Chen Yang, and Yixuan Yuan. Personalized retrogress-resilient framework for real-world medical federated learning. In Marleen de Bruijne, Philippe C. Cattin, Stéphane Cotin, Nicolas Padoy, Stefanie Speidel, Yefeng Zheng, and Caroline Essert, editors, Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part III, volume 12903 of Lecture Notes in Computer Science, pages 347–356. Springer, 2021.
242
+ [7] Don Kurian Dennis, Tian Li, and Virginia Smith. Heterogeneity for the win: One-shot federated clustering, 2021.
243
+ [8] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 11(1):1–17, 2021.
244
+ [9] Jiahua Dong, Yang Cong, Gan Sun, Bineng Zhong, and Xiaowei Xu. What can be transferred: Unsupervised domain adaptation for endoscopic lesions segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4022–4031, June 2020.
245
+ [10] Neel Guha, Ameet Talwalkar, and Virginia Smith. One-shot federated learning. arXiv preprint arXiv:1902.11175, 2019.
246
+ [11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015.
247
+ [12] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge distillation with adversarial samples supporting decision boundary. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3771–3778, 2019.
248
+ [13] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015.
249
+ [14] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015.
250
+ [15] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 603–618, 2017.
251
+ [16] Hao Huang, Yongtao Wang, Zhaoyu Chen, Yuze Zhang, Yuheng Li, Zhi Tang, Wei Chu, Jingdong Chen, Weisi Lin, and Kai-Kuang Ma. Cmua-watermark: A cross-model universal adversarial watermark for combating deepfakes. Proceedings of the AAAI Conference on Artificial Intelligence, 36(1):989–997, Jun. 2022.
252
+ [17] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift, 2015.
253
+
254
+ [18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning, pages 5132–5143. PMLR, 2020.
255
+
256
+ [19] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via majorto-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.
257
+
258
+ [20] Aashish Kolluri, Teodora Baluta, and Prateek Saxena. Private hierarchical clustering in federated networks. In Yongdae Kim, Jong Kim, Giovanni Vigna, and Elaine Shi, editors, CCS ’21: 2021 ACM SIGSAC Conference on Computer and Communications Security, Virtual Event, Republic of Korea, November 15 - 19, 2021, pages 2342–2360. ACM, 2021.
259
+
260
+ [21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.
261
+
262
+ [22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017.
263
+
264
+ [23] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015.
265
+
266
+ [24] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
267
+
268
+ [25] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., volume 2, pages II–104. IEEE, 2004.
269
+
270
+ [26] Bo Li, Zhengxing Sun, and Yuqi Guo. Supervae: Superpixelwise variational autoencoder for salient object detection. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 8569–8576. AAAI Press, 2019.
271
+
272
+ [27] Bo Li, Zhengxing Sun, Lv Tang, Yunhan Sun, and Jinlong Shi. Detecting robust co-saliency with recurrent co-attention neural network. In Sarit Kraus, editor, Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 818–825. ijcai.org, 2019.
273
+
274
+ [28] Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation. arXiv preprint arXiv:1910.03581, 2019.
275
+
276
+ [29] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study, 2021.
277
+
278
+ [30] Qinbin Li, Bingsheng He, and Dawn Song. Practical one-shot federated learning for cross-silo setting. arXiv preprint arXiv:2010.01017, 2020.
279
+
280
+ [31] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks, 2020.
281
+
282
+ [32] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations, 2020.
283
+
284
+ [33] Xiangru Lian, Ce Zhang, Huan Zhang, Cho-Jui Hsieh, Wei Zhang, and Ji Liu. Can decentralized algorithms outperform centralized algorithms? a case study for decentralized parallel stochastic gradient descent, 2017.
285
+
286
+ [34] Feng Liang, Weike Pan, and Zhong Ming. Fedrec++: Lossless federated recommendation with explicit feedback. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021.
287
+
288
+ [35] Tao Lin, Lingjing Kong, Sebastian U. Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020.
289
+
290
+ [36] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1013–1023. Computer Vision Foundation / IEEE, 2021.
291
+
292
+ [37] Ruixuan Liu, Yang Cao, Hong Chen, Ruoyang Guo, and Masatoshi Yoshikawa. FLAME: differentially private federated learning in the shuffle model. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 8688–8696. AAAI Press, 2021.
293
+
294
+ [38] Shuchang Liu, Shuyuan Xu, Wenhui Yu, Zuohui Fu, Yongfeng Zhang, and Amélie Marian. Fedct: Federated collaborative transfer for recommendation. In Fernando Diaz, Chirag Shah, Torsten Suel, Pablo Castells, Rosie Jones, and Tetsuya Sakai, editors, SIGIR ’21: The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, Canada, July 11-15, 2021, pages 716–725. ACM, 2021.
295
+
296
+ [39] Yi Liu, Jialiang Peng, JQ James, and Yi Wu. Ppgan: Privacy-preserving generative adversarial network. In 2019 IEEE 25Th international conference on parallel and distributed systems (ICPADS), pages 985–989. IEEE, 2019.
297
+
298
+ [40] Lingjuan Lyu, Han Yu, Xingjun Ma, Chen Chen, Lichao Sun, Jun Zhao, Qiang Yang, and Philip S Yu. Privacy and robustness in federated learning: Attacks and defenses. arXiv preprint arXiv:2012.06337, 2020.
299
+
300
+ [41] Lingjuan Lyu, Han Yu, Jun Zhao, and Qiang Yang. Threats to federated learning. In Federated Learning, pages 3–16. Springer, 2020.
301
+
302
+ [42] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data, 2017.
303
+
304
+ [43] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew $\mathrm { ~ Y ~ N ~ g ~ } _ { }$ Reading digits in natural images with unsupervised feature learning. 2011.
305
+
306
+ [44] Lichao Sun and Lingjuan Lyu. Federated model distillation with noise-free differential privacy. In IJCAI, 2021.
307
+
308
+ [45] Manasi Vartak. MODELDB: A system for machine learning model management. In 8th Biennial Conference on Innovative Data Systems Research, CIDR 2017, Chaminade, CA, USA, January 8-11, 2017, Online Proceedings. www.cidrdb.org, 2017.
309
+
310
+ [46] Chenghong Wang, Jieren Deng, Xianrui Meng, Yijue Wang, Ji Li, Fei Miao, Sanguthevar Rajasekaran, and Caiwen Ding. A secure and efficient federated learning framework for NLP. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih, editors, Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021.
311
+
312
+ [47] Derui Wang, Chaoran Li, Sheng Wen, Surya Nepal, and Yang Xiang. Man-in-the-middle attacks against machine learning classifiers via malicious generative models. IEEE Transactions on Dependable and Secure Computing, 2020.
313
+
314
+ [48] Hongyi Wang, Saurabh Agarwal, and Dimitris S. Papailiopoulos. Pufferfish: Communicationefficient models at no extra cost. CoRR, abs/2103.03936, 2021.
315
+
316
+ [49] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. arXiv preprint arXiv:2007.07481, 2020.
317
+
318
+ [50] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A. Efros. Dataset distillation, 2020.
319
+
320
+ [51] Kun Wei, Da Chen, Yuhong Li, Xu Yang, Cheng Deng, and Dacheng Tao. Incremental embedding learning with disentangled representation translation. IEEE Transactions on Neural Networks and Learning Systems, 2022.
321
+
322
+ [52] Kun Wei, Cheng Deng, Xu Yang, and Maosen Li. Incremental embedding learning via zero-shot translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10254–10262, 2021.
323
+
324
+ [53] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017.
325
+
326
+ [54] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16337–16346, 2021.
327
+
328
+ [55] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deepinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715–8724, 2020.
329
+
330
+ [56] Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Trong Nghia Hoang, and Yasaman Khazaeni. Bayesian nonparametric federated learning of neural networks, 2019.
331
+
332
+ [57] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks, 2017.
333
+
334
+ [58] Jie Zhang, Chen Chen, Jiahua Dong, Ruoxi Jia, and Lingjuan Lyu. Qekd: Query-efficient and data-free knowledge distillation from black-box models. arXiv preprint arXiv:2205.11158, 2022.
335
+
336
+ [59] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In International Conference on Machine Learning, pages 26311–26329. PMLR, 2022.
337
+
338
+ [60] Shaofeng Zhang, Meng Liu, and Junchi Yan. The diversified ensemble neural network. Advances in Neural Information Processing Systems, 33, 2020.
339
+
340
+ [61] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data, 2018.
341
+
342
+ [62] Yanlin Zhou, George Pu, Xiyao Ma, Xiaolin Li, and Dapeng Wu. Distilled one-shot federated learning. arXiv preprint arXiv:2009.07999, 2020.
343
+
344
+ [63] Xinghua Zhu, Jianzong Wang, Zhenhou Hong, and Jing Xiao. Empirical studies of institutional federated learning for natural language processing. In Trevor Cohn, Yulan He, and Yang Liu, editors, Findings of the Association for Computational Linguistics: EMNLP 2020, Online Event, 16-20 November 2020, volume EMNLP 2020 of Findings of ACL, pages 625–634. Association for Computational Linguistics, 2020.
345
+
346
+ [64] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 12878–12889. PMLR, 2021.
347
+
348
+ # Checklist
349
+
350
+ 1. For all authors...
351
+
352
+ (a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
353
+ (b) Did you describe the limitations of your work? [Yes]
354
+ (c) Did you discuss any potential negative societal impacts of your work? [No]
355
+ (d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
356
+
357
+ 2. If you are including theoretical results...
358
+
359
+ (a) Did you state the full set of assumptions of all theoretical results? [Yes] (b) Did you include complete proofs of all theoretical results? [Yes]
360
+
361
+ 3. If you ran experiments...
362
+
363
+ (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes]
364
+ (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes]
365
+ (c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [No]
366
+ (d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes]
367
+
368
+ 4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
369
+
370
+ (a) If your work uses existing assets, did you cite the creators? [Yes]
371
+ (b) Did you mention the license of the assets? [N/A]
372
+ (c) Did you include any new assets either in the supplemental material or as a URL? [No]
373
+ (d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
374
+ (e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A]
375
+
376
+ 5. If you used crowdsourcing or conducted research with human subjects...
377
+
378
+ (a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
379
+ (b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
380
+ (c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_content_list.json ADDED
@@ -0,0 +1,1866 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "DENSE: Data-Free One-Shot Federated Learning ",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 196,
8
+ 122,
9
+ 802,
10
+ 147
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Jie Zhang1∗ Chen Chen1∗ Bo Li2‡ Lingjuan Lyu3‡ Shuang Wu2 Shouhong Ding2 Chunhua Shen1 Chao Wu1‡ ",
17
+ "bbox": [
18
+ 274,
19
+ 199,
20
+ 717,
21
+ 231
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "1Zhejiang University 2Youtu Lab, Tencent 3 Sony AI ",
28
+ "bbox": [
29
+ 292,
30
+ 236,
31
+ 705,
32
+ 251
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "{zj_zhangjie, chao.wu, cc33} $@$ zju.edu.cn, Lingjuan.Lv $@$ sony.com {libraboli, calvinwu, ericshding}@tencent.com, chunhua $@$ me.com ",
39
+ "bbox": [
40
+ 299,
41
+ 258,
42
+ 697,
43
+ 286
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "Abstract ",
50
+ "text_level": 1,
51
+ "bbox": [
52
+ 462,
53
+ 329,
54
+ 535,
55
+ 345
56
+ ],
57
+ "page_idx": 0
58
+ },
59
+ {
60
+ "type": "text",
61
+ "text": "One-shot Federated Learning (FL) has recently emerged as a promising approach, which allows the central server to learn a model in a single communication round. Despite the low communication cost, existing one-shot FL methods are mostly impractical or face inherent limitations, e.g., a public dataset is required, clients’ models are homogeneous, and additional data/model information need to be uploaded. To overcome these issues, we propose a novel two-stage Data-freE oNeShot federated lEarning (DENSE) framework, which trains the global model by a data generation stage and a model distillation stage. DENSE is a practical one-shot FL method that can be applied in reality due to the following advantages: (1) DENSE requires no additional information compared with other methods (except the model parameters) to be transferred between clients and the server; (2) DENSE does not require any auxiliary dataset for training; (3) DENSE considers model heterogeneity in FL, i.e., different clients can have different model architectures. Experiments on a variety of real-world datasets demonstrate the superiority of our method. For example, DENSE outperforms the best baseline method Fed-ADI by $5 . 0 8 \\%$ on CIFAR10 dataset. ",
62
+ "bbox": [
63
+ 232,
64
+ 361,
65
+ 766,
66
+ 580
67
+ ],
68
+ "page_idx": 0
69
+ },
70
+ {
71
+ "type": "text",
72
+ "text": "1 Introduction ",
73
+ "text_level": 1,
74
+ "bbox": [
75
+ 174,
76
+ 607,
77
+ 310,
78
+ 625
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "text",
84
+ "text": "Deep neural networks (DNNs) have recently gained popularity as a powerful tool for advancing artificial intelligence in both established and emerging fields [25, 22, 13, 9, 8, 51, 52, 58, 5, 4, 16]. Federated learning (FL) [42] has emerged as a promising learning paradigm which allows multiple clients to collaboratively train a global model without exposing their private training data. In FL, each client trains a local model on its own data and is required to periodically share its high-dimensional model parameters with a central server. Recent years, FL has shown its potential to facilitate real-world applications in many fields, including medical image analysis [36, 6], recommender systems [34, 38], natural language processing [63, 46] and computer vision [27, 26]. ",
85
+ "bbox": [
86
+ 174,
87
+ 638,
88
+ 825,
89
+ 750
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "text",
95
+ "text": "The original FL framework requires participants to communicate frequently with the central server in order to exchange models. In real-world FL, such high communication cost may be intolerable and impractical. Reducing the communication cost between clients and the server is desired both for system efficiency and to support the privacy goals of federated learning [40, 41]. Recent research proposed some common methods to reduce communication costs, e.g., utilize multiple local updates [18], employ compression techniques [48], and one-shot FL [10]. Among them, one-shot FL is a promising solution which only allows one communication round. There are several motivations behind one-shot FL: 1) First of all, multi-round training is not practical in some scenarios such as model markets [45], in which users can only buy the pre-trained models from the market without any real data. 2) Furthermore, frequent communication poses a high risk of being attacked. For instance, frequent communication can be easily intercepted by attackers, who can launch man-in-the-middle attacks [47] or even reconstruct the training data from gradients [54]. In this way, one-shot FL can reduce the probability of being intercepted by malicious attackers due to the one-round property. Thus, in this paper, we mainly focus on one-shot FL. ",
96
+ "bbox": [
97
+ 174,
98
+ 756,
99
+ 825,
100
+ 839
101
+ ],
102
+ "page_idx": 0
103
+ },
104
+ {
105
+ "type": "text",
106
+ "text": "",
107
+ "bbox": [
108
+ 174,
109
+ 90,
110
+ 825,
111
+ 202
112
+ ],
113
+ "page_idx": 1
114
+ },
115
+ {
116
+ "type": "text",
117
+ "text": "However, existing one-shot FL studies [10, 30, 62, 7] are still hard to apply in real-world applications, due to impractical settings. For example, Guha et al. [10] and Li et al. [30] involved a public dataset for training, which may be impractical in very sensitive scenarios such as the biomedical domains. Zhu et al. [62] adopted dataset distillation [50] in one-shot FL, but they need to send distilled data to the central server, which causes additional communication cost and potential privacy leakage. Dennis et al. [7] utilized cluster-based method in one-shot FL, which requires to upload the cluster means to the server, causing additional communication cost. Additionally, none of these methods consider model heterogeneity, i.e., different clients have different model architectures [31], which is very common in practical scenarios. For instance, in model market, models sold by different sellers are likely to be heterogeneous. Besides, when several medical institutions participate in FL, they may need to design their own model to meet distinct specifications. Therefore, developing a practical one-shot FL method is in urgent need. ",
118
+ "bbox": [
119
+ 174,
120
+ 208,
121
+ 826,
122
+ 375
123
+ ],
124
+ "page_idx": 1
125
+ },
126
+ {
127
+ "type": "text",
128
+ "text": "In this work, we propose a novel two-stage Data-freE oNe-Shot federated lEarning (DENSE) framework, which trains the global model by a data generation stage and a model distillation stage. In the first stage, we utilize the ensemble models (i.e., ensemble of local models uploaded by clients) to train a generator, which can generate synthetic data for training in the second stage. In the second stage, we distill the knowledge of the ensemble models to the global model. In contrast to traditional FL methods based on FedAvg [42], our method does not require averaging of model parameters, thus it can support heterogeneous models, i.e., clients can have different model architectures. In summary, our main contributions are summarized as follows: ",
129
+ "bbox": [
130
+ 173,
131
+ 381,
132
+ 825,
133
+ 491
134
+ ],
135
+ "page_idx": 1
136
+ },
137
+ {
138
+ "type": "text",
139
+ "text": "• We propose a novel data-free one-shot FL framework named DENSE, which consists of two stages. In the first stage, we train a generator that considers similarity, stability, and transferability at the same time. In the second stage, we use the ensemble models and the data generated by the generator to train a global model. \nThe setting of DENSE is practical in the following aspects. First, DENSE requires no additional information (except the model parameters) to be transferred between clients and the server; Second, DENSE does not require any auxiliary dataset for training; Third, DENSE considers model heterogeneity, i.e., different clients can have different model architectures. \nDENSE is a compatible approach, which can be combined with any local training techniques to further improve the performance of the global model. For instance, we can adopt LDAM [1] to train the clients’ local models, and improve the accuracy of the global model (refer to Section 2.3 and Section 3.2). Extensive experiments on various datasets verify the effectiveness of our proposed DENSE. For example, DENSE outperforms the best baseline method Fed-ADI [55] by $5 . 0 8 \\%$ on CIFAR10 dataset. ",
140
+ "bbox": [
141
+ 217,
142
+ 501,
143
+ 825,
144
+ 733
145
+ ],
146
+ "page_idx": 1
147
+ },
148
+ {
149
+ "type": "text",
150
+ "text": "2 Data-Free One-Shot Federated Learning ",
151
+ "text_level": 1,
152
+ "bbox": [
153
+ 174,
154
+ 751,
155
+ 545,
156
+ 768
157
+ ],
158
+ "page_idx": 1
159
+ },
160
+ {
161
+ "type": "text",
162
+ "text": "2.1 Framework Overview ",
163
+ "text_level": 1,
164
+ "bbox": [
165
+ 174,
166
+ 781,
167
+ 364,
168
+ 796
169
+ ],
170
+ "page_idx": 1
171
+ },
172
+ {
173
+ "type": "text",
174
+ "text": "To tackle the problems in recent one-shot FL methods as mentioned in Sec. 1, we propose a novel method named DENSE, which conducts one-shot FL without the need to share additional information or rely on any auxiliary dataset, while considering model heterogeneity. To simulate real-world applications, we consider a more challenging yet practical setting where the data on each client are not independent and identically distributed (non-IID). ",
175
+ "bbox": [
176
+ 174,
177
+ 806,
178
+ 825,
179
+ 877
180
+ ],
181
+ "page_idx": 1
182
+ },
183
+ {
184
+ "type": "text",
185
+ "text": "The illustration of the learning procedure is demonstrated in Figure 1, and the whole training process of DENSE is shown in Algorithm 1. After clients upload their local models to the server, the server trains a global model with DENSE in two stages. In the data generation stage (first stage), we train an auxiliary generator that can generate synthetic data by the ensemble models, i.e., ensemble of local models uploaded by clients. In the model distillation stage (second stage), we use the ensemble models and the synthetic data (generated by the generator) to train the global model. ",
186
+ "bbox": [
187
+ 173,
188
+ 883,
189
+ 821,
190
+ 911
191
+ ],
192
+ "page_idx": 1
193
+ },
194
+ {
195
+ "type": "image",
196
+ "img_path": "images/2ce9636880a058cd0713d603581fa7f3e6acfe162a440e0ad227547b6544a41c.jpg",
197
+ "image_caption": [
198
+ "Figure 1: An illustration of training process of DENSE on the server, which consists of two stages: (1) In data generation stage, we train an auxiliary generator that considers similarity, stability, and transferability at the same time; (2) In model distillation stage, we distill the knowledge of the ensemble models and transfer to the global model. Note that the fixed global model is used as an additional discriminator in the divergence loss $\\mathcal { L } _ { d i v }$ . "
199
+ ],
200
+ "image_footnote": [],
201
+ "bbox": [
202
+ 174,
203
+ 90,
204
+ 823,
205
+ 287
206
+ ],
207
+ "page_idx": 2
208
+ },
209
+ {
210
+ "type": "text",
211
+ "text": "",
212
+ "bbox": [
213
+ 174,
214
+ 398,
215
+ 825,
216
+ 455
217
+ ],
218
+ "page_idx": 2
219
+ },
220
+ {
221
+ "type": "text",
222
+ "text": "2.2 Data Generation ",
223
+ "text_level": 1,
224
+ "bbox": [
225
+ 174,
226
+ 468,
227
+ 330,
228
+ 482
229
+ ],
230
+ "page_idx": 2
231
+ },
232
+ {
233
+ "type": "text",
234
+ "text": "In the first stage, we aim to train a generator to generate synthetic data. Specifically, given the ensemble of well-trained models uploaded by clients, our goal is to train a generator that can generate data that have similar distribution to the training data of clients. In addition, we aim not to leak private information from our generated data, i.e., attackers are not able to predict any sensitive information of clients from the generated data. Recent work [35] generated data by utilizing a pre-trained generative adversarial network (GAN). However, such a method is unable to generate data as the pre-trained GAN is trained on public datasets, which is likely to have different data distribution from the training data of clients. Moreover, we need to consider model heterogeneity, which makes the problem more complicated. ",
235
+ "bbox": [
236
+ 173,
237
+ 487,
238
+ 825,
239
+ 612
240
+ ],
241
+ "page_idx": 2
242
+ },
243
+ {
244
+ "type": "text",
245
+ "text": "To solve these issues, we propose to train a generator that considers similarity\\*, stability, and transferability. The data generation process is shown in line 8 to 11 in Algorithm 1. In particular, given a random noise $\\mathbf { z }$ (generated from a standard Gaussian distribution) and a random one-hot label y (generated from a uniform distribution), the generator $G ( \\cdot )$ aims to generate a synthetic data $\\hat { \\mathbf { x } } = G ( \\mathbf { z } )$ such that $\\hat { \\bf x }$ is similar to the training data (with label $\\mathbf { y }$ ) of clients. ",
246
+ "bbox": [
247
+ 173,
248
+ 618,
249
+ 825,
250
+ 688
251
+ ],
252
+ "page_idx": 2
253
+ },
254
+ {
255
+ "type": "text",
256
+ "text": "Similarity. First, we need to consider the similarity between synthetic data $\\hat { \\bf x }$ and the training data. Since we are unable to access the training data of clients, we cannot compute the similarity between the synthetic data and the training data directly. Instead, we first compute the average logits (i.e., outputs of the last fully connected layer) of $\\hat { \\bf x }$ computed by the ensemble models. ",
257
+ "bbox": [
258
+ 173,
259
+ 698,
260
+ 826,
261
+ 755
262
+ ],
263
+ "page_idx": 2
264
+ },
265
+ {
266
+ "type": "equation",
267
+ "img_path": "images/6c71d9f5a4178d1b48f2d19bbb98de16ffa57a60378aaed0df0bb58c218884ba.jpg",
268
+ "text": "$$\nD ( \\hat { \\mathbf { x } } ; \\{ \\pmb { \\theta } ^ { k } \\} _ { k = 1 } ^ { m } ) = \\frac { 1 } { m } \\sum _ { k \\in \\mathcal { C } } f ^ { k } \\left( \\hat { \\mathbf { x } } ; \\pmb { \\theta } ^ { k } \\right) ,\n$$",
269
+ "text_format": "latex",
270
+ "bbox": [
271
+ 369,
272
+ 763,
273
+ 627,
274
+ 803
275
+ ],
276
+ "page_idx": 2
277
+ },
278
+ {
279
+ "type": "text",
280
+ "text": "where $m = | \\mathcal { C } |$ , and $D ( \\hat { \\mathbf { x } } ; \\{ \\pmb { \\theta } ^ { k } \\} _ { k = 1 } ^ { m } )$ is the average logits of $\\hat { \\bf x }$ , $\\pmb { \\theta } ^ { k }$ is the parameter of the $k$ -th client. \nAnd $f ^ { k } \\left( { \\hat { \\mathbf { x } } } ; \\pmb { \\theta } ^ { k } \\right)$ is the prediction function of client $k$ that outputs the logits of $\\hat { \\bf x }$ given parameter $\\theta ^ { k }$ . \nFor simplicity, we use $D ( \\hat { \\mathbf { x } } )$ to denote $D ( \\hat { \\mathbf { x } } ; \\{ \\pmb { \\theta } ^ { k } \\} _ { k = 1 } ^ { m } )$ in the rest of the paper. ",
281
+ "bbox": [
282
+ 173,
283
+ 813,
284
+ 826,
285
+ 871
286
+ ],
287
+ "page_idx": 2
288
+ },
289
+ {
290
+ "type": "text",
291
+ "text": "Then, we minimize the average logits and the random label $y$ with the following cross-entropy (CE) loss. ",
292
+ "bbox": [
293
+ 169,
294
+ 90,
295
+ 825,
296
+ 118
297
+ ],
298
+ "page_idx": 3
299
+ },
300
+ {
301
+ "type": "equation",
302
+ "img_path": "images/b9a7f312faf7e4a780e52e674aa461846742d3a0e70f090e75b0124546d4c1e1.jpg",
303
+ "text": "$$\n\\mathcal { L } _ { C E } ( \\hat { \\mathbf { x } } , \\mathbf { y } ; \\pmb { \\theta } _ { G } ) = C E ( D ( \\hat { \\mathbf { x } } ) , \\mathbf { y } ) ,\n$$",
304
+ "text_format": "latex",
305
+ "bbox": [
306
+ 387,
307
+ 117,
308
+ 609,
309
+ 135
310
+ ],
311
+ "page_idx": 3
312
+ },
313
+ {
314
+ "type": "text",
315
+ "text": "It is expected that the synthetic images can be classified into one particular class with a high probability by the ensemble models. In fact, during the training phase, the loss between $D ( \\hat { \\mathbf { x } } )$ and $\\mathbf { y }$ can easily reduce to almost 0, which indicates the synthetic data matches the ensemble models perfectly. Moreover, we do not directly compute the similarity between the synthetic data and the training data, which can reduce the probability of leaking sensitive information of the clients. ",
316
+ "bbox": [
317
+ 173,
318
+ 137,
319
+ 825,
320
+ 208
321
+ ],
322
+ "page_idx": 3
323
+ },
324
+ {
325
+ "type": "text",
326
+ "text": "However, by utilizing only the CE loss, we cannot achieve a high performance (please refer to Section 3.2 for detail). We conjecture this is because the ensemble models are trained on non-IID data, the generator may be unstable and trapped into sub-optimal local minima or overfit to the synthetic data [49, 32]. ",
327
+ "bbox": [
328
+ 174,
329
+ 213,
330
+ 825,
331
+ 268
332
+ ],
333
+ "page_idx": 3
334
+ },
335
+ {
336
+ "type": "text",
337
+ "text": "Stability. Second, to improve the stability of the generator, we propose to add an additional regularization to stabilize the training. In particular, we utilize the Batch Normalization (BN) loss to make the synthetic data conform with the batch normalization statistics [55]. ",
338
+ "bbox": [
339
+ 173,
340
+ 276,
341
+ 823,
342
+ 319
343
+ ],
344
+ "page_idx": 3
345
+ },
346
+ {
347
+ "type": "equation",
348
+ "img_path": "images/de28f63560d2cb651e1066d0a90842918e0e047cf416964e25375636f1fc64fa.jpg",
349
+ "text": "$$\n\\mathcal { L } _ { B N } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { G } ) = \\frac { 1 } { m } \\sum _ { k \\in \\mathcal { C } } \\sum _ { l } \\left( \\lVert \\mu _ { l } ( \\hat { \\mathbf { x } } ) - \\mu _ { k , l } \\rVert + \\lVert \\sigma _ { l } ^ { 2 } ( \\hat { \\mathbf { x } } ) - \\sigma _ { k , l } ^ { 2 } \\rVert \\right) ,\n$$",
350
+ "text_format": "latex",
351
+ "bbox": [
352
+ 282,
353
+ 324,
354
+ 714,
355
+ 363
356
+ ],
357
+ "page_idx": 3
358
+ },
359
+ {
360
+ "type": "text",
361
+ "text": "where $\\mu _ { l } ( \\hat { \\mathbf { x } } )$ and $\\sigma _ { l } ^ { 2 } ( \\hat { \\mathbf { x } } )$ are the batch-wise mean and variance estimates of feature maps corresponding to the $l$ -th BN layer of the generator $G ( \\cdot ) ^ { \\dagger }$ , $\\mu _ { k , l }$ and $\\sigma _ { k , l } ^ { 2 }$ are the mean and variance of the $l$ -th BN layer [17] of $f ^ { k } ( \\cdot )$ . The BN loss minimizes the distance between the feature map statistics of the synthetic data and the training data of clients. As a result, the synthetic data can have a similar distribution to the training data of clients, no matter if the data is non-IID or IID. ",
362
+ "bbox": [
363
+ 173,
364
+ 369,
365
+ 825,
366
+ 445
367
+ ],
368
+ "page_idx": 3
369
+ },
370
+ {
371
+ "type": "text",
372
+ "text": "Transferability. By utilizing the CE loss and BN loss, we can train a generator that can generate synthetic data, but we observed that the synthetic data are likely to be far away from the decision boundary (of the ensemble models), which makes the ensemble models (teachers) hard to transfer their knowledge to the global model (student). We illustrate the observation in the left panel of Figure 2. S and T are the decision boundaries of the global model (the detail of the global model is introduced in Section 2.3) and ensemble models respectively. The essence of knowledge distillation is transferring the information of decision boundary from the teacher model to the student model [12]. We aim to learn the decision boundary of global model and have a high classification accuracy on the real test data (blue diamonds). However, the generated synthetic data (red circles) are likely to be on the same side of the two decision boundaries and unhelpful to the transfer of knowledge [12]. To solve this problem, we ",
373
+ "bbox": [
374
+ 174,
375
+ 453,
376
+ 483,
377
+ 757
378
+ ],
379
+ "page_idx": 3
380
+ },
381
+ {
382
+ "type": "image",
383
+ "img_path": "images/76d9a9a4dd165c97d9c6b81648bf69bc02c6b45b06c64e217312f1914f3b33a5.jpg",
384
+ "image_caption": [
385
+ "Figure 2: The illustration of generated data and decision boundary of ensemble models (teachers) and global model (student). Left panel: Synthetic data (red circles) are far away from the decision boundary, which is less helpful to the transfer of knowledge. Right panel: By utilizing our boundary support loss, we can generate more synthetic data near the decision boundaries (black circles), which helps the student better learn the decision boundary of the teacher. "
386
+ ],
387
+ "image_footnote": [],
388
+ "bbox": [
389
+ 532,
390
+ 472,
391
+ 785,
392
+ 582
393
+ ],
394
+ "page_idx": 3
395
+ },
396
+ {
397
+ "type": "text",
398
+ "text": "argue to generate more synthetic data that fall between the decision boundaries of the ensemble models and the global model. We illustrate our idea in the right panel of Figure 2. Red circles are synthetic data on the same side of the decision boundary, which are less helpful in learning the global model. Black circles are synthetic data between the decision boundaries, i.e., the global model and the ensemble models have different predictions on these data. Black circles can help the global model better learn the decision boundary of the ensemble models. ",
399
+ "bbox": [
400
+ 174,
401
+ 757,
402
+ 825,
403
+ 838
404
+ ],
405
+ "page_idx": 3
406
+ },
407
+ {
408
+ "type": "text",
409
+ "text": "Motivated by the above observations, we introduce a new boundary support loss, which urges the generator to generate more synthetic data between the decision boundaries of the ensemble models and the global model. We divide the synthetic data into 2 sets: (1) the global model and the ensemble models have the same predictions on data in the first set $\\begin{array} { r } { \\arg \\operatorname* { m a x } _ { c } D ^ { ( c ) } ( \\hat { \\mathbf { x } } ) = \\arg \\operatorname* { m a x } _ { c } f _ { S } ^ { ( c ) } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { S } ) ) . } \\end{array}$ ; (2) different predictions on data in the second set ${ ( \\arg \\operatorname* { m a x } _ { c } D ^ { ( c ) } ( \\hat { \\bf x } ) \\neq \\arg \\operatorname* { m a x } _ { c } f _ { S } ^ { ( c ) } ( \\hat { \\bf x } ; \\pmb { \\theta } _ { S } ) ) }$ , where $D ^ { ( c ) } ( \\hat { \\mathbf { x } } )$ and $f _ { S } ^ { ( c ) } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { S } )$ are the logits for the $c$ -th label of the ensemble models and the global model respectively. The data in the first set are on the same side of those two decision boundaries (red circles in Figure 2) while the data in the second set (black circles in Figure 2) are between the decision boundaries of the ensemble models and the global model. We maximize the differences of predictions of the global model and the ensemble models on data in the second set with Kullback-Leibler divergence loss as follows. ",
410
+ "bbox": [
411
+ 176,
412
+ 845,
413
+ 825,
414
+ 887
415
+ ],
416
+ "page_idx": 3
417
+ },
418
+ {
419
+ "type": "text",
420
+ "text": "",
421
+ "bbox": [
422
+ 173,
423
+ 89,
424
+ 825,
425
+ 212
426
+ ],
427
+ "page_idx": 4
428
+ },
429
+ {
430
+ "type": "equation",
431
+ "img_path": "images/520285b0f14a26c736e532c26652dfd9fc8275befa8c250aad7dda976e95bca6.jpg",
432
+ "text": "$$\n\\mathcal { L } _ { d i v } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { G } ) = - \\omega K L \\left( D ( \\hat { \\mathbf { x } } ) , f _ { S } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { S } ) \\right) ,\n$$",
433
+ "text_format": "latex",
434
+ "bbox": [
435
+ 354,
436
+ 218,
437
+ 642,
438
+ 236
439
+ ],
440
+ "page_idx": 4
441
+ },
442
+ {
443
+ "type": "text",
444
+ "text": "where $K L ( \\cdot , \\cdot )$ denotes the Kullback-Leibler (KL) divergence loss, $\\omega = \\mathbb { 1 } ( \\arg \\operatorname* { m a x } _ { c } D ^ { ( c ) } ( \\hat { \\mathbf { x } } ) \\neq$ arg $\\operatorname* { m a x } _ { c } f _ { S } ^ { ( c ) } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { S } ) )$ outputs 0 for data in the first set and 1 for data in the second set, and $\\mathbb { 1 } ( a )$ is the indicator function that outputs 1 if $a$ is true and outputs 0 if $a$ is false. By maximizing the KL divergence loss, the generator can generate more synthetic data that are more helpful to the model distillation stage (refer to Section 2.3 for detail) and further improve the transferability of the ensemble models. ",
445
+ "bbox": [
446
+ 173,
447
+ 244,
448
+ 825,
449
+ 333
450
+ ],
451
+ "page_idx": 4
452
+ },
453
+ {
454
+ "type": "text",
455
+ "text": "By combining the above losses, we can obtain the generator loss as follows, ",
456
+ "bbox": [
457
+ 176,
458
+ 339,
459
+ 671,
460
+ 354
461
+ ],
462
+ "page_idx": 4
463
+ },
464
+ {
465
+ "type": "equation",
466
+ "img_path": "images/2b2e84d45592acc7d50816d7c7dbfcba9119ec8fa3c317113edc2160d6ff6eef.jpg",
467
+ "text": "$$\n\\begin{array} { r } { \\mathcal { L } _ { g e n } ( \\hat { \\mathbf { x } } , \\mathbf { y } ; \\pmb { \\theta } _ { G } ) = \\mathcal { L } _ { C E } ( \\hat { \\mathbf { x } } , \\mathbf { y } ; \\pmb { \\theta } _ { G } ) + \\lambda _ { 1 } \\mathcal { L } _ { B N } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { G } ) + \\lambda _ { 2 } \\mathcal { L } _ { d i v } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { G } ) , } \\end{array}\n$$",
468
+ "text_format": "latex",
469
+ "bbox": [
470
+ 263,
471
+ 363,
472
+ 733,
473
+ 381
474
+ ],
475
+ "page_idx": 4
476
+ },
477
+ {
478
+ "type": "text",
479
+ "text": "where $\\lambda _ { 1 }$ and $\\lambda _ { 2 }$ are scaling factors for the losses. ",
480
+ "bbox": [
481
+ 173,
482
+ 388,
483
+ 501,
484
+ 405
485
+ ],
486
+ "page_idx": 4
487
+ },
488
+ {
489
+ "type": "text",
490
+ "text": "Algorithm 1 Training process of DENSE ",
491
+ "text_level": 1,
492
+ "bbox": [
493
+ 176,
494
+ 420,
495
+ 446,
496
+ 435
497
+ ],
498
+ "page_idx": 4
499
+ },
500
+ {
501
+ "type": "text",
502
+ "text": "Input: Number of client $m$ , clients’ local models $\\{ f ^ { 1 } ( ) , \\cdots , f ^ { m } ( ) \\}$ , generator $G ( \\cdot )$ with parameter $\\theta _ { G }$ , learning rate of the generator $\\eta _ { G }$ , number of training rounds $T _ { G }$ for generator in each epoch, global model $f _ { S } ( \\ u )$ with parameter $\\theta _ { S }$ , learning rate of the global model $\\eta _ { S }$ , global model training epochs $T$ , and batch size $b$ . ",
503
+ "bbox": [
504
+ 174,
505
+ 438,
506
+ 825,
507
+ 493
508
+ ],
509
+ "page_idx": 4
510
+ },
511
+ {
512
+ "type": "text",
513
+ "text": "for each client $k \\in { \\mathcal { C } }$ in parallel do $\\pmb { \\theta } ^ { k } \\mathbf { L o c a l U p d a t e } ( k )$ \nend for \nInitialize parameter $\\theta _ { G }$ and $\\theta _ { S }$ \nfor epoch $\\scriptstyle 1 = 1 , \\cdots , T$ do Sample a batch of noises and labels $\\{ \\mathbf { z } _ { i } , \\mathbf { y } _ { i } \\} _ { i = 1 } ^ { b }$ // data generation stage for $j = 1 , \\cdots , T _ { G }$ do Generate $\\{ \\hat { \\mathbf { x } } _ { i } \\} _ { i = 1 } ^ { b }$ with $\\{ { \\mathbf { z } } _ { i } \\} _ { i = 1 } ^ { b }$ and $G ( \\cdot )$ enc $\\begin{array} { r } { \\pmb { \\theta } _ { G } \\pmb { \\theta } _ { G } - \\eta _ { G } \\frac { 1 } { b } \\sum _ { i = 1 } ^ { b } \\nabla _ { \\pmb { \\theta } _ { G } } \\ell _ { g e n } ( \\hat { \\mathbf { x } } _ { i } , \\mathbf { y } _ { i } ; \\pmb { \\theta } _ { G } ) } \\end{array}$ $/ /$ model distillation stage Generate $\\{ \\hat { \\mathbf { x } } _ { i } \\} _ { i = 1 } ^ { b }$ with $\\mathbf { \\bar { \\{ z } } _ { i } \\} _ { i = 1 } ^ { b }$ and $G ( \\cdot )$ $\\begin{array} { r } { \\pmb { \\theta } _ { S } = \\pmb { \\theta } _ { S } - \\eta _ { S } \\frac { 1 } { b } \\sum _ { i = 1 } ^ { b } \\nabla _ { \\pmb { \\theta } _ { S } } \\ell _ { d i s } ( \\hat { \\bf x } _ { i } ; \\pmb { \\theta } _ { S } ) } \\end{array}$ \nend for ",
514
+ "bbox": [
515
+ 194,
516
+ 496,
517
+ 553,
518
+ 713
519
+ ],
520
+ "page_idx": 4
521
+ },
522
+ {
523
+ "type": "text",
524
+ "text": "Note that the generated synthetic data have similar features but different from the training data (of clients), which reduces the probability of leaking sensitive information of clients. More discussions of the privacy issues are in Section 3.3.3. ",
525
+ "bbox": [
526
+ 176,
527
+ 729,
528
+ 823,
529
+ 772
530
+ ],
531
+ "page_idx": 4
532
+ },
533
+ {
534
+ "type": "text",
535
+ "text": "2.3 Model Distillation ",
536
+ "text_level": 1,
537
+ "bbox": [
538
+ 174,
539
+ 781,
540
+ 338,
541
+ 796
542
+ ],
543
+ "page_idx": 4
544
+ },
545
+ {
546
+ "type": "text",
547
+ "text": "In the second stage, we train the global model with the generator (discussed in the previous section) and the ensemble models. Previous research [60, 35] showed that model ensemble provides a general method for improving the accuracy and stability of learning models. Motivated by [60], we propose to use the ensemble models as a teacher to train a student (global) model. A straightforward method is to obtain the global model by aggregating the parameters of all client models (e.g., by FedAvg [42]). However, in real-world applications, clients are likely to have different model architectures [44], making FedAvg useless. Moreover, since the data in different clients are non-IID, FedAvg cannot deliver a good performance or even diverge [59, 32]. ",
548
+ "bbox": [
549
+ 173,
550
+ 800,
551
+ 826,
552
+ 911
553
+ ],
554
+ "page_idx": 4
555
+ },
556
+ {
557
+ "type": "text",
558
+ "text": "To this end, we follow [35] to distill the knowledge of the ensemble models to the global model by minimizing the predictions between the ensemble models (teacher) and the global model (student) on the same synthetic data. The model distillation process is shown in line 13 to 14 in Algorithm 1. First, we compute the average logits of the synthetic data according to Eq. (1), i.e., $D ( \\hat { \\mathbf { x } } ) =$ $\\begin{array} { r } { \\frac { 1 } { m } \\sum _ { k \\in \\mathcal { C } } f ^ { k } \\left( \\hat { \\mathbf { x } } ; \\pmb { \\theta } ^ { k } \\right) } \\end{array}$ . In contrast to traditional aggregation methods (e.g., FedAvg) that are unable to aggregate heterogeneous models, averaging logits can be easily applied to both heterogeneous and homogeneous FL systems. ",
559
+ "bbox": [
560
+ 173,
561
+ 90,
562
+ 825,
563
+ 198
564
+ ],
565
+ "page_idx": 5
566
+ },
567
+ {
568
+ "type": "text",
569
+ "text": "Then, we use the average logits to distill the knowledge of the ensemble models by minimizing the following objective function. ",
570
+ "bbox": [
571
+ 173,
572
+ 203,
573
+ 823,
574
+ 232
575
+ ],
576
+ "page_idx": 5
577
+ },
578
+ {
579
+ "type": "equation",
580
+ "img_path": "images/c963de924077ad4d47b94f877b9169c34bb7f0e549bd8ea7f5d0b06ddaa6a9a1.jpg",
581
+ "text": "$$\n\\mathcal { L } _ { d i s } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { S } ) = K L \\left( D ( \\hat { \\mathbf { x } } ) , f _ { S } ( \\hat { \\mathbf { x } } ; \\pmb { \\theta } _ { S } ) \\right) .\n$$",
582
+ "text_format": "latex",
583
+ "bbox": [
584
+ 367,
585
+ 234,
586
+ 630,
587
+ 251
588
+ ],
589
+ "page_idx": 5
590
+ },
591
+ {
592
+ "type": "text",
593
+ "text": "By minimizing the KL loss, we can train a global model with the knowledge of the ensemble models and the synthetic data regardless of data and model heterogeneity. ",
594
+ "bbox": [
595
+ 173,
596
+ 252,
597
+ 823,
598
+ 280
599
+ ],
600
+ "page_idx": 5
601
+ },
602
+ {
603
+ "type": "text",
604
+ "text": "Note that DENSE has no restriction on the clients’ local models, i.e., clients can train models with arbitrary techniques. Thus, DENSE is a compatible approach, which can be combined with any local training techniques to further improve the performance of the global model. We further discuss the combination of local training techniques in Section 3.2. ",
605
+ "bbox": [
606
+ 174,
607
+ 286,
608
+ 825,
609
+ 342
610
+ ],
611
+ "page_idx": 5
612
+ },
613
+ {
614
+ "type": "text",
615
+ "text": "Discussions on privacy-preserving Research [15] has shown that it is possible to launch an attack where a malicious user uses GANs to recreate samples of another participant’s private datasets. Besides, in FL, exchanging models between the server and clients can result in potential privacy leakage. Note that our method prohibits the generator from seeing the real data directly, and there is only one communication round, which reduces the risk of privacy leakage. In addition, we display our generated images in Section 3.3.3, which does not directly reveal the information of real data. Several existing privacy-preserving methods can be incorporated into our framework to better protect clients from adversaries [37, 20]. We leave this as our future work. ",
616
+ "bbox": [
617
+ 174,
618
+ 348,
619
+ 825,
620
+ 459
621
+ ],
622
+ "page_idx": 5
623
+ },
624
+ {
625
+ "type": "text",
626
+ "text": "Discussions on Knowledge Distillation in FL In traditional FL frameworks, all users have to agree on the specific architecture of the global model. To support model heterogeneity, Li et al. [28] proposed a new federated learning framework that enables participants to independently design their models by knowledge distillation [14]. With the use of a proxy dataset, knowledge distillation alleviates the model drift issue induced by non-IID data. However, the requirement of proxy data renders such a method impractical for many applications, since a carefully designed dataset is not always available on the server. Data-free knowledge distillation is a promising approach, which can transfer knowledge of a teacher model to a student model without any real data [2, 55]. Lin et al. [35] proposed data-free ensemble distillation for model fusion through synthetic data in each communication round, which requires high communication costs and computational costs. However, in this paper, we are more concerned with obtaining a good global model through only one round of communication in cases of heterogeneous models, which is more challenging and practical. Zhu et al. [64] also proposed a data-free knowledge distillation approach for FL, which learns a generator derived from the prediction of local models. However, the learned generator is later broadcasted to all clients, and then clients need to send their generators to the server, which increases the communication burden. More seriously, the generator has direct access to the local data (the generator can easily remember the training samples [39]), which can cause privacy concerns. As the generator used in our method is always stored in the central server, it never sees any real local data. ",
627
+ "bbox": [
628
+ 173,
629
+ 464,
630
+ 825,
631
+ 714
632
+ ],
633
+ "page_idx": 5
634
+ },
635
+ {
636
+ "type": "text",
637
+ "text": "3 Experiments ",
638
+ "text_level": 1,
639
+ "bbox": [
640
+ 174,
641
+ 726,
642
+ 312,
643
+ 742
644
+ ],
645
+ "page_idx": 5
646
+ },
647
+ {
648
+ "type": "text",
649
+ "text": "3.1 Experimental Setup ",
650
+ "text_level": 1,
651
+ "bbox": [
652
+ 174,
653
+ 748,
654
+ 351,
655
+ 763
656
+ ],
657
+ "page_idx": 5
658
+ },
659
+ {
660
+ "type": "text",
661
+ "text": "3.1.1 Datasets ",
662
+ "text_level": 1,
663
+ "bbox": [
664
+ 174,
665
+ 767,
666
+ 284,
667
+ 781
668
+ ],
669
+ "page_idx": 5
670
+ },
671
+ {
672
+ "type": "text",
673
+ "text": "Our experiments are conducted on the following 6 real-world datasets: MNIST [24], FMNIST [53], SVHN [43], CIFAR10 [21], CIFAR100 [21], and Tiny-ImageNet [23].MNIST dataset contains binary images of handwritten digits. There are 60,000 training images and 10,000 testing images in MNIST dataset. CIFAR10 dataset consists of $6 0 , 0 0 0 \\ 3 2 \\mathrm { x } \\ 3 2$ color images in 10 classes, with 6,000 images per class. There are 50,000 training images and 10,000 test images in CIFAR10 dataset. CIFAR100 dataset is similar to CIFAR10 dataset, except it has 100 classes containing 600 images each. There are 500 training images and 100 testing images per class. Tiny-ImageNet contains 100000 images of 200 classes (500 for each class) downsized to $6 4 { \\times } 6 4$ colored images. Each class has 500 training images, 50 validation images and 50 test images. ",
674
+ "bbox": [
675
+ 174,
676
+ 786,
677
+ 825,
678
+ 911
679
+ ],
680
+ "page_idx": 5
681
+ },
682
+ {
683
+ "type": "table",
684
+ "img_path": "images/bc06aa78c13f13ab8794fd3beec2575b3d720ca22c180a49252fd14c96d37d5f.jpg",
685
+ "table_caption": [
686
+ "Table 1: Accuracy of different methods across $\\alpha = \\{ 0 . 1 , 0 . 3 , 0 . 5 \\}$ on different datasets. "
687
+ ],
688
+ "table_footnote": [],
689
+ "table_body": "<table><tr><td>Dataset</td><td colspan=\"3\">MNIST</td><td colspan=\"3\">FMNIST</td><td colspan=\"3\">CIFAR10</td><td colspan=\"3\">SVHN</td><td colspan=\"3\">CIFAR100</td><td colspan=\"3\">Tiny-ImageNet</td></tr><tr><td>Method</td><td>a=0.1</td><td>Q=0.3</td><td>a=0.5</td><td>α=0.1</td><td>Q=0.3</td><td>Q=0.5</td><td>α=0.1</td><td>a=0.3</td><td>α=0.5</td><td>α=0.1</td><td>Q=0.3</td><td>α=0.5</td><td>Q=0.1</td><td>a=0.3</td><td>α=0.5</td><td></td><td>a=0.1 a=0.3</td><td>a=0.5</td></tr><tr><td>FedAvg</td><td>48.24</td><td>72.94</td><td>90.55</td><td>41.69</td><td>82.96</td><td>83.72</td><td>23.93</td><td>27.72</td><td>43.67</td><td>31.65</td><td>61.51</td><td>56.09</td><td>4.58</td><td>11.61</td><td>12.11</td><td>3.12</td><td>10.46</td><td>11.89</td></tr><tr><td>FedDF</td><td>60.15</td><td>74.01</td><td>92.18</td><td>43.58</td><td>80.67</td><td>84.67</td><td>40.58</td><td>46.78</td><td>53.56</td><td>49.13</td><td>73.34</td><td>73.98</td><td>28.17</td><td>30.28</td><td>36.35</td><td>15.34</td><td>18.22</td><td>27.43</td></tr><tr><td>Fed-DAFL</td><td>64.38</td><td>74.18</td><td>93.01</td><td>47.14</td><td>80.59</td><td>84.02</td><td>47.34</td><td>53.89</td><td>58.59</td><td>53.23</td><td>76.56</td><td>78.03</td><td>28.89</td><td>34.89</td><td>38.19</td><td>18.38</td><td>22.18</td><td>28.22</td></tr><tr><td>Fed-ADI</td><td>64.13</td><td>75.03</td><td>93.49</td><td>48.49</td><td>81.15</td><td>84.19</td><td>48.59</td><td>54.68</td><td>59.34</td><td>53.45</td><td>77.45</td><td>78.85</td><td>30.13</td><td>35.18</td><td>40.28</td><td>19.59</td><td>25.34</td><td>30.21</td></tr><tr><td>DENSE (ours)</td><td>66.61</td><td>76.48</td><td>95.82</td><td>50.29</td><td>83.96</td><td>85.94</td><td>50.26</td><td>59.76</td><td>62.19</td><td>55.34</td><td>79.59</td><td>80.03</td><td>32.03</td><td>37.32</td><td>42.07</td><td>22.44</td><td>28.14</td><td>32.34</td></tr></table>",
690
+ "bbox": [
691
+ 176,
692
+ 112,
693
+ 831,
694
+ 181
695
+ ],
696
+ "page_idx": 6
697
+ },
698
+ {
699
+ "type": "text",
700
+ "text": "3.1.2 Data partition ",
701
+ "text_level": 1,
702
+ "bbox": [
703
+ 174,
704
+ 210,
705
+ 325,
706
+ 224
707
+ ],
708
+ "page_idx": 6
709
+ },
710
+ {
711
+ "type": "text",
712
+ "text": "To simulate real-world applications, we use Dirichlet distribution to generate non-IID data partition among clients [56, 29]. In particular, we sample $p _ { k } \\sim D i r ( \\alpha )$ and allocate a $p _ { k } ^ { i }$ proportion of the data of class $k$ to client $i$ . By varying the parameter $\\alpha$ , we can change the degree of imbalance. A small $\\alpha$ generates highly skewed data. We set $\\alpha = 0 . 5$ as default. ",
713
+ "bbox": [
714
+ 174,
715
+ 234,
716
+ 825,
717
+ 290
718
+ ],
719
+ "page_idx": 6
720
+ },
721
+ {
722
+ "type": "text",
723
+ "text": "3.1.3 Baselines ",
724
+ "text_level": 1,
725
+ "bbox": [
726
+ 174,
727
+ 306,
728
+ 289,
729
+ 320
730
+ ],
731
+ "page_idx": 6
732
+ },
733
+ {
734
+ "type": "text",
735
+ "text": "To ensure fair comparisons, we neglect the comparison with methods that require to download auxiliary models or datasets, such as FedBE [3] and FedGen [64]. Moreover, since there is only one communication round, aggregation methods that are based on regularization have no effect. Thus, we also omit the comparison with these regularization-based methods, e.g., FedProx [31], FedNova [49], and Scaffold [18]. Instead, we compare our proposed DENSE with FedAvg [42] and FedDF [35]. Furthermore, since DENSE is a data-free method, we derive some baselines from prevailing data-free knowledge distillation methods, including: 1) DAFL [2], a novel data-free learning framework based on generative adversarial networks; 2) ADI [55], an image synthesizing method that utilizes the image distribution to train a deep neural network without real data. We apply these methods to one-shot FL, and name these two baselines as Fed-DAFL and Fed-ADI. ",
736
+ "bbox": [
737
+ 173,
738
+ 332,
739
+ 826,
740
+ 469
741
+ ],
742
+ "page_idx": 6
743
+ },
744
+ {
745
+ "type": "text",
746
+ "text": "3.1.4 Settings ",
747
+ "text_level": 1,
748
+ "bbox": [
749
+ 174,
750
+ 486,
751
+ 281,
752
+ 501
753
+ ],
754
+ "page_idx": 6
755
+ },
756
+ {
757
+ "type": "text",
758
+ "text": "For clients’ local training, we use the SGD optimizer with momentum $= 0 . 9$ and learning rate ${ \\mathrm { \\Omega } } { = } 0 . 0 1$ . We set the batch size $b = 1 2 8$ , the number of local epochs $E = 2 0 0$ , and the client number $m = 5$ . Following the setting of [2], we train the auxiliary generator $G ( \\cdot )$ with a deep convolutional network. We use Adam optimizer with learning rate $\\eta _ { G } = 0 . 0 0 1$ . We set the number of training rounds in each epoch as $T _ { G } = 3 0$ , and set the scaling factor $\\lambda _ { 1 } = 1$ and $\\lambda _ { 2 } = 0 . 5$ . For the training of the server model $f _ { S } ( \\ v r )$ , we use the SGD optimizer with learning rate $\\eta _ { S } = 0 . 0 1$ and momentum $\\scriptstyle 1 = 0 . 9$ . The number of epochs for distillation $T = 2 0 0$ . All baseline methods use the same setting as ours. ",
759
+ "bbox": [
760
+ 173,
761
+ 510,
762
+ 826,
763
+ 608
764
+ ],
765
+ "page_idx": 6
766
+ },
767
+ {
768
+ "type": "text",
769
+ "text": "3.2 Results ",
770
+ "text_level": 1,
771
+ "bbox": [
772
+ 174,
773
+ 626,
774
+ 264,
775
+ 640
776
+ ],
777
+ "page_idx": 6
778
+ },
779
+ {
780
+ "type": "text",
781
+ "text": "3.2.1 Evaluation on real-world datasets ",
782
+ "text_level": 1,
783
+ "bbox": [
784
+ 176,
785
+ 652,
786
+ 459,
787
+ 666
788
+ ],
789
+ "page_idx": 6
790
+ },
791
+ {
792
+ "type": "text",
793
+ "text": "To evaluate the effectiveness of our method, we conduct experiments under different non-IID settings by varying $\\alpha = \\{ 0 . 1 , 0 . 3 , 0 . 5 \\}$ and report the performance on different datasets and different methods in Table 1. The results show that: (1) Our DENSE achieves the highest accuracy across all datasets. In particular, DENSE outperforms the best baseline method Fed-ADI [55] by $5 . 0 8 \\%$ when $\\alpha = 0 . 3$ on CIFAR10 dataset. (2) FedAvg has the worst performance, which implies that directly averaging the model parameters cannot achieve a good performance under nonIID setting in one-shot FL. (3) As $\\alpha$ becomes smaller (i.e., data become more imbalanced), the performance of all methods decrease significantly, which shows that all methods suffer ",
794
+ "bbox": [
795
+ 174,
796
+ 676,
797
+ 475,
798
+ 911
799
+ ],
800
+ "page_idx": 6
801
+ },
802
+ {
803
+ "type": "image",
804
+ "img_path": "images/7afe915a39613e67bcc59c3ee4f8eec8f1486e67c0d910606b0ebda3e3ab2535.jpg",
805
+ "image_caption": [
806
+ "Figure 3: Left panel: Accuracy of FedAvg and clients’ local models across different local training epochs $E = \\{ 2 0 , 4 0 , 6 0 , \\cdots , 4 0 0 \\}$ . Right panel: The accuracy curve for local training. The dotted lines represent the best results of two one-shot FL methods (FedAvg and DENSE). Our DENSE outperforms FedAvg and local models consistently. "
807
+ ],
808
+ "image_footnote": [],
809
+ "bbox": [
810
+ 493,
811
+ 696,
812
+ 813,
813
+ 791
814
+ ],
815
+ "page_idx": 6
816
+ },
817
+ {
818
+ "type": "table",
819
+ "img_path": "images/27c99bde5f56a13daec53f95a059c921d6106a82bd4cb6ae2766b67a7e1d46cf.jpg",
820
+ "table_caption": [
821
+ "Table 2: Accuracy comparisons across heterogeneous client models on CIFAR10. There are five clients in total, and each client has a personalized model. "
822
+ ],
823
+ "table_footnote": [],
824
+ "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"5\">Client</td><td colspan=\"4\">Server (ResNet-18)</td></tr><tr><td>ResNet-18</td><td>CNN1</td><td>CNN2</td><td>WRN-16-1</td><td>WRN-40-1</td><td>FedDF</td><td>Fed-DAFL</td><td>Fed-ADI</td><td>DENSE (ours)</td></tr><tr><td>Q=0.1</td><td>40.83</td><td>33.67</td><td>35.21</td><td>27.73</td><td>32.93</td><td>42.35</td><td>43.12</td><td>44.63</td><td>49.76</td></tr><tr><td>Q=0.3</td><td>51.49</td><td>52.78</td><td>44.96</td><td>47.35</td><td>37.24</td><td>52.72</td><td>57.72</td><td>58.96</td><td>63.25</td></tr><tr><td>α=0.5</td><td>59.96</td><td>58.67</td><td>54.28</td><td>53.39</td><td>58.14</td><td>60.05</td><td>61.56</td><td>63.24</td><td>67.42</td></tr></table>",
825
+ "bbox": [
826
+ 176,
827
+ 125,
828
+ 831,
829
+ 202
830
+ ],
831
+ "page_idx": 7
832
+ },
833
+ {
834
+ "type": "image",
835
+ "img_path": "images/997f84347ee2f7b0394beb308bb0fc8e563a958887577e7c3d644b5d494bb0e8.jpg",
836
+ "image_caption": [
837
+ "Figure 4: Visualization of the test accuracy and data distribution for CIFAR10 with $\\alpha = \\{ 0 . 3 , 0 . 5 \\}$ "
838
+ ],
839
+ "image_footnote": [],
840
+ "bbox": [
841
+ 178,
842
+ 218,
843
+ 823,
844
+ 339
845
+ ],
846
+ "page_idx": 7
847
+ },
848
+ {
849
+ "type": "text",
850
+ "text": "from highly skewed data. Even under highly skewed setting, DENSE still significantly outperforms other methods, which further demonstrates the superiority of our proposed method. ",
851
+ "bbox": [
852
+ 176,
853
+ 395,
854
+ 821,
855
+ 422
856
+ ],
857
+ "page_idx": 7
858
+ },
859
+ {
860
+ "type": "text",
861
+ "text": "3.2.2 Impact of model distillation ",
862
+ "text_level": 1,
863
+ "bbox": [
864
+ 176,
865
+ 438,
866
+ 416,
867
+ 452
868
+ ],
869
+ "page_idx": 7
870
+ },
871
+ {
872
+ "type": "text",
873
+ "text": "We show the impact of model distillation by comparing with FedAvg. We first conduct one-shot FL and use FedAvg to aggregate the local models. We show the results of the global model and clients’ local models across different local training epochs $E = \\{ 2 0 , 4 0 , 6 0 , \\cdots , 4 0 0 \\}$ in the left panel of Figure 3. The global model achieves the best performance (test accuracy $= 3 4 \\%$ ) when $E = 4 0$ , while a larger value of $E$ can cause the model to degrade even collapse. This result can be attributed to the inconsistent optimization objectives with non-IID data [49], which leads to weight divergence [61]. Then, we show the results of one-shot FL when $E = 4 0 0$ and report the performance of FedAvg and DENSE in the right panel of Figure 3. We also plot the performance of clients’ local models. DENSE outperforms each client’s local model while FedAvg underperforms each client’s local model. This validates that model distillation can enhance training while directly aggregating is harmful to the training under non-IID setting in one-shot FL. ",
874
+ "bbox": [
875
+ 173,
876
+ 460,
877
+ 825,
878
+ 613
879
+ ],
880
+ "page_idx": 7
881
+ },
882
+ {
883
+ "type": "text",
884
+ "text": "3.2.3 Results in heterogeneous FL ",
885
+ "text_level": 1,
886
+ "bbox": [
887
+ 176,
888
+ 628,
889
+ 419,
890
+ 642
891
+ ],
892
+ "page_idx": 7
893
+ },
894
+ {
895
+ "type": "text",
896
+ "text": "Note that our proposed DENSE can support heterogeneous models. We apply five different CNN models on CIFAR10 dataset with Dirichlet distribution $\\alpha = \\{ 0 . 1 , 0 . 3 , 0 . 5 \\}$ . The heterogeneous models include: 1) one ResNet-18 [11], 2) two small CNNs: CNN1 and CNN2; 3) two Wide-ResNets (WRN) [57]: WRN-16-1 and WRN-40-1. For knowledge distillation, we use ResNet-18 as the server’s global model. Detailed architecture information of the given deep networks can be found in Appendix. Table 2 evaluates all methods in heterogeneous one-shot FL under practical non-IID data settings. We omit the results for FedAvg as FedAvg does not support heterogeneous models. We remark that FL under both the non-IID data distribution and different model architecture setting is a quite challenging task. Even under this setting, our DENSE still significantly outperforms other baselines. In addition, we report the accuracy curve of global distillation. As shown in Figure 4, our method outperforms other baselines by a large margin. ",
897
+ "bbox": [
898
+ 173,
899
+ 651,
900
+ 825,
901
+ 804
902
+ ],
903
+ "page_idx": 7
904
+ },
905
+ {
906
+ "type": "text",
907
+ "text": "3.3 Analysis of Our Method ",
908
+ "text_level": 1,
909
+ "bbox": [
910
+ 174,
911
+ 820,
912
+ 380,
913
+ 835
914
+ ],
915
+ "page_idx": 7
916
+ },
917
+ {
918
+ "type": "text",
919
+ "text": "3.3.1 Impact of the number of clients ",
920
+ "text_level": 1,
921
+ "bbox": [
922
+ 176,
923
+ 845,
924
+ 442,
925
+ 861
926
+ ],
927
+ "page_idx": 7
928
+ },
929
+ {
930
+ "type": "text",
931
+ "text": "Furthermore, we evaluate the performance of these methods on CIFAR10 and SVHN datasets by varying the number of clients $m = \\{ 5 , 1 0 , 2 0 , 5 0 , 1 0 0 \\}$ . According to [33], the server can become a bottleneck when the number of clients is very large, we are also concerned with the model performance when $m$ increases. Table 3 shows the results of different methods across different $m$ . The accuracy of all methods decreases as the number of clients $m$ increases, which is consistent with observations in [33, 42]. Even though the number of clients can affect the performance of one-shot FL, our method still outperforms other baselines. The increasing number of clients poses new challenges for ensemble distillation, which we leave for future investigation. ",
932
+ "bbox": [
933
+ 176,
934
+ 869,
935
+ 823,
936
+ 911
937
+ ],
938
+ "page_idx": 7
939
+ },
940
+ {
941
+ "type": "table",
942
+ "img_path": "images/0908ac74d76efb399be6b4e6ed74bf3d3fca08f06344eae5ac521211e38a743a.jpg",
943
+ "table_caption": [
944
+ "Table 3: Accuracy across different number of clients $m = \\{ 5 , 1 0 , 2 0 , 5 0 , 1 0 0 \\}$ on CIFAR10 and SVHN datasets. "
945
+ ],
946
+ "table_footnote": [],
947
+ "table_body": "<table><tr><td>Dataset</td><td colspan=\"5\">CIFAR10</td><td colspan=\"5\">SVHN</td></tr><tr><td>m</td><td>FedAvg</td><td>FedDF</td><td>Fed-DAFL</td><td>Fed-ADI</td><td>DENSE (ours)</td><td>FedAvg</td><td>FedDF</td><td>Fed-DAFL</td><td>Fed-ADI</td><td>DENSE (ours)</td></tr><tr><td>5</td><td>1 43.67</td><td>53.56</td><td>55.46</td><td>58.59</td><td>62.19</td><td>56.09</td><td>73.98</td><td>78.03</td><td>78.85</td><td>80.03</td></tr><tr><td>10</td><td>38.29</td><td>54.44</td><td>56.34</td><td>57.13</td><td>61.42</td><td>45.34</td><td>62.12</td><td>63.34</td><td>65.45</td><td>67.57</td></tr><tr><td>20</td><td>36.03</td><td>43.15</td><td>45.98</td><td>46.45</td><td>52.71</td><td>47.79</td><td>60.45</td><td>62.19</td><td>63.98</td><td>66.42</td></tr><tr><td>50</td><td>37.03</td><td>40.89</td><td>43.02</td><td>44.47</td><td>48.47</td><td>36.53</td><td>51.44</td><td>54.23</td><td>57.35</td><td>59.27</td></tr><tr><td>100</td><td>33.54</td><td>36.89</td><td>37.55</td><td>36.98</td><td>43.28</td><td>30.18</td><td>46.58</td><td>47.19</td><td>48.33</td><td>52.48</td></tr></table>",
948
+ "bbox": [
949
+ 178,
950
+ 126,
951
+ 823,
952
+ 228
953
+ ],
954
+ "page_idx": 8
955
+ },
956
+ {
957
+ "type": "table",
958
+ "img_path": "images/1a988e48c53129ddc9758f443bfe72aa03843d05f07004aaa83c1cf5cc583af6.jpg",
959
+ "table_caption": [
960
+ "Table 4: Performance analysis of DENSE+LDAM. "
961
+ ],
962
+ "table_footnote": [],
963
+ "table_body": "<table><tr><td>Dataset</td><td colspan=\"3\">CIFAR10</td><td colspan=\"3\">SVHN</td></tr><tr><td>Method</td><td>α=0.1</td><td>a=0.3</td><td>a=0.5</td><td>a=0.1</td><td>α=0.3</td><td>Q=0.5</td></tr><tr><td>DENSE</td><td>50.26</td><td>59.76</td><td>62.19</td><td>55.34</td><td>79.59</td><td>80.03</td></tr><tr><td>DENSE+LDAM</td><td>57.24</td><td>63.13</td><td>64.76</td><td>58.04</td><td>81.28</td><td>81.77</td></tr></table>",
964
+ "bbox": [
965
+ 290,
966
+ 273,
967
+ 709,
968
+ 345
969
+ ],
970
+ "page_idx": 8
971
+ },
972
+ {
973
+ "type": "text",
974
+ "text": "",
975
+ "bbox": [
976
+ 173,
977
+ 381,
978
+ 825,
979
+ 452
980
+ ],
981
+ "page_idx": 8
982
+ },
983
+ {
984
+ "type": "text",
985
+ "text": "3.3.2 Combination with imbalanced learning ",
986
+ "text_level": 1,
987
+ "bbox": [
988
+ 173,
989
+ 476,
990
+ 500,
991
+ 491
992
+ ],
993
+ "page_idx": 8
994
+ },
995
+ {
996
+ "type": "text",
997
+ "text": "The accuracy of federated learning reduces significantly with non-IID data, which has been broadly discussed in recent studies [29, 49]. Additionally, previous studies [1, 19] have demonstrated their superiority on imbalanced data. The combination of our method with these techniques to address imbalanced local data can lead to a more effective FL system. For example, by using LDAM [1] in clients’ local training, we can mitigate the impact of data imbalance, and thereby build a more powerful ensemble model. We compare the performance of the original DENSE and DENSE combined with LDAM (DENSE+LDAM) across $\\alpha = \\{ 0 . 1 , 0 . 3 , 0 . 5 \\}$ on CIFAR10 and SVHN datasets. ",
998
+ "bbox": [
999
+ 174,
1000
+ 505,
1001
+ 485,
1002
+ 710
1003
+ ],
1004
+ "page_idx": 8
1005
+ },
1006
+ {
1007
+ "type": "text",
1008
+ "text": "As demonstrated in Table 4, DENSE+LDAM can significantly improve the performance, especially for highly skewed non-IID data (i.e. $\\alpha = 0 . 1$ ). To help understand the performance gap and data skewness, in Figure 5, we visualize the accuracy curve and data distribution of CIFAR10 $( \\alpha { = } 0 . 1 )$ in the left panel and right panel respectively. The number in the right panel stands for the number of examples associated with the corresponding label in one particular client. These figures imply that significant improvement can be achieved by combining DENSE with LDAM on highly skewed data. ",
1009
+ "bbox": [
1010
+ 174,
1011
+ 718,
1012
+ 485,
1013
+ 910
1014
+ ],
1015
+ "page_idx": 8
1016
+ },
1017
+ {
1018
+ "type": "image",
1019
+ "img_path": "images/86839b766d8522061366b4faa54518774846b2b1c44f4ba7188558d299c00bfc.jpg",
1020
+ "image_caption": [
1021
+ "Figure 5: Left panel: Accuracy curves of DENSE and DENSE $+$ LDAM. Right panel: Data distribution of different clients for CIFAR10 dataset $( \\alpha { = } 0 . 1 )$ . "
1022
+ ],
1023
+ "image_footnote": [],
1024
+ "bbox": [
1025
+ 509,
1026
+ 508,
1027
+ 810,
1028
+ 622
1029
+ ],
1030
+ "page_idx": 8
1031
+ },
1032
+ {
1033
+ "type": "image",
1034
+ "img_path": "images/c560833d9a5c700338ba1d8898f10c61b70db8bc9e23feb190e95096a4188af9.jpg",
1035
+ "image_caption": [
1036
+ "Figure 6: Visualization of synthetic data on CIFAR10 and SVHN datasets. "
1037
+ ],
1038
+ "image_footnote": [],
1039
+ "bbox": [
1040
+ 542,
1041
+ 719,
1042
+ 772,
1043
+ 885
1044
+ ],
1045
+ "page_idx": 8
1046
+ },
1047
+ {
1048
+ "type": "text",
1049
+ "text": "3.3.3 Visualization of synthetic data ",
1050
+ "text_level": 1,
1051
+ "bbox": [
1052
+ 174,
1053
+ 90,
1054
+ 434,
1055
+ 106
1056
+ ],
1057
+ "page_idx": 9
1058
+ },
1059
+ {
1060
+ "type": "text",
1061
+ "text": "To compare the synthetic data with the training data, we visualize the synthetic data on CIFAR10 and SVHN datasets in Figure 6. As shown in the figure, the first / third row is the original data of CIFAR10 / SVHN dataset, and the second / last row is the synthetic data generated by the model trained on CIFAR10 / SVHN dataset. The synthetic data are not similar to the original data, which can effectively reduce the probability of leaking sensitive information of clients. Note that although the synthetic data look much different from the original data, our method still achieves a higher performance than other baseline methods by training with these synthetic data (as shown in Table 1). Note that the ideal synthetic data should be visually distinct from the real data. ",
1062
+ "bbox": [
1063
+ 173,
1064
+ 117,
1065
+ 825,
1066
+ 228
1067
+ ],
1068
+ "page_idx": 9
1069
+ },
1070
+ {
1071
+ "type": "text",
1072
+ "text": "3.3.4 Extend to multiple rounds ",
1073
+ "text_level": 1,
1074
+ "bbox": [
1075
+ 174,
1076
+ 248,
1077
+ 406,
1078
+ 263
1079
+ ],
1080
+ "page_idx": 9
1081
+ },
1082
+ {
1083
+ "type": "text",
1084
+ "text": "We also extend DENSE to multi-round FL to test its effectiveness, i.e., there are multiple communication rounds between clients and server. Table 5 demonstrates the results of DENSE across different communication rounds $T _ { c } =$ $\\{ 1 , 2 , 3 , 4 , 5 \\}$ on CIFAR10 and SVHN datasets. The local training epoch is fixed as $E = 1 0$ . The performance of ",
1085
+ "bbox": [
1086
+ 174,
1087
+ 273,
1088
+ 437,
1089
+ 398
1090
+ ],
1091
+ "page_idx": 9
1092
+ },
1093
+ {
1094
+ "type": "table",
1095
+ "img_path": "images/7a8fc861426d5b4f5057042fc82a4f484bcbc4cdbccc40842aeea7e1202f7fbe.jpg",
1096
+ "table_caption": [
1097
+ "Table 5: Accuracy for multiple communication rounds. "
1098
+ ],
1099
+ "table_footnote": [],
1100
+ "table_body": "<table><tr><td>Dataset</td><td colspan=\"3\">CIFAR10</td><td colspan=\"3\">SVHN</td></tr><tr><td>Communication rounds</td><td>α=0.1</td><td>α=0.3</td><td>α=0.5</td><td>a=0.1</td><td>α=0.3</td><td>α=0.5</td></tr><tr><td>Tc=1</td><td>50.72</td><td>59.41</td><td>63.89</td><td>54.344</td><td>79.87</td><td>80.14</td></tr><tr><td>T=2</td><td>63.08</td><td>65.90</td><td>71.16</td><td>56.13</td><td>79.75</td><td>85.18</td></tr><tr><td>Te=3</td><td>61.61</td><td>69.73</td><td>73.91</td><td>74.41</td><td>86.42</td><td>86.18</td></tr><tr><td>T=4</td><td>66.26</td><td>69.40</td><td>74.39</td><td>78.67</td><td>86.36</td><td>86.43</td></tr><tr><td>T=5</td><td>67.65</td><td>71.42</td><td>76.01</td><td>80.28</td><td>86.25</td><td>86.55</td></tr></table>",
1101
+ "bbox": [
1102
+ 452,
1103
+ 295,
1104
+ 820,
1105
+ 376
1106
+ ],
1107
+ "page_idx": 9
1108
+ },
1109
+ {
1110
+ "type": "text",
1111
+ "text": "DENSE improves as $T _ { c }$ increases, and DENSE achieves the best performance when $T _ { c } = 5$ . This shows that DENSE can be extended to multi-round FL and the performance can be further enhanced by increasing the communication rounds. ",
1112
+ "bbox": [
1113
+ 174,
1114
+ 398,
1115
+ 826,
1116
+ 440
1117
+ ],
1118
+ "page_idx": 9
1119
+ },
1120
+ {
1121
+ "type": "text",
1122
+ "text": "3.3.5 Contribution of $\\mathcal { L } _ { B N }$ and $\\mathcal { L } _ { d i v }$ ",
1123
+ "text_level": 1,
1124
+ "bbox": [
1125
+ 174,
1126
+ 460,
1127
+ 434,
1128
+ 476
1129
+ ],
1130
+ "page_idx": 9
1131
+ },
1132
+ {
1133
+ "type": "text",
1134
+ "text": "We investigate the contributions of different loss functions used in data generation. We conduct leave-one-out testing and show the results by removing $\\mathcal { L } _ { d i v }$ (w/o $\\mathcal { L } _ { d i v , }$ ), and removing $\\mathcal { L } _ { B N }$ (w/o $\\mathcal { L } _ { B N } \\mathrm { . }$ ). Additionally, we report the result by removing both $\\mathcal { L } _ { d i v }$ and $\\mathcal { L } _ { B N }$ , i.e., using only $\\mathcal { L } _ { C E }$ (w/ $\\mathcal { L } _ { C E } )$ ). As illustrated in Table 6, using only $\\mathcal { L } _ { C E }$ to train the generator leads to poor performance. Be",
1135
+ "bbox": [
1136
+ 173,
1137
+ 486,
1138
+ 439,
1139
+ 625
1140
+ ],
1141
+ "page_idx": 9
1142
+ },
1143
+ {
1144
+ "type": "table",
1145
+ "img_path": "images/6cb24af2fdb71e7a2fae8cfbeb64dcd6225000df4d23a9f2c3fa4e998eb09be9.jpg",
1146
+ "table_caption": [
1147
+ "Table 6: Impact of loss functions in data generation. "
1148
+ ],
1149
+ "table_footnote": [],
1150
+ "table_body": "<table><tr><td>Dataset</td><td>CIFAR10</td><td>SVHN</td><td>CIFAR100</td></tr><tr><td>DENSE</td><td>62.19</td><td>80.03</td><td>42.07</td></tr><tr><td>wlLcE</td><td>53.12</td><td>73.11</td><td>36.47</td></tr><tr><td>W/o LBN</td><td>61.05</td><td>78.36</td><td>39.89</td></tr><tr><td>w/o Ldiv</td><td>59.18</td><td>77.59</td><td>39.14</td></tr></table>",
1151
+ "bbox": [
1152
+ 472,
1153
+ 508,
1154
+ 797,
1155
+ 613
1156
+ ],
1157
+ "page_idx": 9
1158
+ },
1159
+ {
1160
+ "type": "text",
1161
+ "text": "sides, removing either the $\\mathcal { L } _ { B N }$ loss or $\\mathcal { L } _ { d i v }$ loss also affects the accuracy of the global model. A combination of these loss functions leads to a high performance of global model, which shows that each part of the loss function plays an important role in enhancing the generator. ",
1162
+ "bbox": [
1163
+ 173,
1164
+ 625,
1165
+ 825,
1166
+ 666
1167
+ ],
1168
+ "page_idx": 9
1169
+ },
1170
+ {
1171
+ "type": "text",
1172
+ "text": "4 Conclusion ",
1173
+ "text_level": 1,
1174
+ "bbox": [
1175
+ 174,
1176
+ 691,
1177
+ 299,
1178
+ 708
1179
+ ],
1180
+ "page_idx": 9
1181
+ },
1182
+ {
1183
+ "type": "text",
1184
+ "text": "In this paper, we propose an effective one-shot federated learning method called DENSE, which trains the global model by a data generation stage and a model distillation stage. Extensive experiments across various settings validate the efficacy of our method. Overall, DENSE is by far the most practical framework that can conduct data-free one-shot FL with model heterogeneity. A promising future direction is to consider the potential privacy attacks in one-shot FL. ",
1185
+ "bbox": [
1186
+ 174,
1187
+ 726,
1188
+ 825,
1189
+ 796
1190
+ ],
1191
+ "page_idx": 9
1192
+ },
1193
+ {
1194
+ "type": "text",
1195
+ "text": "5 Acknowledgement ",
1196
+ "text_level": 1,
1197
+ "bbox": [
1198
+ 174,
1199
+ 820,
1200
+ 359,
1201
+ 838
1202
+ ],
1203
+ "page_idx": 9
1204
+ },
1205
+ {
1206
+ "type": "text",
1207
+ "text": "This work was supported by Sony AI, the National Key Research and Development Project of China (2021ZD0110400 No. 2018AAA0101900), National Natural Science Foundation of China (U19B2042), Zhejiang Lab (2021KE0AC02), Academy Of Social Governance Zhejiang University, Fundamental Research Funds for the Central Universities. ",
1208
+ "bbox": [
1209
+ 174,
1210
+ 854,
1211
+ 825,
1212
+ 911
1213
+ ],
1214
+ "page_idx": 9
1215
+ },
1216
+ {
1217
+ "type": "text",
1218
+ "text": "References \n[1] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss, 2019. \n[2] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Dafl: Data-free learning of student networks. In ICCV, 2019. \n[3] Hong-You Chen and Wei-Lun Chao. Fedbe: Making bayesian model ensemble applicable to federated learning, 2021. \n[4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: Deformable patch attack. In European conference on computer vision, 2022. \n[5] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15148–15158, June 2022. \n[6] Zhen Chen, Meilu Zhu, Chen Yang, and Yixuan Yuan. Personalized retrogress-resilient framework for real-world medical federated learning. In Marleen de Bruijne, Philippe C. Cattin, Stéphane Cotin, Nicolas Padoy, Stefanie Speidel, Yefeng Zheng, and Caroline Essert, editors, Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part III, volume 12903 of Lecture Notes in Computer Science, pages 347–356. Springer, 2021. \n[7] Don Kurian Dennis, Tian Li, and Virginia Smith. Heterogeneity for the win: One-shot federated clustering, 2021. \n[8] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 11(1):1–17, 2021. \n[9] Jiahua Dong, Yang Cong, Gan Sun, Bineng Zhong, and Xiaowei Xu. What can be transferred: Unsupervised domain adaptation for endoscopic lesions segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4022–4031, June 2020. \n[10] Neel Guha, Ameet Talwalkar, and Virginia Smith. One-shot federated learning. arXiv preprint arXiv:1902.11175, 2019. \n[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. \n[12] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge distillation with adversarial samples supporting decision boundary. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3771–3778, 2019. \n[13] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. \n[14] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015. \n[15] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 603–618, 2017. \n[16] Hao Huang, Yongtao Wang, Zhaoyu Chen, Yuze Zhang, Yuheng Li, Zhi Tang, Wei Chu, Jingdong Chen, Weisi Lin, and Kai-Kuang Ma. Cmua-watermark: A cross-model universal adversarial watermark for combating deepfakes. Proceedings of the AAAI Conference on Artificial Intelligence, 36(1):989–997, Jun. 2022. \n[17] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift, 2015. ",
1219
+ "bbox": [
1220
+ 173,
1221
+ 68,
1222
+ 828,
1223
+ 916
1224
+ ],
1225
+ "page_idx": 10
1226
+ },
1227
+ {
1228
+ "type": "text",
1229
+ "text": "[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning, pages 5132–5143. PMLR, 2020. ",
1230
+ "bbox": [
1231
+ 171,
1232
+ 90,
1233
+ 823,
1234
+ 133
1235
+ ],
1236
+ "page_idx": 11
1237
+ },
1238
+ {
1239
+ "type": "text",
1240
+ "text": "[19] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via majorto-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. ",
1241
+ "bbox": [
1242
+ 173,
1243
+ 140,
1244
+ 823,
1245
+ 184
1246
+ ],
1247
+ "page_idx": 11
1248
+ },
1249
+ {
1250
+ "type": "text",
1251
+ "text": "[20] Aashish Kolluri, Teodora Baluta, and Prateek Saxena. Private hierarchical clustering in federated networks. In Yongdae Kim, Jong Kim, Giovanni Vigna, and Elaine Shi, editors, CCS ’21: 2021 ACM SIGSAC Conference on Computer and Communications Security, Virtual Event, Republic of Korea, November 15 - 19, 2021, pages 2342–2360. ACM, 2021. ",
1252
+ "bbox": [
1253
+ 174,
1254
+ 191,
1255
+ 823,
1256
+ 247
1257
+ ],
1258
+ "page_idx": 11
1259
+ },
1260
+ {
1261
+ "type": "text",
1262
+ "text": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. ",
1263
+ "bbox": [
1264
+ 171,
1265
+ 256,
1266
+ 825,
1267
+ 285
1268
+ ],
1269
+ "page_idx": 11
1270
+ },
1271
+ {
1272
+ "type": "text",
1273
+ "text": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017. ",
1274
+ "bbox": [
1275
+ 174,
1276
+ 291,
1277
+ 823,
1278
+ 321
1279
+ ],
1280
+ "page_idx": 11
1281
+ },
1282
+ {
1283
+ "type": "text",
1284
+ "text": "[23] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. ",
1285
+ "bbox": [
1286
+ 179,
1287
+ 329,
1288
+ 810,
1289
+ 344
1290
+ ],
1291
+ "page_idx": 11
1292
+ },
1293
+ {
1294
+ "type": "text",
1295
+ "text": "[24] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ",
1296
+ "bbox": [
1297
+ 173,
1298
+ 352,
1299
+ 821,
1300
+ 381
1301
+ ],
1302
+ "page_idx": 11
1303
+ },
1304
+ {
1305
+ "type": "text",
1306
+ "text": "[25] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., volume 2, pages II–104. IEEE, 2004. ",
1307
+ "bbox": [
1308
+ 174,
1309
+ 387,
1310
+ 825,
1311
+ 444
1312
+ ],
1313
+ "page_idx": 11
1314
+ },
1315
+ {
1316
+ "type": "text",
1317
+ "text": "[26] Bo Li, Zhengxing Sun, and Yuqi Guo. Supervae: Superpixelwise variational autoencoder for salient object detection. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 8569–8576. AAAI Press, 2019. ",
1318
+ "bbox": [
1319
+ 174,
1320
+ 452,
1321
+ 826,
1322
+ 522
1323
+ ],
1324
+ "page_idx": 11
1325
+ },
1326
+ {
1327
+ "type": "text",
1328
+ "text": "[27] Bo Li, Zhengxing Sun, Lv Tang, Yunhan Sun, and Jinlong Shi. Detecting robust co-saliency with recurrent co-attention neural network. In Sarit Kraus, editor, Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 818–825. ijcai.org, 2019. ",
1329
+ "bbox": [
1330
+ 174,
1331
+ 530,
1332
+ 825,
1333
+ 587
1334
+ ],
1335
+ "page_idx": 11
1336
+ },
1337
+ {
1338
+ "type": "text",
1339
+ "text": "[28] Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation. arXiv preprint arXiv:1910.03581, 2019. ",
1340
+ "bbox": [
1341
+ 173,
1342
+ 594,
1343
+ 823,
1344
+ 623
1345
+ ],
1346
+ "page_idx": 11
1347
+ },
1348
+ {
1349
+ "type": "text",
1350
+ "text": "[29] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study, 2021. ",
1351
+ "bbox": [
1352
+ 174,
1353
+ 631,
1354
+ 823,
1355
+ 660
1356
+ ],
1357
+ "page_idx": 11
1358
+ },
1359
+ {
1360
+ "type": "text",
1361
+ "text": "[30] Qinbin Li, Bingsheng He, and Dawn Song. Practical one-shot federated learning for cross-silo setting. arXiv preprint arXiv:2010.01017, 2020. ",
1362
+ "bbox": [
1363
+ 171,
1364
+ 667,
1365
+ 823,
1366
+ 696
1367
+ ],
1368
+ "page_idx": 11
1369
+ },
1370
+ {
1371
+ "type": "text",
1372
+ "text": "[31] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks, 2020. ",
1373
+ "bbox": [
1374
+ 171,
1375
+ 704,
1376
+ 823,
1377
+ 733
1378
+ ],
1379
+ "page_idx": 11
1380
+ },
1381
+ {
1382
+ "type": "text",
1383
+ "text": "[32] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations, 2020. ",
1384
+ "bbox": [
1385
+ 173,
1386
+ 739,
1387
+ 825,
1388
+ 784
1389
+ ],
1390
+ "page_idx": 11
1391
+ },
1392
+ {
1393
+ "type": "text",
1394
+ "text": "[33] Xiangru Lian, Ce Zhang, Huan Zhang, Cho-Jui Hsieh, Wei Zhang, and Ji Liu. Can decentralized algorithms outperform centralized algorithms? a case study for decentralized parallel stochastic gradient descent, 2017. ",
1395
+ "bbox": [
1396
+ 173,
1397
+ 791,
1398
+ 823,
1399
+ 833
1400
+ ],
1401
+ "page_idx": 11
1402
+ },
1403
+ {
1404
+ "type": "text",
1405
+ "text": "[34] Feng Liang, Weike Pan, and Zhong Ming. Fedrec++: Lossless federated recommendation with explicit feedback. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021. ",
1406
+ "bbox": [
1407
+ 174,
1408
+ 842,
1409
+ 826,
1410
+ 911
1411
+ ],
1412
+ "page_idx": 11
1413
+ },
1414
+ {
1415
+ "type": "text",
1416
+ "text": "[35] Tao Lin, Lingjing Kong, Sebastian U. Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. ",
1417
+ "bbox": [
1418
+ 174,
1419
+ 90,
1420
+ 826,
1421
+ 161
1422
+ ],
1423
+ "page_idx": 12
1424
+ },
1425
+ {
1426
+ "type": "text",
1427
+ "text": "[36] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1013–1023. Computer Vision Foundation / IEEE, 2021. ",
1428
+ "bbox": [
1429
+ 174,
1430
+ 169,
1431
+ 825,
1432
+ 226
1433
+ ],
1434
+ "page_idx": 12
1435
+ },
1436
+ {
1437
+ "type": "text",
1438
+ "text": "[37] Ruixuan Liu, Yang Cao, Hong Chen, Ruoyang Guo, and Masatoshi Yoshikawa. FLAME: differentially private federated learning in the shuffle model. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 8688–8696. AAAI Press, 2021. ",
1439
+ "bbox": [
1440
+ 174,
1441
+ 233,
1442
+ 825,
1443
+ 318
1444
+ ],
1445
+ "page_idx": 12
1446
+ },
1447
+ {
1448
+ "type": "text",
1449
+ "text": "[38] Shuchang Liu, Shuyuan Xu, Wenhui Yu, Zuohui Fu, Yongfeng Zhang, and Amélie Marian. Fedct: Federated collaborative transfer for recommendation. In Fernando Diaz, Chirag Shah, Torsten Suel, Pablo Castells, Rosie Jones, and Tetsuya Sakai, editors, SIGIR ’21: The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, Canada, July 11-15, 2021, pages 716–725. ACM, 2021. ",
1450
+ "bbox": [
1451
+ 173,
1452
+ 325,
1453
+ 825,
1454
+ 395
1455
+ ],
1456
+ "page_idx": 12
1457
+ },
1458
+ {
1459
+ "type": "text",
1460
+ "text": "[39] Yi Liu, Jialiang Peng, JQ James, and Yi Wu. Ppgan: Privacy-preserving generative adversarial network. In 2019 IEEE 25Th international conference on parallel and distributed systems (ICPADS), pages 985–989. IEEE, 2019. ",
1461
+ "bbox": [
1462
+ 173,
1463
+ 404,
1464
+ 823,
1465
+ 446
1466
+ ],
1467
+ "page_idx": 12
1468
+ },
1469
+ {
1470
+ "type": "text",
1471
+ "text": "[40] Lingjuan Lyu, Han Yu, Xingjun Ma, Chen Chen, Lichao Sun, Jun Zhao, Qiang Yang, and Philip S Yu. Privacy and robustness in federated learning: Attacks and defenses. arXiv preprint arXiv:2012.06337, 2020. ",
1472
+ "bbox": [
1473
+ 173,
1474
+ 454,
1475
+ 823,
1476
+ 497
1477
+ ],
1478
+ "page_idx": 12
1479
+ },
1480
+ {
1481
+ "type": "text",
1482
+ "text": "[41] Lingjuan Lyu, Han Yu, Jun Zhao, and Qiang Yang. Threats to federated learning. In Federated Learning, pages 3–16. Springer, 2020. ",
1483
+ "bbox": [
1484
+ 171,
1485
+ 505,
1486
+ 825,
1487
+ 534
1488
+ ],
1489
+ "page_idx": 12
1490
+ },
1491
+ {
1492
+ "type": "text",
1493
+ "text": "[42] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data, 2017. ",
1494
+ "bbox": [
1495
+ 173,
1496
+ 541,
1497
+ 823,
1498
+ 571
1499
+ ],
1500
+ "page_idx": 12
1501
+ },
1502
+ {
1503
+ "type": "text",
1504
+ "text": "[43] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew $\\mathrm { ~ Y ~ N ~ g ~ } _ { }$ Reading digits in natural images with unsupervised feature learning. 2011. ",
1505
+ "bbox": [
1506
+ 171,
1507
+ 579,
1508
+ 823,
1509
+ 608
1510
+ ],
1511
+ "page_idx": 12
1512
+ },
1513
+ {
1514
+ "type": "text",
1515
+ "text": "[44] Lichao Sun and Lingjuan Lyu. Federated model distillation with noise-free differential privacy. In IJCAI, 2021. ",
1516
+ "bbox": [
1517
+ 171,
1518
+ 616,
1519
+ 825,
1520
+ 645
1521
+ ],
1522
+ "page_idx": 12
1523
+ },
1524
+ {
1525
+ "type": "text",
1526
+ "text": "[45] Manasi Vartak. MODELDB: A system for machine learning model management. In 8th Biennial Conference on Innovative Data Systems Research, CIDR 2017, Chaminade, CA, USA, January 8-11, 2017, Online Proceedings. www.cidrdb.org, 2017. ",
1527
+ "bbox": [
1528
+ 173,
1529
+ 651,
1530
+ 825,
1531
+ 695
1532
+ ],
1533
+ "page_idx": 12
1534
+ },
1535
+ {
1536
+ "type": "text",
1537
+ "text": "[46] Chenghong Wang, Jieren Deng, Xianrui Meng, Yijue Wang, Ji Li, Fei Miao, Sanguthevar Rajasekaran, and Caiwen Ding. A secure and efficient federated learning framework for NLP. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih, editors, Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021. ",
1538
+ "bbox": [
1539
+ 173,
1540
+ 703,
1541
+ 826,
1542
+ 772
1543
+ ],
1544
+ "page_idx": 12
1545
+ },
1546
+ {
1547
+ "type": "text",
1548
+ "text": "[47] Derui Wang, Chaoran Li, Sheng Wen, Surya Nepal, and Yang Xiang. Man-in-the-middle attacks against machine learning classifiers via malicious generative models. IEEE Transactions on Dependable and Secure Computing, 2020. ",
1549
+ "bbox": [
1550
+ 173,
1551
+ 781,
1552
+ 823,
1553
+ 824
1554
+ ],
1555
+ "page_idx": 12
1556
+ },
1557
+ {
1558
+ "type": "text",
1559
+ "text": "[48] Hongyi Wang, Saurabh Agarwal, and Dimitris S. Papailiopoulos. Pufferfish: Communicationefficient models at no extra cost. CoRR, abs/2103.03936, 2021. ",
1560
+ "bbox": [
1561
+ 173,
1562
+ 832,
1563
+ 821,
1564
+ 861
1565
+ ],
1566
+ "page_idx": 12
1567
+ },
1568
+ {
1569
+ "type": "text",
1570
+ "text": "[49] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. arXiv preprint arXiv:2007.07481, 2020. ",
1571
+ "bbox": [
1572
+ 173,
1573
+ 868,
1574
+ 823,
1575
+ 911
1576
+ ],
1577
+ "page_idx": 12
1578
+ },
1579
+ {
1580
+ "type": "text",
1581
+ "text": "[50] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A. Efros. Dataset distillation, 2020. ",
1582
+ "bbox": [
1583
+ 171,
1584
+ 90,
1585
+ 826,
1586
+ 121
1587
+ ],
1588
+ "page_idx": 13
1589
+ },
1590
+ {
1591
+ "type": "text",
1592
+ "text": "[51] Kun Wei, Da Chen, Yuhong Li, Xu Yang, Cheng Deng, and Dacheng Tao. Incremental embedding learning with disentangled representation translation. IEEE Transactions on Neural Networks and Learning Systems, 2022. ",
1593
+ "bbox": [
1594
+ 174,
1595
+ 127,
1596
+ 823,
1597
+ 171
1598
+ ],
1599
+ "page_idx": 13
1600
+ },
1601
+ {
1602
+ "type": "text",
1603
+ "text": "[52] Kun Wei, Cheng Deng, Xu Yang, and Maosen Li. Incremental embedding learning via zero-shot translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10254–10262, 2021. ",
1604
+ "bbox": [
1605
+ 176,
1606
+ 179,
1607
+ 823,
1608
+ 223
1609
+ ],
1610
+ "page_idx": 13
1611
+ },
1612
+ {
1613
+ "type": "text",
1614
+ "text": "[53] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. ",
1615
+ "bbox": [
1616
+ 173,
1617
+ 231,
1618
+ 823,
1619
+ 261
1620
+ ],
1621
+ "page_idx": 13
1622
+ },
1623
+ {
1624
+ "type": "text",
1625
+ "text": "[54] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16337–16346, 2021. ",
1626
+ "bbox": [
1627
+ 173,
1628
+ 268,
1629
+ 825,
1630
+ 313
1631
+ ],
1632
+ "page_idx": 13
1633
+ },
1634
+ {
1635
+ "type": "text",
1636
+ "text": "[55] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deepinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715–8724, 2020. ",
1637
+ "bbox": [
1638
+ 173,
1639
+ 319,
1640
+ 826,
1641
+ 376
1642
+ ],
1643
+ "page_idx": 13
1644
+ },
1645
+ {
1646
+ "type": "text",
1647
+ "text": "[56] Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Trong Nghia Hoang, and Yasaman Khazaeni. Bayesian nonparametric federated learning of neural networks, 2019. ",
1648
+ "bbox": [
1649
+ 173,
1650
+ 386,
1651
+ 825,
1652
+ 428
1653
+ ],
1654
+ "page_idx": 13
1655
+ },
1656
+ {
1657
+ "type": "text",
1658
+ "text": "[57] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks, 2017. ",
1659
+ "bbox": [
1660
+ 174,
1661
+ 436,
1662
+ 691,
1663
+ 452
1664
+ ],
1665
+ "page_idx": 13
1666
+ },
1667
+ {
1668
+ "type": "text",
1669
+ "text": "[58] Jie Zhang, Chen Chen, Jiahua Dong, Ruoxi Jia, and Lingjuan Lyu. Qekd: Query-efficient and data-free knowledge distillation from black-box models. arXiv preprint arXiv:2205.11158, 2022. ",
1670
+ "bbox": [
1671
+ 173,
1672
+ 460,
1673
+ 826,
1674
+ 502
1675
+ ],
1676
+ "page_idx": 13
1677
+ },
1678
+ {
1679
+ "type": "text",
1680
+ "text": "[59] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In International Conference on Machine Learning, pages 26311–26329. PMLR, 2022. ",
1681
+ "bbox": [
1682
+ 171,
1683
+ 512,
1684
+ 825,
1685
+ 555
1686
+ ],
1687
+ "page_idx": 13
1688
+ },
1689
+ {
1690
+ "type": "text",
1691
+ "text": "[60] Shaofeng Zhang, Meng Liu, and Junchi Yan. The diversified ensemble neural network. Advances in Neural Information Processing Systems, 33, 2020. ",
1692
+ "bbox": [
1693
+ 173,
1694
+ 563,
1695
+ 823,
1696
+ 593
1697
+ ],
1698
+ "page_idx": 13
1699
+ },
1700
+ {
1701
+ "type": "text",
1702
+ "text": "[61] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data, 2018. ",
1703
+ "bbox": [
1704
+ 176,
1705
+ 601,
1706
+ 825,
1707
+ 630
1708
+ ],
1709
+ "page_idx": 13
1710
+ },
1711
+ {
1712
+ "type": "text",
1713
+ "text": "[62] Yanlin Zhou, George Pu, Xiyao Ma, Xiaolin Li, and Dapeng Wu. Distilled one-shot federated learning. arXiv preprint arXiv:2009.07999, 2020. ",
1714
+ "bbox": [
1715
+ 173,
1716
+ 638,
1717
+ 825,
1718
+ 667
1719
+ ],
1720
+ "page_idx": 13
1721
+ },
1722
+ {
1723
+ "type": "text",
1724
+ "text": "[63] Xinghua Zhu, Jianzong Wang, Zhenhou Hong, and Jing Xiao. Empirical studies of institutional federated learning for natural language processing. In Trevor Cohn, Yulan He, and Yang Liu, editors, Findings of the Association for Computational Linguistics: EMNLP 2020, Online Event, 16-20 November 2020, volume EMNLP 2020 of Findings of ACL, pages 625–634. Association for Computational Linguistics, 2020. ",
1725
+ "bbox": [
1726
+ 174,
1727
+ 676,
1728
+ 825,
1729
+ 746
1730
+ ],
1731
+ "page_idx": 13
1732
+ },
1733
+ {
1734
+ "type": "text",
1735
+ "text": "[64] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 12878–12889. PMLR, 2021. ",
1736
+ "bbox": [
1737
+ 174,
1738
+ 756,
1739
+ 828,
1740
+ 813
1741
+ ],
1742
+ "page_idx": 13
1743
+ },
1744
+ {
1745
+ "type": "text",
1746
+ "text": "Checklist ",
1747
+ "text_level": 1,
1748
+ "bbox": [
1749
+ 174,
1750
+ 89,
1751
+ 254,
1752
+ 106
1753
+ ],
1754
+ "page_idx": 14
1755
+ },
1756
+ {
1757
+ "type": "text",
1758
+ "text": "1. For all authors... ",
1759
+ "bbox": [
1760
+ 214,
1761
+ 116,
1762
+ 339,
1763
+ 131
1764
+ ],
1765
+ "page_idx": 14
1766
+ },
1767
+ {
1768
+ "type": "text",
1769
+ "text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] \n(c) Did you discuss any potential negative societal impacts of your work? [No] \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
1770
+ "bbox": [
1771
+ 238,
1772
+ 135,
1773
+ 825,
1774
+ 227
1775
+ ],
1776
+ "page_idx": 14
1777
+ },
1778
+ {
1779
+ "type": "text",
1780
+ "text": "2. If you are including theoretical results... ",
1781
+ "bbox": [
1782
+ 214,
1783
+ 231,
1784
+ 493,
1785
+ 244
1786
+ ],
1787
+ "page_idx": 14
1788
+ },
1789
+ {
1790
+ "type": "text",
1791
+ "text": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] (b) Did you include complete proofs of all theoretical results? [Yes] ",
1792
+ "bbox": [
1793
+ 236,
1794
+ 248,
1795
+ 735,
1796
+ 280
1797
+ ],
1798
+ "page_idx": 14
1799
+ },
1800
+ {
1801
+ "type": "text",
1802
+ "text": "3. If you ran experiments... ",
1803
+ "bbox": [
1804
+ 212,
1805
+ 285,
1806
+ 393,
1807
+ 299
1808
+ ],
1809
+ "page_idx": 14
1810
+ },
1811
+ {
1812
+ "type": "text",
1813
+ "text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [No] \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] ",
1814
+ "bbox": [
1815
+ 238,
1816
+ 303,
1817
+ 826,
1818
+ 421
1819
+ ],
1820
+ "page_idx": 14
1821
+ },
1822
+ {
1823
+ "type": "text",
1824
+ "text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
1825
+ "bbox": [
1826
+ 214,
1827
+ 426,
1828
+ 823,
1829
+ 440
1830
+ ],
1831
+ "page_idx": 14
1832
+ },
1833
+ {
1834
+ "type": "text",
1835
+ "text": "(a) If your work uses existing assets, did you cite the creators? [Yes] \n(b) Did you mention the license of the assets? [N/A] \n(c) Did you include any new assets either in the supplemental material or as a URL? [No] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] ",
1836
+ "bbox": [
1837
+ 238,
1838
+ 444,
1839
+ 823,
1840
+ 551
1841
+ ],
1842
+ "page_idx": 14
1843
+ },
1844
+ {
1845
+ "type": "text",
1846
+ "text": "5. If you used crowdsourcing or conducted research with human subjects... ",
1847
+ "bbox": [
1848
+ 214,
1849
+ 556,
1850
+ 705,
1851
+ 570
1852
+ ],
1853
+ "page_idx": 14
1854
+ },
1855
+ {
1856
+ "type": "text",
1857
+ "text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ",
1858
+ "bbox": [
1859
+ 238,
1860
+ 574,
1861
+ 825,
1862
+ 665
1863
+ ],
1864
+ "page_idx": 14
1865
+ }
1866
+ ]
parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_layout.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5ebad17bc58f39697990bfbe8ae1dc78f623c6487fbefe1cd14d06dbb40ee307
3
+ size 1971141
parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70146daeb5d07cde8a15901b01dd2c887903bde8baa3c9e68197aedd8b6690cc
3
+ size 1797555
parse/dev/QFQoxCFYEkA/QFQoxCFYEkA_span.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f3bccd7fc59ddb08881058cc1866bcfa1adf75637537262d6cc9c9dc59599b8d
3
+ size 1977640
parse/dev/QFQoxCFYEkA/images/0908ac74d76efb399be6b4e6ed74bf3d3fca08f06344eae5ac521211e38a743a.jpg ADDED

Git LFS Details

  • SHA256: 0ca6616bebdf812226f5a820380448e4cbf99b9a82913ec514229e63c17f0ef9
  • Pointer size: 130 Bytes
  • Size of remote file: 51.7 kB
parse/dev/QFQoxCFYEkA/images/1a988e48c53129ddc9758f443bfe72aa03843d05f07004aaa83c1cf5cc583af6.jpg ADDED

Git LFS Details

  • SHA256: ba5842ad51d0547e739dd40aaa96fb0d73a05b176683ff773a0d010fa40feea5
  • Pointer size: 130 Bytes
  • Size of remote file: 26.8 kB
parse/dev/QFQoxCFYEkA/images/27c99bde5f56a13daec53f95a059c921d6106a82bd4cb6ae2766b67a7e1d46cf.jpg ADDED

Git LFS Details

  • SHA256: c3fd8017d59c764ea5cb50f93e71e19ddf0b723e2a6db0f43dea44d2e787f64f
  • Pointer size: 130 Bytes
  • Size of remote file: 38.5 kB
parse/dev/QFQoxCFYEkA/images/2b2e84d45592acc7d50816d7c7dbfcba9119ec8fa3c317113edc2160d6ff6eef.jpg ADDED

Git LFS Details

  • SHA256: 14ff2a7a78c80246595ac51604191a4914a418f301d99d2cf183e462d3403dd4
  • Pointer size: 129 Bytes
  • Size of remote file: 8.4 kB
parse/dev/QFQoxCFYEkA/images/2ce9636880a058cd0713d603581fa7f3e6acfe162a440e0ad227547b6544a41c.jpg ADDED

Git LFS Details

  • SHA256: add1619647f75e219045f250045a87dc9324f31105a54902ee7875448950c461
  • Pointer size: 130 Bytes
  • Size of remote file: 66.5 kB
parse/dev/QFQoxCFYEkA/images/47710160045685a73f2907778920be78282559b5f39d434ffbd97b2a67637d92.jpg ADDED

Git LFS Details

  • SHA256: 2a47c7d0fe85993f2e6c18e040fc165c1aa84554f76de353753baf29d80bd0f8
  • Pointer size: 129 Bytes
  • Size of remote file: 3.16 kB
parse/dev/QFQoxCFYEkA/images/520285b0f14a26c736e532c26652dfd9fc8275befa8c250aad7dda976e95bca6.jpg ADDED

Git LFS Details

  • SHA256: f95c4935b66ae73387b1c1ac01421852d5fb7ec44f05d495defd0c7a2bf8358e
  • Pointer size: 129 Bytes
  • Size of remote file: 5.69 kB
parse/dev/QFQoxCFYEkA/images/6c71d9f5a4178d1b48f2d19bbb98de16ffa57a60378aaed0df0bb58c218884ba.jpg ADDED

Git LFS Details

  • SHA256: 03a2a92d6bed11f5ef19f2200c4ad0190ae96273144e11351def007410ed77be
  • Pointer size: 129 Bytes
  • Size of remote file: 6.27 kB
parse/dev/QFQoxCFYEkA/images/6cb24af2fdb71e7a2fae8cfbeb64dcd6225000df4d23a9f2c3fa4e998eb09be9.jpg ADDED

Git LFS Details

  • SHA256: d9852a367b0a665c2a7adaaf7881f3232bdd670ad4c8ddfb805aac4a12987e06
  • Pointer size: 130 Bytes
  • Size of remote file: 26.6 kB
parse/dev/QFQoxCFYEkA/images/76d9a9a4dd165c97d9c6b81648bf69bc02c6b45b06c64e217312f1914f3b33a5.jpg ADDED

Git LFS Details

  • SHA256: 8c345c7a1aa8820ad9388fb3f8f88097ffdd5bcaae984d20d87cee9d39a11d37
  • Pointer size: 130 Bytes
  • Size of remote file: 23.8 kB
parse/dev/QFQoxCFYEkA/images/7a8fc861426d5b4f5057042fc82a4f484bcbc4cdbccc40842aeea7e1202f7fbe.jpg ADDED

Git LFS Details

  • SHA256: d31ad72b18f31f39c7fb6533923713f3a7b5289971a5ec2f4efe53e4bc9c5d51
  • Pointer size: 130 Bytes
  • Size of remote file: 26.2 kB
parse/dev/QFQoxCFYEkA/images/7afe915a39613e67bcc59c3ee4f8eec8f1486e67c0d910606b0ebda3e3ab2535.jpg ADDED

Git LFS Details

  • SHA256: 597e859a096622f01e44d644a967c22610476ad20c984110e7d32b444b061751
  • Pointer size: 130 Bytes
  • Size of remote file: 22.5 kB
parse/dev/QFQoxCFYEkA/images/86839b766d8522061366b4faa54518774846b2b1c44f4ba7188558d299c00bfc.jpg ADDED

Git LFS Details

  • SHA256: 8b0965333fa1b4a6a52908efa14f7f9f4be213ad75f143db3d2def744893e095
  • Pointer size: 130 Bytes
  • Size of remote file: 20.6 kB
parse/dev/QFQoxCFYEkA/images/997f84347ee2f7b0394beb308bb0fc8e563a958887577e7c3d644b5d494bb0e8.jpg ADDED

Git LFS Details

  • SHA256: 53789083bdc317a3a9d4d3e1dbce5a1452c455c7331b07aeb81c678558d11d68
  • Pointer size: 130 Bytes
  • Size of remote file: 48.3 kB
parse/dev/QFQoxCFYEkA/images/b9a7f312faf7e4a780e52e674aa461846742d3a0e70f090e75b0124546d4c1e1.jpg ADDED

Git LFS Details

  • SHA256: 56b8b5bc6230cfd287f6246898d0b8de28c17a2e901e00693b591a2c029c7a14
  • Pointer size: 129 Bytes
  • Size of remote file: 4.55 kB
parse/dev/QFQoxCFYEkA/images/bc06aa78c13f13ab8794fd3beec2575b3d720ca22c180a49252fd14c96d37d5f.jpg ADDED

Git LFS Details

  • SHA256: 5fcdf351a0e61dcc6816c7e25280315be60a9fa7192c40459b41f39ae535d9cc
  • Pointer size: 130 Bytes
  • Size of remote file: 52.7 kB
parse/dev/QFQoxCFYEkA/images/c560833d9a5c700338ba1d8898f10c61b70db8bc9e23feb190e95096a4188af9.jpg ADDED

Git LFS Details

  • SHA256: f43bf9d756ff35177a49a804effe4d1c50186dc0af42077d93d60e88cf01aa48
  • Pointer size: 130 Bytes
  • Size of remote file: 40.4 kB
parse/dev/QFQoxCFYEkA/images/c963de924077ad4d47b94f877b9169c34bb7f0e549bd8ea7f5d0b06ddaa6a9a1.jpg ADDED

Git LFS Details

  • SHA256: 9a60975fb08b2ff6411fc95779b8e200593990cd6c86321889f50f8a215e984d
  • Pointer size: 129 Bytes
  • Size of remote file: 5.23 kB
parse/dev/QFQoxCFYEkA/images/de28f63560d2cb651e1066d0a90842918e0e047cf416964e25375636f1fc64fa.jpg ADDED

Git LFS Details

  • SHA256: 480d7c4d4302e022514d54a402bf3e6b1aa6c3d93150e8dcd0e02d48f63d3a31
  • Pointer size: 129 Bytes
  • Size of remote file: 9.14 kB
parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux.md ADDED
@@ -0,0 +1,351 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # DETECTING ADVERSARIAL EXAMPLES IS (NEARLY) AS HARD AS CLASSIFYING THEM
2
+
3
+ Anonymous authors Paper under double-blind review
4
+
5
+ # ABSTRACT
6
+
7
+ Making classifiers robust to adversarial examples is challenging. Thus, many defenses tackle the seemingly easier task of detecting perturbed inputs.
8
+
9
+ We show a barrier towards this goal. We prove a general hardness reduction between detection and classification of adversarial examples: given a robust detector for attacks at distance $\epsilon$ (in some metric), we show how to build a similarly robust (but inefficient) classifier for attacks at distance $\epsilon / 2$ —and vice-versa.
10
+
11
+ Our reduction is computationally inefficient, and thus cannot be used to build practical classifiers. Instead, it is a useful sanity check to test whether empirical detection results imply something much stronger than the authors presumably anticipated.
12
+
13
+ To illustrate, we revisit 14 empirical detector defenses published over the past years. For $1 2 / 1 4$ defenses, we show that the claimed detection results imply an inefficient classifier with robustness far beyond the state-of-the-art— thus casting some doubts on the results’ validity.
14
+
15
+ Finally, we show that our reduction applies in both directions: a robust classifier for attacks at distance $\epsilon / 2$ implies an inefficient robust detector at distance . Thus, we argue that robust classification and robust detection should be regarded as (near)-equivalent problems, if we disregard their computational complexity.
16
+
17
+ # 1 INTRODUCTION
18
+
19
+ Building models that are robust to adversarial examples (Szegedy et al., 2014; Biggio et al., 2013) is a major challenge and open-problem in machine learning. Due to the inherent difficulty in building robust classifiers, researchers have attempted to build techniques to at least detect adversarial examples, a weaker task that is largely considered easier than robust classification (Xu et al., 2018; Pang et al., 2021; Sheikholeslami et al., 2021).
20
+
21
+ Yet, evaluating the robustness of empirical detector defenses is challenging. This is in part due to a lack of strong evaluation guidelines and benchmarks—akin to those developed for robust classifiers (Carlini et al., 2019; Croce et al., 2020)—as well as to a lack of long-standing comparative baselines such as adversarial training (Madry et al., 2018).
22
+
23
+ To illustrate, consider the following (fictitious) claims about two defenses against adversarial examples on CIFAR-10:
24
+
25
+ • defense A is a classifier that achieves robust accuracy of $9 0 \%$ under $\ell _ { \infty }$ -perturbations bounded by $\epsilon = { } ^ { 4 } / 2 5 5$ ;
26
+ • defense B also has a “rejection” option, and achieves robust accuracy of $9 0 \%$ under $\ell _ { \infty }$ - perturbations bounded by $\epsilon = { \bf 8 } / 2 5 5$ (we say that defense B is robust for some example if it classifies that example correctly, and either rejects/detects or correctly classifies all perturbed examples at distance $\epsilon .$ .)
27
+
28
+ Which of these two (empirical) claims are you more likely to believe to be correct?
29
+
30
+ Defense A claims much higher robustness than the current best result achieved with adversarial training (Madry et al., 2018; Rebuffi et al., 2021), the only empirical defense against adversarial examples that has stood the test of time. Indeed, the state-of-the-art $\ell _ { \infty }$ robustness for $\epsilon = { ^ 4 } / { 2 5 } 5$ on CIFAR-10 (without external data) is $\approx 7 9 \%$ (Rebuffi et al., 2021). Thus, the claim of defense A would likely be met with some initial skepticism and heightened scrutiny, as could be expected for such a claimed breakthrough result.
31
+
32
+ The claim of defense B is harder to assess, due to a lack of long-standing baselines for robust detectors (many detection defenses have been shown to be broken (Carlini & Wagner, 2017; Tramer\` et al., 2020)). On one hand, detection of adversarial examples has largely been considered to be an easier task than classification $\mathrm { { X u } }$ et al., 2018; Pang et al., 2021; Sheikholeslami et al., 2021). On the other hand, defense B claims robustness to perturbations that are twice as large as defense A ${ \bf \zeta } ( \epsilon = 8 / 2 5 5$ vs. $\epsilon = 4 / 2 5 5$ ).
33
+
34
+ In this paper, we show that the claims of defenses A and B are, in fact, equivalent! (up to computational efficiency.)
35
+
36
+ We prove a general hardness reduction between classification and detection of adversarial examples. Given a detector defense that achieves robust risk $\alpha$ for attacks at distance $\epsilon$ (under any metric), we show how to build an explicit but inefficient classifier that achieves robust risk $\alpha$ for classifying attacks at distance $\epsilon / 2$ . The reverse implication also holds: a classifier robust at distance $\epsilon / 2$ implies an explicit but inefficient robust detector at distance $\epsilon$ .
37
+
38
+ To the authors knowledge, there is no known way of leveraging computational inefficiency to build more robust models. We should thus be as “surprised” by the claim made by defense B as by the claim made by defense A.
39
+
40
+ Our reduction provides a way of assessing the plausibility of new robust detection claims, by contrasting them with results from the more mature literature on robust classification. To illustrate, we revisit 14 published detection defenses across three datasets, and show that in 12/14 cases the defense’s robust detection claims would imply an inefficient classifier with robustness far superior to the current state-of-the-art. Yet, none of these detection papers make the claim that their techniques should imply such a breakthrough in robust classification.
41
+
42
+ Using our reduction, it is obvious that many detection defenses are claiming much stronger robustness than we believe feasible with current techniques. And indeed, many of these defenses were later shown to have overestimated their robustness (Carlini & Wagner, 2017; Tramer et al., 2020). \`
43
+
44
+ Remarkably, we find that for certified defenses, the state-of-the-art results for provable robust classification and detection perfectly match the results implied by our reduction. For example, Sheikholeslami et al. (2021) recently proposed a certified detector on CIFAR-10 with provable robust error that is within $3 \%$ of the provable error of the inefficient detector obtained by combining our result with the state-of-the-art robust classifier of Zhang et al. (2020a).
45
+
46
+ In summary, we prove that giving classifiers access to a detection option does not help robustness (or at least, not much). Our work provides, to our knowledge, the first example of a hardness reduction between different approaches for robust machine learning. As in the case of computational complexity, we believe that such reductions can be useful for identifying research questions or areas that are unlikely to bear fruit (bar a significant breakthrough)—so that the majority of the community’s efforts can be redirected elsewhere.
47
+
48
+ On a technical level, our reduction exposes a natural connection between robustness and error correcting codes, which may be of independent interest.
49
+
50
+ # 2 HARDNESS REDUCTIONS BETWEEN ROBUST CLASSIFIERS AND DETECTORS
51
+
52
+ In this section, we prove our main result: a reduction between robust detectors and robust classifiers, and vice-versa. We first introduce some useful notation and define the (robust) risk of classifiers with and without a detection option.
53
+
54
+ # 2.1 PRELIMINARIES
55
+
56
+ We consider a classification task with a distribution $\mathcal { D }$ over examples $x \in \mathbb { R } ^ { d }$ with labels $y \in$ $[ C ]$ . A classifier is a function $f : \mathbb { R } ^ { d } [ C ]$ . A detector is a classifier with an extra “rejection”
57
+
58
+ or ”detection” option $\perp$ , that indicates the absence of a classification. We assume for simplicity that classifiers and detectors are deterministic. Our results can easily be extended to randomized functions as well. The binary indicator function $\mathbb { 1 } _ { \{ A \} }$ is 1 if and only if the predicate $A$ is true.
59
+
60
+ We first define a classifier’s risk, i.e., its classification error on unperturbed samples.
61
+
62
+ Definition 1 (Risk). Let $f : \mathbb { R } ^ { d } [ C ] \cup \{ \bot \}$ be a classifier (optionally with a detection output $\perp$ ). The risk of $f$ is the expected rate at which $f$ fails to correctly classify a sample:
63
+
64
+ $$
65
+ R ( f ) : = \operatorname* { \mathbb { E } } _ { ( x , y ) \sim \mathcal { D } } \left[ \mathbb { 1 } _ { \{ f ( x ) \neq y \} } \right]
66
+ $$
67
+
68
+ Note that for a detector, rejecting an unperturbed example sampled from the distribution $\mathcal { D }$ is counted as an error.
69
+
70
+ For classifiers without a rejection option, we define the robust risk as the risk on worst-case adversarial examples (Madry et al., 2018). Given an input $x$ sampled from $\mathcal { D }$ , an adversarial example $\hat { x }$ is constrained to being within distance $d ( x , { \hat { x } } ) \leq { \bar { \epsilon } }$ from $x$ , where $d$ is some distance measure.
71
+
72
+ Definition 2 (Robust risk). Let $f : \mathbb { R } ^ { d } [ C ]$ be a classifier. The robust risk at distance $\epsilon$ is:
73
+
74
+ $$
75
+ R _ { a d \nu } ^ { \epsilon } ( f ) : = \operatorname* { \mathbb { E } } _ { ( x , y ) \sim \mathcal { D } } \left[ \operatorname* { m a x } _ { d ( x , \hat { x } ) \leq \epsilon } \mathbb { 1 } _ { \{ f ( \hat { x } ) \neq y \} } \right]
76
+ $$
77
+
78
+ Thus, a sample $( x , y )$ is robustly classified if and only if every point within distance $\epsilon$ of $x$ (including $x$ itself) is correctly classified as $y$ .
79
+
80
+ For a detector (a classifier with an extra detection/rejection output), we analogously define the robust risk with detection. The classifier is now allowed to reject adversarial examples.
81
+
82
+ Definition 3 (Robust risk with detection). Let $f : \mathbb { R } ^ { d } [ C ] \cup \{ \bot \}$ be a classifier with an extra detection output $\perp$ . The robust risk with detection at distance $\epsilon$ is:
83
+
84
+ $$
85
+ R _ { a d \nu - d e t } ^ { \epsilon } ( f ) : = \underset { ( x , y ) \sim \mathcal { D } } { \mathbb { E } } \left[ \operatorname* { m a x } _ { d ( x , \hat { x } ) \leq \epsilon } \mathbb { 1 } _ { \left\{ f ( x ) \neq y \vee f ( \hat { x } ) \notin \left\{ y , \perp \right\} \right\} } \right]
86
+ $$
87
+
88
+ That is, a detector defense $f$ is robust on a natural input $x$ if and only if the defense classifies the natural input $x$ correctly, and the defense either rejects or correctly classifies every perturbed input $\hat { x }$ within distance $\epsilon$ from $x$ . The requirement that the defense correctly classify natural examples eliminates pathological defenses that reject all inputs.
89
+
90
+ # 2.2 ROBUST DETECTION IMPLIES INEFFICIENT ROBUST CLASSIFICATION
91
+
92
+ We are now ready to introduce our main result, a reduction from a robust detector for adversarial examples at distance $\epsilon$ , to an inefficient robust classifier at distance $\epsilon / 2$ . We later prove that this reduction also holds in the reverse direction, thereby demonstrating the equivalence between robust detection and classification—up to computational hardness.
93
+
94
+ Theorem 4 $\epsilon$ -robust detection implies inefficient $\epsilon / 2$ -robust classification). Let $d ( \cdot , \cdot )$ be an arbitrary metric. Let $f$ be a detector that achieves risk $R ( f ) = \alpha _ { \mathrm { { } } }$ , and robust risk with detection $R _ { a d \nu - d e t } ^ { \epsilon } ( f ) = \beta$ . Then, we can construct an explicit (but inefficient) classifier $g$ that achieves risk $R ( g ) \leq \alpha$ and robust risk $R _ { a d \nu } ^ { \epsilon / 2 } ( g ) \le \beta$ .
95
+
96
+ The classifier $g$ is constructed as follows on input $x$ :
97
+
98
+ • Run the detector model $y f ( x )$ . If the input is not rejected, i.e., $y \ne \perp$ , then output the label y that was predicted by the detector.
99
+
100
+ • Otherwise, find an input $x ^ { \prime }$ within distance $\epsilon / 2$ of $x$ that is not rejected, i.e., $d ( x , x ^ { \prime } ) \leq \epsilon / 2$ and $f ( x ^ { \prime } ) \neq \bot$ . If such an input $x ^ { \prime }$ exists, output the label $y \gets f ( x ^ { \prime } )$ . Else, output a uniformly random label $y \in [ C ]$ .
101
+
102
+ An intuitive illustration for our construction, and for the proof of the theorem (see below) is in Figure 1.
103
+
104
+ ![](images/1703fa013d6a1ee6b00378a016c2a93fc0e3073c9d6734ac06a5bcc5e5e8f741.jpg)
105
+ Figure 1: Illustration of the construction of a robust classifier from a robust detector in Theorem 4. The outer blue circle represents all inputs at distance at most $\epsilon$ from the input $x$ . For a detector $f$ , the areas in green correspond to correctly classified inputs, and ratcheted gray areas correspond to rejected inputs. The detector $f$ is thus robust on $x$ up to distance $\epsilon$ . The classifier $g$ classifies a perturbed input $\hat { x }$ , at distance $\epsilon / 2$ from $x$ , by finding any input within distance $\epsilon / 2$ from $\hat { x }$ (the red dashed circle) that is not rejected by $f$ . Such an input necessarily exists and is correctly labeled by $f$ . The classifier $g$ is thus robust on $x$ up to distance $\epsilon / 2$ .
106
+
107
+ Our construction can be viewed as an analog of minimum distance decoding in coding theory. We can view a clean data point sampled from $\mathcal { D }$ as a codeword, and an adversarial example $\hat { x }$ as a noisy message with a certain number of errors (where the error magnitude is measured using an arbitrary metric on $\mathbb { R } ^ { d }$ rather than the Hamming distance that is typically used for error correcting codes). A standard result in coding theory states that if a code can detect $\alpha$ errors, then it can correct $\alpha / 2$ errors. This result follows from a “ball-packing” argument: if $\alpha$ errors can be detected, then any two valid codewords must be at least at distance $\alpha$ from each other, and therefore $\alpha / 2$ errors can be corrected via minimum distance decoding.
108
+
109
+ Proof of Theorem 4. First, note that the natural accuracy of our constructed classifier $g$ is at least as high as that of the detector $f$ , since $g$ always mimics the output of $f$ whenever $f$ does not reject an input sampled from $\mathcal { D }$ . Thus, $R ( g ) \dot { \leq } R ( f ) = \alpha$ .
110
+
111
+ Now, for the sake of contradiction, consider an input $( x , y ) \sim \mathcal { D }$ for which the constructed classifier $g$ is not robust at distance $\epsilon / 2$ . By construction, this means that there exists some input $\hat { x }$ at distance $\epsilon / 2$ from $x$ such that $\hat { x }$ is misclassified, i.e., $g ( \hat { x } ) = \hat { y } \neq y$ . We will show that the detector $f$ is not robust with detection for $x$ either (for attacks at distance up to $\epsilon$ ).
112
+
113
+ By definition of the classifier $g$ , if $g ( \hat { x } ) = \hat { y } \neq y$ then either:
114
+
115
+ • The detector $f$ also misclassifies $\hat { x }$ , i.e., $f ( \hat { x } ) = \hat { y }$ . So $f$ is not robust with detection for $x$ at distance $\epsilon$ .
116
+ • There exists an input $x ^ { \prime }$ within distance $\epsilon / 2$ of $x$ , such that the detector $f$ misclassifies $x ^ { \prime }$ , i.e. $\boldsymbol { f } ( \boldsymbol { x } ^ { \prime } ) = \boldsymbol { \hat { y } }$ . Note that by the triangular inequality, $d ( x , x ^ { \prime } ) \leq d ( x , \hat { x } ) + d ( \hat { x } , x ^ { \prime } ) \leq \epsilon / 2 + \epsilon / 2 = \epsilon ,$ and thus $f$ is not robust with detection for $x$ at distance $\epsilon$ .
117
+ • The detector $f$ rejects all inputs $x ^ { \prime }$ within distance $\epsilon / 2$ of $x$ (and thus $g$ has output $\hat { y }$ by sampling a label at random). Since $d ( x , { \hat { x } } ) \leq \epsilon / 2$ , this implies that the detector also rejects the clean input $x$ , i.e., $f ( x ) = \bot$ , and thus $f$ is not robust with detection for $x$ .
118
+
119
+ In summary, whenever the constructed classifier $g$ fails to robustly classify an input $x$ up to distance $\epsilon / 2$ , the detector $f$ also fails to robustly classify $x$ with detection up to distance $\epsilon$ . Taking expectations over the entire distribution $\mathcal { D }$ concludes the proof. □
120
+
121
+ Note that the classifier $g$ constructed in Theorem 4 is computationally inefficient. Indeed, the second step of the defense consists in finding a non-rejected input within some metric ball. If the original detector $f$ is a non-convex function (e.g., a deep neural network), then this step consists in solving an intractable non-convex optimization problem. Our reduction is thus typically not suitable for building a practical robust classifier. Instead, it demonstrates the existence of an inefficient but explicit robust classifier. We discuss the implications of this result more thoroughly in Section 3.
122
+
123
+ A corollary to our reduction is that many “information theoretic” results about robust classifiers can
124
+ be directly extended to robust detectors. For example, Tsipras et al. (2019) prove that there exists a
125
+ formal tradeoff between a classifier’s clean accuracy and robust accuracy for certain natural tasks.
126
+ Since their result applies to any classifier (including inefficient ones), combining their result with our
127
+ reduction implies that a similar accuracy-robustness tradeoff exists for detectors. More precisely,
128
+ Tsipras et al. (2019) show that for certain classification tasks and suitable choices of parameters
129
+ $\alpha , \beta , \epsilon$ , any classifier $g$ which achieves risk $R ( g ) \leq \alpha$ must have robust risk at least $R _ { \mathrm { a d v } } ^ { \bar { \epsilon } } ( g ) \geq \beta$ inst k at $\ell _ { \infty }$ -t s boumust d by o hav $\epsilon$ . By our reduction, this implies t robust risk with detection at least $f$ withainst $R ( f ) \leq \alpha$ $R _ { \mathrm { a d v - d e t } } ^ { \epsilon / 2 } ( f ) \geq \beta$
130
+ $\ell _ { \infty }$ $\epsilon / 2$
131
+
132
+ Similar arguments can be applied to show, for instance, that the increased data complexity of robust generalization from Schmidt et al. (2018), or the tradeoff between robustness to multiple perturbation types from Tramer & Boneh (2019), also apply to robust detectors. \`
133
+
134
+ Our reduction does not apply for “computational” hardness results that have been shown for robust classification. For example, Garg et al. (2020) and Bubeck et al. (2018) show (“unnatural”) distributions where learning a robust classifier is computationally hard—under standard cryptographic assumptions. We cannot use Theorem 4 to conclude that learning a robust detector is hard for these distributions, since the existence of such a detector would only imply an inefficient robust classifier which does not contradict the results of Garg et al. (2020) or Bubeck et al. (2018).
135
+
136
+ # 2.3 ROBUST CLASSIFICATION IMPLIES INEFFICIENT ROBUST DETECTION
137
+
138
+ A similar argument as in Theorem 4 can be used in the opposite direction, to show that a robust classifier at distance $\epsilon / 2$ implies an inefficient robust detector at distance $\epsilon$ . Taken together, Theorem 4 and Theorem 5 show that robust detection and classification are equivalent, up to a factor 2 in the norm bound and up to computational constraints.
139
+
140
+ Theorem 5 $( \epsilon / 2$ robust-classification implies inefficient $\epsilon$ -robust detection). Let $d ( \cdot , \cdot )$ be an arbi
141
+ trary metric. Let g be a defense than explicit (but inefficient) defense achieves robust rithat achieves risk $R _ { a d \nu } ^ { \epsilon / 2 } ( f ) = \beta$ . Then, we can constructrobust risk with detection $f$ $R ( f ) \le \beta$
142
+ $R _ { a d \nu - d e t } ^ { \epsilon } ( f ) \leq \beta$ .
143
+
144
+ The defense $f$ is constructed as follows on input $x$ :
145
+
146
+ • Run the classifier $y g ( x )$ .
147
+
148
+ • Find a perturbed input $x ^ { \prime }$ withing distance $\epsilon / 2$ of $x$ that is classified differently, i.e., $d ( x , x ^ { \prime } ) \leq \epsilon / 2$ and $g ( x ^ { \prime } ) \neq y$ . If such an input $x ^ { \prime }$ exists, reject the input and output ⊥. Else, output the class $y$ .
149
+
150
+ We provide the proof of Theorem 5 in Appendix A.
151
+
152
+ A main distinction between Theorem 4 and Theorem 5 is that the construction in Theorem 4 preserves clean accuracy, but the construction in Theorem 5 does not. That is, the constructed robust detector in Theorem 5 has clean accuracy that is equal to the robust classifier’s robust accuracy.
153
+
154
+ The construction in Theorem 5 can be efficiently (but approximately) instantiated by a certifiably robust classifier (Wong & Kolter, 2018; Raghunathan et al., 2018). These defenses can certify that a classifier’s output is constant for all points within some distance $\epsilon$ of the input. For an adversarial example $\hat { x }$ for $g$ , the certification always fails and thus the constructed detector $f$ will reject $\hat { x }$ . If $g$ is robust and the certification succeeds, the detector $f$ copies the output of $g$ . However, a certified defense may fail to certify a robust input (a false negative), and thus the detector $f$ may reject more inputs than with the “optimal” construction in Theorem 5. This reduction from a certified classifier to a detector is implicit in (Wong & Kolter, 2018, Section 3.1).
155
+
156
+ # 3 WHAT ARE DETECTION DEFENSES CLAIMING?
157
+
158
+ We now survey 14 detection defenses, and consider the robust classification performance that these defenses implicitly claim (via Theorem 4). As we will see, in 12/14 cases, the defenses’ detection results imply an inefficient classifier with far better robust accuracy than the state-of-the-art.
159
+
160
+ Before presenting our experimental setup and the explicit results from the reduction, we first discuss how we believe these results should be interpreted.
161
+
162
+ Interpreting our reduction. Suppose that some detector defense claims a robust accuracy that implies—via our reduction—an inefficient classifier with much higher robustness that the state-ofthe-art (e.g., the defense A described in the introduction of this paper).
163
+
164
+ A first possible interpretation of our reduction is that this robust detector implies the existence of a robust classifier. This interpretation is rather weak however, since it is typically presumed that robust classification is possible, and that human perception is one concrete example of a robust classifier. The mere existence of a robust classifier is thus typically already assumed to be true.
165
+
166
+ Our reduction yields a stronger result. It provides an explicit construction of an (inefficient) robust classifier from a robust detector. The question then is whether we should expect the construction of inefficient robust classifiers to be easier than the construction of efficient ones. That is, do we expect that we can leverage computational inefficiency to build more robust classifiers that the current state-of-the-art?
167
+
168
+ We do not know of a positive answer to this question, and there is evidence to suggest that the answer may be negative.1 For example, the work of Schmidt et al. (2018) proves that for a synthetic classification task between Gaussian distributions, building more robust classifiers requires additional data regardless of the amount of computation power. Their results are corroborated by current state-ofthe-art robust classifiers based on adversarial training (Madry et al., 2018), which do not appear to be limited by computational constraints. On CIFAR-10 for example, adversarial training achieves $100 \%$ robust training accuracy (Schmidt et al., 2018). Thus, it is unclear how computational inefficiency could be leveraged to build more robust classifiers using existing techniques.
169
+
170
+ Candidate approaches could be to train much larger models (e.g., with an exponential number of parameters), or to perform an exhaustive architecture search to find more robust models. Yet, note that the robust classifier constructed in our reduction only uses its unbounded computational power at inference time. That is, the classifier that is built in Theorem 4 uses a trained detector model as a subroutine (which is presumed to be efficient), and then solves a non-convex optimization problem at inference time. The classifier built in our reduction is thus presumably weaker than a robust classifier that can be trained with unbounded computational power.
171
+
172
+ To summarize, when a detector defense claims a certain robust accuracy, this implies the existence of a concretely instantiatable robust classifier with an inefficient inference procedure. If this inefficient classifier is much more robust than the current state-of-the-art, this does not necessarily mean that the defense’s claims is wrong. But given how challenging robust classification is proving to be, we have reason to be skeptical of such a major breakthrough (even for inefficient classifiers). To compound this, many proposed detection defenses are quite simple, and reject adversarial inputs based on some standard statistical test over a neural network’s features. It would thus be particularly surprising if such simple techniques could yield robust classifiers, given that “simple” approaches to adversarial robustness (denoising, compression, randomness, etc.) are ineffective (He et al., 2017).
173
+
174
+ As a result, it is not too surprising that a number of the detector defenses that we survey have already been broken by stronger attacks (Carlini & Wagner, 2017; Tramer et al., 2020). Our reduction would \` have already suggested that such a break was likely to happen.
175
+
176
+ Experimental setup. We choose 14 detector defenses from the literature (see Table 1). Our selection of these defenses was partially motivated by a pragmatic consideration on the easiness of translating the defenses’ claims into a bound on the robust risk with detection $R _ { \mathrm { a d v - d e t } } ^ { \epsilon }$ . Indeed, some defenses simply report a single AUC score for the detector’s performance, from which we cannot derive a useful bound on the robust risk. We thus focus on defenses that either directly report a robust error akin to Definition 3, or that provide concrete pairs of false-positive and false-negative rates (e.g., a full ROC curve). In the latter case, we compute a “best-effort” bound on the robust risk with detection2 as:
177
+
178
+ $$
179
+ R _ { \mathrm { a d v - d e t } } ^ { \epsilon } ( f ) \leq \mathrm { F P R } + \mathrm { F N R } + R ( f ) ,
180
+ $$
181
+
182
+ where FPR and FNR are the detector’s false-positive and false-negative rates for a fixed detection threshold, and $R ( f )$ is the defense’s standard risk (i.e., the test error on natural examples).
183
+
184
+ The above union bound in Equation (4) is quite pessimistic, as we may over-count examples that lead to multiple sources of errors (e.g., a natural input that is misclassified and erroneously detected). The true robustness claim made by these detector defenses might thus be stronger than what we obtain from our bound. We encourage future detection papers to report their adversarial risk with detection, $R _ { \mathrm { a d v - d e t } } ^ { \epsilon }$ , to facilitate direct comparisons with robust classifiers using our reduction.
185
+
186
+ The 14 detector defenses use three datasets: MNIST, CIFAR-10 and ImageNet, and consider adversarial examples under the $\ell _ { \infty }$ or $\ell _ { 2 }$ norms. Given a claim of robust detection at distance , we contrast it to a state-of-the-art robust classification result for distance $\epsilon / 2$ :
187
+
188
+ • On MNIST with $\ell _ { \infty }$ attacks, we use the adversarially-trained TRADES classifier (Zhang et al., 2019) and measure robust error with the Square attack (Andriushchenko et al., 2020).
189
+ • On MNIST with $\ell _ { 2 }$ attacks, we use the adversarially-trained classifier from Tramer & \` Boneh (2019) and measure robust error with PGD (Madry et al., 2018).
190
+ • On CIFAR-10, for both $\ell _ { \infty }$ and $\ell _ { 2 }$ attacks we use the adversarially-trained classifier of Rebuffi et al. (2021) (trained without external data), and attack it using the APGD-CE attack from AutoAttack (Croce & Hein, 2020).
191
+ • For ImageNet, for both $\ell _ { \infty }$ and $\ell _ { 2 }$ attacks we use adversarially-trained classifiers and PGD attacks from Engstrom et al. (2019).
192
+
193
+ We also consider two certified defenses for $\ell _ { \infty }$ attacks on CIFAR-10: the robust classifier of Zhang et al. (2020a), and a recent certified detector of Sheikholeslami et al. (2021).
194
+
195
+ Results. As we can see from Table 1, most defenses claim a detection performance that implies a far greater robust accuracy than our current best robust classifiers. To illustrate with a concrete example, the CIFAR-10 detector of Miller et al. (2019) claims to achieve robust accuracy with detection of $7 5 \%$ for $\ell _ { 2 }$ attacks with $\epsilon = 2 . 9$ . Using Theorem 4, this implies an inefficient classifier with robust accuracy of $7 5 \%$ for $\ell _ { 2 }$ attacks with $\epsilon = { } ^ { 2 . 9 } / 2 = 1 . 4 5$ . Yet, the current state-of-the-art robust accuracy for such a perturbation budget is only $3 0 \%$ (Rebuffi et al., 2021). If this detector defense’s robustness claim were correct, it would imply a remarkable breakthrough in robust classification.
196
+
197
+ Why do many of these defenses claim robust accuracies that appear “too good to be true”? A primary reason is that the vast majority of the above detector defenses do not consider evaluations against adaptive attacks (Carlini et al., 2019; Athalye et al., 2018; Tramer et al., 2020). That is, \` these defenses show that they can detect some fixed attacks, and thereafter conclude that the detector is robust against all attacks. As in the case of robust classifiers, such an evaluation is clearly insufficient! Some defenses do evaluate against adaptive adversaries, but fail to build a sufficiently strong attack to reliably approximate the worst-case robust risk. Because of the lack of a strong comparative baseline, it is not always immediately clear that these results are overly strong.
198
+
199
+ Table 1: For each detector defense, we compute a (best-effort) bound on the claimed robust risk with detection $R _ { \mathrm { a d v - d e t } } ^ { \epsilon }$ using Equation (4), and report the complement (the robust accuracy with detection), $1 - R _ { \mathrm { a d v - d e t } } ^ { \epsilon }$ . For each detector’s robustness claim (at distance $\epsilon$ ), we report the stateof-the-art robust classification accuracy for attacks at distance $\epsilon / 2$ , denoted $1 - R _ { \mathrm { a d v } } ^ { \epsilon / 2 }$ . Detection defense claims that imply a higher robust classification accuracy than the current state-of-the-art are highlighted in red.
200
+
201
+ <table><tr><td>Dataset</td><td>Defense</td><td>Norm</td><td>E</td><td>1-Radv-det</td><td>1- R</td></tr><tr><td rowspan="3">MNIST</td><td>Grosse et al. (2017)</td><td>l8</td><td>0.5</td><td>≥ 98%</td><td>94%</td></tr><tr><td>Ma et al. (2018)</td><td>l2</td><td>4.2</td><td>≥ 99%</td><td>72%</td></tr><tr><td>Raghuram et al. (2021)</td><td>l2</td><td>8.9</td><td>≥ 74%</td><td>0%</td></tr><tr><td rowspan="8">CIFAR-10</td><td>Yin et al. (2020)</td><td>l2</td><td>1.7</td><td>≥ 90%</td><td>66%</td></tr><tr><td>Feinman et al. (2017)</td><td>l2</td><td>2.7</td><td>≥ 43%</td><td>36%</td></tr><tr><td>Miller et al. (2019)</td><td>l2</td><td>2.9</td><td>≥ 75%</td><td>30%</td></tr><tr><td>Raghuram et al. (2021)</td><td>l</td><td>4.0</td><td>≥ 56%</td><td>10%</td></tr><tr><td>Ma &amp; Liu (2019)</td><td>l8</td><td>4/255</td><td>≥ 96%</td><td>85%</td></tr><tr><td>Roth et al. (2019)</td><td>l8</td><td>8/255</td><td>≥ 66%</td><td>79%</td></tr><tr><td>Lee et al. (2018)</td><td>lo</td><td>20/255</td><td>≥ 81%</td><td>59%</td></tr><tr><td>Li et al. (2019)</td><td>lo</td><td>26/255</td><td>≥ 80%</td><td>44%</td></tr><tr><td rowspan="5">ImageNet</td><td>Xu et al. (2018)</td><td>l</td><td>1.0</td><td>≥ 67%</td><td>54%</td></tr><tr><td>Ma &amp; Liu (2019)</td><td>l8</td><td>2/255</td><td>≥ 68%</td><td>55%</td></tr><tr><td>Jha et al. (2019)</td><td>lo</td><td>2/255</td><td>≥ 30%</td><td>55%</td></tr><tr><td>Hendrycks &amp; Gimpel (2017)</td><td>l8</td><td>10/255</td><td>≥ 76%</td><td>30%</td></tr><tr><td>Yu et al. (2019)</td><td>l8</td><td>26/255</td><td>M 7%</td><td>5%</td></tr></table>
202
+
203
+ For example, the recent work of Raghuram et al. (2021, ICML Long Talk) builds a detector on MNIST with a FNR of $\leq 5 \%$ at a FPR of $\leq 2 0 \%$ , for adaptive $\ell _ { 2 }$ attacks bounded by $\epsilon = 8 . 9$ . Yet, this perturbation bound is much larger than the average distance between an MNIST image and the nearest image from a different class! Thus, an attack within this perturbation bound can trivially reduce the detector’s accuracy to chance. On CIFAR-10, the same detector achieves $9 5 \%$ clean accuracy, and a FNR of $\leq 1 9 \%$ at a FPR of $\leq 2 0 \%$ for adaptive $\ell _ { 2 }$ attacks bounded by $\epsilon = 4$ . Using Equation (4), this yields a bound on the robust accuracy with detection of $1 - R _ { \mathrm { a d v - d e t } } ^ { \epsilon } ( f ) \geq$ $1 - \bar { ( 5 \% + 1 9 \% + 2 0 \% ) } = 5 6 \%$ . In contrast, the best robust classifier we are aware of for $\ell _ { 2 }$ attacks bounded by $\epsilon = 2$ achieves robust accuracy of only $1 0 \%$ (Rebuffi et al., 2021). In summary, the adaptive attack considered in this detector defense’s evaluation is highly unlikely to be good approximation of a worst-case attack, and this defense can likely be broken by stronger attacks.
204
+
205
+ Certifiably robust detection and classification. In Table 2, we look at the robust accuracy with detection, and standard robust accuracy achieved by certified defenses (for which the claimed robustness numbers are necessarily mathematically correct).
206
+
207
+ We note that our reduction is not as meaningful in the case of certified defenses, since it is highly plausible that computational inefficiency can be leveraged to build better certified classifiers. Indeed, given any robust classifier (e.g., an adversarially trained model), the classifier’s robustness can always be certified inefficiently (by enumerating over all points within an $\epsilon$ -ball). Thus, the existence of an inefficient classifier with higher certified robustness than the state-of-the-art is to be expected.
208
+
209
+ Nevertheless, we find that existing results for certified classifiers and detectors perfectly match what is implied by our reduction (up to $\pm 2 \%$ error). For example, Zhang et al. (2020a) follow a long line of results on robust classifiers and achieve $3 9 \%$ robust accuracy on CIFAR-10 for perturbations of $\ell _ { \infty }$ -norm below $^ { 4 } / 2 5 5$ . Together with Theorem 5, this implies an inefficient detector with $3 9 \%$ robust detection accuracy for perturbations of $\ell _ { \infty }$ -norm below $^ { 8 } / 2 5 5$ . The recent work of Sheikholeslami et al. (2021) nearly matches that bound $3 7 \%$ robust accuracy with detection), with a defense that has the advantage of being concretely efficient.
210
+
211
+ Table 2: Certified robust accuracy $1 - R _ { \mathrm { a d v } } ^ { \epsilon / 2 }$ for the defense of Zhang et al. (2020a), and certified robust accuracy with detection $1 - R _ { \mathrm { a d v - d e t } } ^ { \epsilon }$ dv for the defense of Sheikholeslami et al. (2021).
212
+
213
+ <table><tr><td>E</td><td>1-Radv-det</td><td></td></tr><tr><td>8/255</td><td>37%</td><td>39%</td></tr><tr><td>16/255</td><td>32%</td><td>33%</td></tr></table>
214
+
215
+ These results give additional credence to our thesis: with current techniques, robust classification is indeed approximately twice as hard (in terms of the perturbation bounds covered) than robust detection.
216
+
217
+ Extensions and open problems. The main open problem raised by our work is of course whether it could be possible to show an efficient reduction between classification and detection of adversarial examples, but this seems implausible (at least with our minimum distance decoding approach).
218
+
219
+ Another interesting question is whether a similar reduction can be shown for robustness to less “structured” perturbations than $\ell _ { p }$ balls and other metric spaces. For example, there has been a line of research on defending against adversarial patches (Brown et al., 2017), using empirical (Hayes, 2018; Naseer et al., 2019; Chou et al., 2020) and certifiable techniques (Chiang et al., 2020; Zhang et al., 2020b; Xiang et al., 2021). To use our result, we would have to define some metric to measure the size of an adversarial patch’s perturbation. Yet, the size of a patch is typically defined by the number of contiguously perturbed pixels, which does not define a metric (in particular, it does not satisfy the triangular inequality which our reduction relies on).
220
+
221
+ Finally, similar hardness reductions might exist between other candidate approaches for building robust classifiers. For example, the question of whether (test-time) randomness can be leveraged to build more robust models is also intriguing .Empirical defenses that use randomness can be notoriously hard to evaluate (Athalye et al., 2018; Tramer et al., 2020), so a reduction similar to ours \` might be useful in showing that we should not expect such approaches to bare fruit.
222
+
223
+ # 4 CONCLUSION
224
+
225
+ We have shown formal reductions between robust classification with, and without, a detection option. Our results show that significant progress on one of these two tasks implies similar progress on the other—unless computational inefficiency can somehow be leveraged to build more robust models. This raises the question on whether we should spend our efforts on studying both of these tasks, or focus our efforts on a single one.
226
+
227
+ On one hand, the two tasks represent different ways of tackling a common goal, and working on either task might result in new techniques or ideas that apply to the other task as well. On the other hand, our reductions show that unless we make progress on both tasks, work on one of the tasks can merely aim to match the robustness of our inefficient constructions, whilst improving their computational complexity.
228
+
229
+ We believe our reduction will serve as a useful sanity-check when assessing the claims of future detector defenses. Detector defenses’ robustness evaluations have received less stringent scrutiny than robust classifiers over the past years, perhaps in part due to a lack of strong comparative baselines. Instead of having to wait until some detector defense’s claims pass the test-of-time, we show that detection results can be directly contrasted against long-standing results for robust classification.
230
+
231
+ When applying this approach to past detector defenses, we find that many make robustness claims that imply significant breakthroughs in robust classification. We believe our reduction could have been useful in highlighting the suspiciously strong claims made by many of these defenses—before they were explicitly broken by stronger attacks.
232
+
233
+ Ethics statement. Our paper demonstrates a fundamental barrier towards detecting against adversarial examples, under the assumption that our current techniques are insufficient to achieve strong (inefficient) robust classification. We do not however explicitly break any existing defenses (our results merely strongly suggest that many existing detector defenses’ claims are suspiciously high). Our paper therefore cannot lead to any explicit harms, but aims to further our understanding of the hardness of robust classification and detection.
234
+
235
+ Reproducibility statement. Our paper’s contribution is mainly of theoretical nature. Section 2 is self-contained and clearly states our assumptions, results and proofs (except for the proof of Theorem 5 in Appendix A). The experiments in Section 3 use only public datasets and pre-trained models, with clearly indicates hyper-parameters for all attacks that we evaluate.
236
+
237
+ # REFERENCES
238
+
239
+ Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, 2020.
240
+
241
+ Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, 2018.
242
+
243
+ Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndi ˇ c, Pavel Laskov, Gior- ´ gio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 387–402. Springer, 2013.
244
+
245
+ Tom B Brown, Dandelion Mane, Aurko Roy, Mart ´ ´ın Abadi, and Justin Gilmer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2017.
246
+
247
+ Sebastien Bubeck, Yin Tat Lee, Eric Price, and Ilya Razenshteyn. Adversarial examples from cryp-´ tographic pseudo-random generators. arXiv preprint arXiv:1811.06418, 2018.
248
+
249
+ Nicholas Carlini and David Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In AISec, pp. 3–14. ACM, 2017.
250
+
251
+ Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, and Aleksander Madry. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705, 2019.
252
+
253
+ Ping-yeh Chiang, Renkun Ni, Ahmed Abdelkader, Chen Zhu, Christoph Studer, and Tom Goldstein. Certified defenses for adversarial patches. arXiv preprint arXiv:2003.06693, 2020.
254
+
255
+ Edward Chou, Florian Tramer, and Giancarlo Pellegrino. Sentinet: Detecting physical attacks \` against deep learning systems. In Workshop on Deep Learning Security, 2020.
256
+
257
+ Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020.
258
+
259
+ Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020.
260
+
261
+ Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. URL https://github.com/MadryLab/robustness.
262
+
263
+ Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017.
264
+
265
+ Sanjam Garg, Somesh Jha, Saeed Mahloujifar, and Mahmoody Mohammad. Adversarially robust learning could leverage computational hardness. In Algorithmic Learning Theory, pp. 364–385. PMLR, 2020.
266
+
267
+ Kathrin Grosse, Praveen Manoharan, Nicolas Papernot, Michael Backes, and Patrick McDaniel. On the (statistical) detection of adversarial examples. arXiv preprint arXiv:1702.06280, 2017.
268
+
269
+ Jamie Hayes. On visible adversarial perturbations & digital watermarking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 1597–1604, 2018.
270
+
271
+ Warren He, James Wei, Xinyun Chen, Nicholas Carlini, and Dawn Song. Adversarial example defenses: Ensembles of weak defenses are not strong. In USENIX Workshop on Offensive Technologies, 2017.
272
+
273
+ Dan Hendrycks and Kevin Gimpel. Early methods for detecting adversarial images. In International Conference on Learning Representations, 2017.
274
+
275
+ Susmit Jha, Sunny Raj, Steven Lawrence Fernandes, Sumit Kumar Jha, Somesh Jha, Gunjan Verma, Brian Jalaian, and Ananthram Swami. Attribution-driven causal analysis for detection of adversarial examples. arXiv preprint arXiv:1903.05821, 2019.
276
+
277
+ Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems, 2018.
278
+
279
+ Yingzhen Li, John Bradshaw, and Yash Sharma. Are generative classifiers more robust to adversarial attacks? In International Conference on Machine Learning, pp. 3804–3814. PMLR, 2019.
280
+
281
+ Shiqing Ma and Yingqi Liu. Nic: Detecting adversarial samples with neural network invariant checking. In Proceedings of the 26th Network and Distributed System Security Symposium (NDSS 2019), 2019.
282
+
283
+ Xingjun Ma, Bo Li, Yisen Wang, Sarah M Erfani, Sudanthi Wijewickrema, Grant Schoenebeck, Dawn Song, Michael E Houle, and James Bailey. Characterizing adversarial subspaces using local intrinsic dimensionality. In International Conference on Learning Representations, 2018.
284
+
285
+ Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018.
286
+
287
+ David Miller, Yujia Wang, and George Kesidis. When not to classify: Anomaly detection of attacks (ada) on dnn classifiers at test time. Neural computation, 31(8):1624–1670, 2019.
288
+
289
+ Muzammal Naseer, Salman Khan, and Fatih Porikli. Local gradients smoothing: Defense against localized adversarial attacks. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1300–1307. IEEE, 2019.
290
+
291
+ Tianyu Pang, Huishuai Zhang, Di He, Yinpeng Dong, Hang Su, Wei Chen, Jun Zhu, and Tie-Yan Liu. Adversarial training with rectified rejection. arXiv preprint arXiv:2105.14785, 2021.
292
+
293
+ Aditi Raghunathan, Jacob Steinhardt, and Percy Liang. Certified defenses against adversarial examples. In International Conference on Learning Representations, 2018.
294
+
295
+ Jayaram Raghuram, Varun Chandrasekaran, Somesh Jha, and Suman Banerjee. A general framework for detecting anomalous inputs to dnn classifiers. In International Conference on Machine Learning, 2021.
296
+
297
+ Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021.
298
+
299
+ Kevin Roth, Yannic Kilcher, and Thomas Hofmann. The odds are odd: A statistical test for detecting adversarial examples. In International Conference on Machine Learning, 2019.
300
+
301
+ Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarially robust generalization requires more data. In Advances In Neural Information Processing Systems, pp. 5019–5031, 2018.
302
+
303
+ Fatemeh Sheikholeslami, Ali Lotfi, and J Zico Kolter. Provably robust classification of adversarial examples with detection. In International Conference on Learning Representations, 2021.
304
+
305
+ Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014.
306
+
307
+ Florian Tramer and Dan Boneh. Adversarial training and robustness for multiple perturbations. In \` Advances In Neural Information Processing Systems, 2019.
308
+
309
+ Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to \` adversarial example defenses. In Conference on Neural Information Processing Systems, 2020.
310
+
311
+ Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In International Conference on Learning Representations, 2019.
312
+
313
+ Eric Wong and Zico Kolter. Provable defenses against adversarial examples via the convex outer adversarial polytope. In International Conference on Machine Learning, pp. 5283–5292, 2018.
314
+
315
+ Chong Xiang, Arjun Nitin Bhagoji, Vikash Sehwag, and Prateek Mittal. Patchguard: A provably robust defense against adversarial patches via small receptive fields and masking. In 30th {USENIX} Security Symposium ({USENIX} Security 21), 2021.
316
+
317
+ Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In Network and Distributed System Security Symposium, 2018.
318
+
319
+ Xuwang Yin, Soheil Kolouri, and Gustavo K Rohde. Gat: Generative adversarial training for adversarial example detection and robust classification. In International Conference on Learning Representations, 2020.
320
+
321
+ Tao Yu, Shengyuan Hu, Chuan Guo, Wei-Lun Chao, and Kilian Q Weinberger. A new defense against adversarial images: Turning a weakness into a strength. In Advances in Neural Information Processing Systems, 2019.
322
+
323
+ Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P. Xing, Laurent El Ghaoui, and Michael I. Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning, 2019.
324
+
325
+ Huan Zhang, Hongge Chen, Chaowei Xiao, Bo Li, Duane Boning, and Cho-Jui Hsieh. Towards stable and efficient training of verifiably robust neural networks. In International Conference on Learning Representations, 2020a.
326
+
327
+ Zhanyuan Zhang, Benson Yuan, Michael McCoyd, and David Wagner. Clipped bagnet: Defending against sticker attacks with clipped bag-of-features. In 2020 IEEE Security and Privacy Workshops (SPW), pp. 55–61. IEEE, 2020b.
328
+
329
+ # A PROOF OF THEOREM 5.
330
+
331
+ We recall Theorem 5:
332
+
333
+ Theorem 5 $( \epsilon / 2$ robust-classification implies inefficient $\epsilon$ -robust detection). Let $d ( \cdot , \cdot )$ be an arbi
334
+ trary metric. Let g be a defense than explicit (but inefficient) defense achieves robust rithat achieves risk $R _ { a d \nu } ^ { \epsilon / 2 } ( f ) = \beta$ . Then, we can constructrobust risk with detection $f$ $R ( f ) \leq \beta$
335
+ $R _ { a d \nu - d e t } ^ { \epsilon } ( f ) \leq \beta$ .
336
+
337
+ The defense $f$ is constructed as follows on input $x$ :
338
+
339
+ • Run the classifier $y g ( x )$ .
340
+
341
+ • Find a perturbed input $x ^ { \prime }$ withing distance $\epsilon / 2$ of $x$ that is classified differently, i.e., $\bar { d } ( x , x ^ { \prime } ) \bar { \ } \leq \ \epsilon / 2$ and $g ( x ^ { \prime } ) \neq y$ . If such an input $x ^ { \prime }$ exists, reject the input and output ⊥. Else, output the class $y$ .
342
+
343
+ Proof of Theorem 5. Note that for any input $( x , y )$ for which the classifier $g$ is robust at distance $\epsilon / 2$ , no input $x ^ { \prime }$ above exists and so $f ( x ) = y$ . Thus, the risk of $f$ is at most the robust risk of $g$ , so $R ( f ) \leq \beta$ .
344
+
345
+ Now, consider an input $( x , y ) \sim \mathcal { D }$ for which $f$ is not robust with detection at distance $\epsilon$ . That is, either $f ( x ) \neq y$ , or there exists an input $\hat { x }$ at distance $d ( x , { \hat { x } } ) \leq \epsilon$ such that $f ( \hat { x } ) = \hat { y } \not \in \{ y , \bot \}$ . We will show that the defense $g$ is not robust for $x$ either (for attacks at distance up to $\epsilon / 2$ .)
346
+
347
+ If $f ( x ) \neq y$ , then by the same argument as above it cannot be the case that $g$ is robust at distance $\epsilon / 2$ for $x$ .
348
+
349
+ So let us consider the case where $f ( \hat { x } ) = \hat { y } \not \in \{ y , \bot \}$ . By the definition of $f$ , this means that for all $x ^ { \prime }$ at distance at most $\epsilon / 2$ from $\hat { x }$ , we have $g ( x ^ { \prime } ) = \hat { y }$ . But, note that there exists a point $x ^ { * }$ that is at distance at most $\epsilon / 2$ from both $\hat { x }$ and $x$ . Since we must have $g ( x ^ { * } ) = \hat { y }$ , we conclude that $g$ is not robust at distance $\epsilon / 2$ for $x$ .
350
+
351
+ Taking expectations over the distribution $\mathcal { D }$ concludes the proof.
parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_content_list.json ADDED
@@ -0,0 +1,1839 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "DETECTING ADVERSARIAL EXAMPLES IS (NEARLY) AS HARD AS CLASSIFYING THEM ",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 176,
8
+ 99,
9
+ 823,
10
+ 146
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Anonymous authors Paper under double-blind review ",
17
+ "bbox": [
18
+ 183,
19
+ 171,
20
+ 398,
21
+ 198
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "ABSTRACT ",
28
+ "text_level": 1,
29
+ "bbox": [
30
+ 454,
31
+ 234,
32
+ 544,
33
+ 251
34
+ ],
35
+ "page_idx": 0
36
+ },
37
+ {
38
+ "type": "text",
39
+ "text": "Making classifiers robust to adversarial examples is challenging. Thus, many defenses tackle the seemingly easier task of detecting perturbed inputs. ",
40
+ "bbox": [
41
+ 232,
42
+ 266,
43
+ 763,
44
+ 294
45
+ ],
46
+ "page_idx": 0
47
+ },
48
+ {
49
+ "type": "text",
50
+ "text": "We show a barrier towards this goal. We prove a general hardness reduction between detection and classification of adversarial examples: given a robust detector for attacks at distance $\\epsilon$ (in some metric), we show how to build a similarly robust (but inefficient) classifier for attacks at distance $\\epsilon / 2$ —and vice-versa. ",
51
+ "bbox": [
52
+ 232,
53
+ 296,
54
+ 764,
55
+ 352
56
+ ],
57
+ "page_idx": 0
58
+ },
59
+ {
60
+ "type": "text",
61
+ "text": "Our reduction is computationally inefficient, and thus cannot be used to build practical classifiers. Instead, it is a useful sanity check to test whether empirical detection results imply something much stronger than the authors presumably anticipated. ",
62
+ "bbox": [
63
+ 233,
64
+ 353,
65
+ 764,
66
+ 410
67
+ ],
68
+ "page_idx": 0
69
+ },
70
+ {
71
+ "type": "text",
72
+ "text": "To illustrate, we revisit 14 empirical detector defenses published over the past years. For $1 2 / 1 4$ defenses, we show that the claimed detection results imply an inefficient classifier with robustness far beyond the state-of-the-art— thus casting some doubts on the results’ validity. ",
73
+ "bbox": [
74
+ 233,
75
+ 411,
76
+ 764,
77
+ 467
78
+ ],
79
+ "page_idx": 0
80
+ },
81
+ {
82
+ "type": "text",
83
+ "text": "Finally, we show that our reduction applies in both directions: a robust classifier for attacks at distance $\\epsilon / 2$ implies an inefficient robust detector at distance \u000f. Thus, we argue that robust classification and robust detection should be regarded as (near)-equivalent problems, if we disregard their computational complexity. ",
84
+ "bbox": [
85
+ 232,
86
+ 469,
87
+ 764,
88
+ 525
89
+ ],
90
+ "page_idx": 0
91
+ },
92
+ {
93
+ "type": "text",
94
+ "text": "1 INTRODUCTION ",
95
+ "text_level": 1,
96
+ "bbox": [
97
+ 176,
98
+ 550,
99
+ 336,
100
+ 565
101
+ ],
102
+ "page_idx": 0
103
+ },
104
+ {
105
+ "type": "text",
106
+ "text": "Building models that are robust to adversarial examples (Szegedy et al., 2014; Biggio et al., 2013) is a major challenge and open-problem in machine learning. Due to the inherent difficulty in building robust classifiers, researchers have attempted to build techniques to at least detect adversarial examples, a weaker task that is largely considered easier than robust classification (Xu et al., 2018; Pang et al., 2021; Sheikholeslami et al., 2021). ",
107
+ "bbox": [
108
+ 174,
109
+ 582,
110
+ 825,
111
+ 651
112
+ ],
113
+ "page_idx": 0
114
+ },
115
+ {
116
+ "type": "text",
117
+ "text": "Yet, evaluating the robustness of empirical detector defenses is challenging. This is in part due to a lack of strong evaluation guidelines and benchmarks—akin to those developed for robust classifiers (Carlini et al., 2019; Croce et al., 2020)—as well as to a lack of long-standing comparative baselines such as adversarial training (Madry et al., 2018). ",
118
+ "bbox": [
119
+ 174,
120
+ 657,
121
+ 825,
122
+ 714
123
+ ],
124
+ "page_idx": 0
125
+ },
126
+ {
127
+ "type": "text",
128
+ "text": "To illustrate, consider the following (fictitious) claims about two defenses against adversarial examples on CIFAR-10: ",
129
+ "bbox": [
130
+ 173,
131
+ 720,
132
+ 821,
133
+ 750
134
+ ],
135
+ "page_idx": 0
136
+ },
137
+ {
138
+ "type": "text",
139
+ "text": "• defense A is a classifier that achieves robust accuracy of $9 0 \\%$ under $\\ell _ { \\infty }$ -perturbations bounded by $\\epsilon = { } ^ { 4 } / 2 5 5$ ; \n• defense B also has a “rejection” option, and achieves robust accuracy of $9 0 \\%$ under $\\ell _ { \\infty }$ - perturbations bounded by $\\epsilon = { \\bf 8 } / 2 5 5$ (we say that defense B is robust for some example if it classifies that example correctly, and either rejects/detects or correctly classifies all perturbed examples at distance $\\epsilon .$ .) ",
140
+ "bbox": [
141
+ 218,
142
+ 760,
143
+ 825,
144
+ 849
145
+ ],
146
+ "page_idx": 0
147
+ },
148
+ {
149
+ "type": "text",
150
+ "text": "Which of these two (empirical) claims are you more likely to believe to be correct? ",
151
+ "bbox": [
152
+ 174,
153
+ 859,
154
+ 714,
155
+ 875
156
+ ],
157
+ "page_idx": 0
158
+ },
159
+ {
160
+ "type": "text",
161
+ "text": "Defense A claims much higher robustness than the current best result achieved with adversarial training (Madry et al., 2018; Rebuffi et al., 2021), the only empirical defense against adversarial examples that has stood the test of time. Indeed, the state-of-the-art $\\ell _ { \\infty }$ robustness for $\\epsilon = { ^ 4 } / { 2 5 } 5$ on CIFAR-10 (without external data) is $\\approx 7 9 \\%$ (Rebuffi et al., 2021). Thus, the claim of defense A would likely be met with some initial skepticism and heightened scrutiny, as could be expected for such a claimed breakthrough result. ",
162
+ "bbox": [
163
+ 176,
164
+ 882,
165
+ 825,
166
+ 922
167
+ ],
168
+ "page_idx": 0
169
+ },
170
+ {
171
+ "type": "text",
172
+ "text": "",
173
+ "bbox": [
174
+ 176,
175
+ 103,
176
+ 821,
177
+ 145
178
+ ],
179
+ "page_idx": 1
180
+ },
181
+ {
182
+ "type": "text",
183
+ "text": "The claim of defense B is harder to assess, due to a lack of long-standing baselines for robust detectors (many detection defenses have been shown to be broken (Carlini & Wagner, 2017; Tramer\\` et al., 2020)). On one hand, detection of adversarial examples has largely been considered to be an easier task than classification $\\mathrm { { X u } }$ et al., 2018; Pang et al., 2021; Sheikholeslami et al., 2021). On the other hand, defense B claims robustness to perturbations that are twice as large as defense A ${ \\bf \\zeta } ( \\epsilon = 8 / 2 5 5$ vs. $\\epsilon = 4 / 2 5 5$ ). ",
184
+ "bbox": [
185
+ 174,
186
+ 152,
187
+ 825,
188
+ 237
189
+ ],
190
+ "page_idx": 1
191
+ },
192
+ {
193
+ "type": "text",
194
+ "text": "In this paper, we show that the claims of defenses A and B are, in fact, equivalent! (up to computational efficiency.) ",
195
+ "bbox": [
196
+ 176,
197
+ 243,
198
+ 821,
199
+ 271
200
+ ],
201
+ "page_idx": 1
202
+ },
203
+ {
204
+ "type": "text",
205
+ "text": "We prove a general hardness reduction between classification and detection of adversarial examples. Given a detector defense that achieves robust risk $\\alpha$ for attacks at distance $\\epsilon$ (under any metric), we show how to build an explicit but inefficient classifier that achieves robust risk $\\alpha$ for classifying attacks at distance $\\epsilon / 2$ . The reverse implication also holds: a classifier robust at distance $\\epsilon / 2$ implies an explicit but inefficient robust detector at distance $\\epsilon$ . ",
206
+ "bbox": [
207
+ 174,
208
+ 279,
209
+ 825,
210
+ 348
211
+ ],
212
+ "page_idx": 1
213
+ },
214
+ {
215
+ "type": "text",
216
+ "text": "To the authors knowledge, there is no known way of leveraging computational inefficiency to build more robust models. We should thus be as “surprised” by the claim made by defense B as by the claim made by defense A. ",
217
+ "bbox": [
218
+ 178,
219
+ 356,
220
+ 821,
221
+ 397
222
+ ],
223
+ "page_idx": 1
224
+ },
225
+ {
226
+ "type": "text",
227
+ "text": "Our reduction provides a way of assessing the plausibility of new robust detection claims, by contrasting them with results from the more mature literature on robust classification. To illustrate, we revisit 14 published detection defenses across three datasets, and show that in 12/14 cases the defense’s robust detection claims would imply an inefficient classifier with robustness far superior to the current state-of-the-art. Yet, none of these detection papers make the claim that their techniques should imply such a breakthrough in robust classification. ",
228
+ "bbox": [
229
+ 174,
230
+ 404,
231
+ 823,
232
+ 488
233
+ ],
234
+ "page_idx": 1
235
+ },
236
+ {
237
+ "type": "text",
238
+ "text": "Using our reduction, it is obvious that many detection defenses are claiming much stronger robustness than we believe feasible with current techniques. And indeed, many of these defenses were later shown to have overestimated their robustness (Carlini & Wagner, 2017; Tramer et al., 2020). \\` ",
239
+ "bbox": [
240
+ 174,
241
+ 494,
242
+ 825,
243
+ 536
244
+ ],
245
+ "page_idx": 1
246
+ },
247
+ {
248
+ "type": "text",
249
+ "text": "Remarkably, we find that for certified defenses, the state-of-the-art results for provable robust classification and detection perfectly match the results implied by our reduction. For example, Sheikholeslami et al. (2021) recently proposed a certified detector on CIFAR-10 with provable robust error that is within $3 \\%$ of the provable error of the inefficient detector obtained by combining our result with the state-of-the-art robust classifier of Zhang et al. (2020a). ",
250
+ "bbox": [
251
+ 174,
252
+ 544,
253
+ 823,
254
+ 613
255
+ ],
256
+ "page_idx": 1
257
+ },
258
+ {
259
+ "type": "text",
260
+ "text": "In summary, we prove that giving classifiers access to a detection option does not help robustness (or at least, not much). Our work provides, to our knowledge, the first example of a hardness reduction between different approaches for robust machine learning. As in the case of computational complexity, we believe that such reductions can be useful for identifying research questions or areas that are unlikely to bear fruit (bar a significant breakthrough)—so that the majority of the community’s efforts can be redirected elsewhere. ",
261
+ "bbox": [
262
+ 174,
263
+ 621,
264
+ 825,
265
+ 704
266
+ ],
267
+ "page_idx": 1
268
+ },
269
+ {
270
+ "type": "text",
271
+ "text": "On a technical level, our reduction exposes a natural connection between robustness and error correcting codes, which may be of independent interest. ",
272
+ "bbox": [
273
+ 174,
274
+ 712,
275
+ 823,
276
+ 739
277
+ ],
278
+ "page_idx": 1
279
+ },
280
+ {
281
+ "type": "text",
282
+ "text": "2 HARDNESS REDUCTIONS BETWEEN ROBUST CLASSIFIERS AND DETECTORS ",
283
+ "text_level": 1,
284
+ "bbox": [
285
+ 174,
286
+ 760,
287
+ 733,
288
+ 794
289
+ ],
290
+ "page_idx": 1
291
+ },
292
+ {
293
+ "type": "text",
294
+ "text": "In this section, we prove our main result: a reduction between robust detectors and robust classifiers, and vice-versa. We first introduce some useful notation and define the (robust) risk of classifiers with and without a detection option. ",
295
+ "bbox": [
296
+ 176,
297
+ 809,
298
+ 823,
299
+ 852
300
+ ],
301
+ "page_idx": 1
302
+ },
303
+ {
304
+ "type": "text",
305
+ "text": "2.1 PRELIMINARIES ",
306
+ "text_level": 1,
307
+ "bbox": [
308
+ 174,
309
+ 869,
310
+ 326,
311
+ 883
312
+ ],
313
+ "page_idx": 1
314
+ },
315
+ {
316
+ "type": "text",
317
+ "text": "We consider a classification task with a distribution $\\mathcal { D }$ over examples $x \\in \\mathbb { R } ^ { d }$ with labels $y \\in$ $[ C ]$ . A classifier is a function $f : \\mathbb { R } ^ { d } [ C ]$ . A detector is a classifier with an extra “rejection” ",
318
+ "bbox": [
319
+ 174,
320
+ 895,
321
+ 821,
322
+ 924
323
+ ],
324
+ "page_idx": 1
325
+ },
326
+ {
327
+ "type": "text",
328
+ "text": "or ”detection” option $\\perp$ , that indicates the absence of a classification. We assume for simplicity that classifiers and detectors are deterministic. Our results can easily be extended to randomized functions as well. The binary indicator function $\\mathbb { 1 } _ { \\{ A \\} }$ is 1 if and only if the predicate $A$ is true. ",
329
+ "bbox": [
330
+ 173,
331
+ 103,
332
+ 825,
333
+ 147
334
+ ],
335
+ "page_idx": 2
336
+ },
337
+ {
338
+ "type": "text",
339
+ "text": "We first define a classifier’s risk, i.e., its classification error on unperturbed samples. ",
340
+ "bbox": [
341
+ 173,
342
+ 152,
343
+ 723,
344
+ 167
345
+ ],
346
+ "page_idx": 2
347
+ },
348
+ {
349
+ "type": "text",
350
+ "text": "Definition 1 (Risk). Let $f : \\mathbb { R } ^ { d } [ C ] \\cup \\{ \\bot \\}$ be a classifier (optionally with a detection output $\\perp$ ). The risk of $f$ is the expected rate at which $f$ fails to correctly classify a sample: ",
351
+ "bbox": [
352
+ 173,
353
+ 171,
354
+ 825,
355
+ 202
356
+ ],
357
+ "page_idx": 2
358
+ },
359
+ {
360
+ "type": "equation",
361
+ "img_path": "images/44c657694cd74a56c1870443ece67a7d806cd61222c843330a77a072ca61eca5.jpg",
362
+ "text": "$$\nR ( f ) : = \\operatorname* { \\mathbb { E } } _ { ( x , y ) \\sim \\mathcal { D } } \\left[ \\mathbb { 1 } _ { \\{ f ( x ) \\neq y \\} } \\right]\n$$",
363
+ "text_format": "latex",
364
+ "bbox": [
365
+ 400,
366
+ 207,
367
+ 596,
368
+ 233
369
+ ],
370
+ "page_idx": 2
371
+ },
372
+ {
373
+ "type": "text",
374
+ "text": "Note that for a detector, rejecting an unperturbed example sampled from the distribution $\\mathcal { D }$ is counted as an error. ",
375
+ "bbox": [
376
+ 176,
377
+ 248,
378
+ 826,
379
+ 277
380
+ ],
381
+ "page_idx": 2
382
+ },
383
+ {
384
+ "type": "text",
385
+ "text": "For classifiers without a rejection option, we define the robust risk as the risk on worst-case adversarial examples (Madry et al., 2018). Given an input $x$ sampled from $\\mathcal { D }$ , an adversarial example $\\hat { x }$ is constrained to being within distance $d ( x , { \\hat { x } } ) \\leq { \\bar { \\epsilon } }$ from $x$ , where $d$ is some distance measure. ",
386
+ "bbox": [
387
+ 174,
388
+ 284,
389
+ 825,
390
+ 327
391
+ ],
392
+ "page_idx": 2
393
+ },
394
+ {
395
+ "type": "text",
396
+ "text": "Definition 2 (Robust risk). Let $f : \\mathbb { R } ^ { d } [ C ]$ be a classifier. The robust risk at distance $\\epsilon$ is: ",
397
+ "bbox": [
398
+ 173,
399
+ 330,
400
+ 779,
401
+ 347
402
+ ],
403
+ "page_idx": 2
404
+ },
405
+ {
406
+ "type": "equation",
407
+ "img_path": "images/bf3313c2eebf9b088d68940d3fcdf78503af5aea78a5151d35b998f8ac2a42af.jpg",
408
+ "text": "$$\nR _ { a d \\nu } ^ { \\epsilon } ( f ) : = \\operatorname* { \\mathbb { E } } _ { ( x , y ) \\sim \\mathcal { D } } \\left[ \\operatorname* { m a x } _ { d ( x , \\hat { x } ) \\leq \\epsilon } \\mathbb { 1 } _ { \\{ f ( \\hat { x } ) \\neq y \\} } \\right]\n$$",
409
+ "text_format": "latex",
410
+ "bbox": [
411
+ 364,
412
+ 353,
413
+ 633,
414
+ 388
415
+ ],
416
+ "page_idx": 2
417
+ },
418
+ {
419
+ "type": "text",
420
+ "text": "Thus, a sample $( x , y )$ is robustly classified if and only if every point within distance $\\epsilon$ of $x$ (including $x$ itself) is correctly classified as $y$ . ",
421
+ "bbox": [
422
+ 171,
423
+ 404,
424
+ 825,
425
+ 433
426
+ ],
427
+ "page_idx": 2
428
+ },
429
+ {
430
+ "type": "text",
431
+ "text": "For a detector (a classifier with an extra detection/rejection output), we analogously define the robust risk with detection. The classifier is now allowed to reject adversarial examples. ",
432
+ "bbox": [
433
+ 173,
434
+ 439,
435
+ 823,
436
+ 467
437
+ ],
438
+ "page_idx": 2
439
+ },
440
+ {
441
+ "type": "text",
442
+ "text": "Definition 3 (Robust risk with detection). Let $f : \\mathbb { R } ^ { d } [ C ] \\cup \\{ \\bot \\}$ be a classifier with an extra detection output $\\perp$ . The robust risk with detection at distance $\\epsilon$ is: ",
443
+ "bbox": [
444
+ 173,
445
+ 470,
446
+ 821,
447
+ 501
448
+ ],
449
+ "page_idx": 2
450
+ },
451
+ {
452
+ "type": "equation",
453
+ "img_path": "images/280c3197d6ce35502a31b2d1bc26122b4c360c6fa70eabe111b6e26a291e0bec.jpg",
454
+ "text": "$$\nR _ { a d \\nu - d e t } ^ { \\epsilon } ( f ) : = \\underset { ( x , y ) \\sim \\mathcal { D } } { \\mathbb { E } } \\left[ \\operatorname* { m a x } _ { d ( x , \\hat { x } ) \\leq \\epsilon } \\mathbb { 1 } _ { \\left\\{ f ( x ) \\neq y \\vee f ( \\hat { x } ) \\notin \\left\\{ y , \\perp \\right\\} \\right\\} } \\right]\n$$",
455
+ "text_format": "latex",
456
+ "bbox": [
457
+ 313,
458
+ 508,
459
+ 678,
460
+ 542
461
+ ],
462
+ "page_idx": 2
463
+ },
464
+ {
465
+ "type": "text",
466
+ "text": "That is, a detector defense $f$ is robust on a natural input $x$ if and only if the defense classifies the natural input $x$ correctly, and the defense either rejects or correctly classifies every perturbed input $\\hat { x }$ within distance $\\epsilon$ from $x$ . The requirement that the defense correctly classify natural examples eliminates pathological defenses that reject all inputs. ",
467
+ "bbox": [
468
+ 174,
469
+ 556,
470
+ 825,
471
+ 614
472
+ ],
473
+ "page_idx": 2
474
+ },
475
+ {
476
+ "type": "text",
477
+ "text": "2.2 ROBUST DETECTION IMPLIES INEFFICIENT ROBUST CLASSIFICATION ",
478
+ "text_level": 1,
479
+ "bbox": [
480
+ 176,
481
+ 631,
482
+ 694,
483
+ 646
484
+ ],
485
+ "page_idx": 2
486
+ },
487
+ {
488
+ "type": "text",
489
+ "text": "We are now ready to introduce our main result, a reduction from a robust detector for adversarial examples at distance $\\epsilon$ , to an inefficient robust classifier at distance $\\epsilon / 2$ . We later prove that this reduction also holds in the reverse direction, thereby demonstrating the equivalence between robust detection and classification—up to computational hardness. ",
490
+ "bbox": [
491
+ 173,
492
+ 657,
493
+ 825,
494
+ 714
495
+ ],
496
+ "page_idx": 2
497
+ },
498
+ {
499
+ "type": "text",
500
+ "text": "Theorem 4 $\\epsilon$ -robust detection implies inefficient $\\epsilon / 2$ -robust classification). Let $d ( \\cdot , \\cdot )$ be an arbitrary metric. Let $f$ be a detector that achieves risk $R ( f ) = \\alpha _ { \\mathrm { { } } }$ , and robust risk with detection $R _ { a d \\nu - d e t } ^ { \\epsilon } ( f ) = \\beta$ . Then, we can construct an explicit (but inefficient) classifier $g$ that achieves risk $R ( g ) \\leq \\alpha$ and robust risk $R _ { a d \\nu } ^ { \\epsilon / 2 } ( g ) \\le \\beta$ . ",
501
+ "bbox": [
502
+ 173,
503
+ 718,
504
+ 825,
505
+ 781
506
+ ],
507
+ "page_idx": 2
508
+ },
509
+ {
510
+ "type": "text",
511
+ "text": "The classifier $g$ is constructed as follows on input $x$ : ",
512
+ "bbox": [
513
+ 174,
514
+ 785,
515
+ 517,
516
+ 801
517
+ ],
518
+ "page_idx": 2
519
+ },
520
+ {
521
+ "type": "text",
522
+ "text": "• Run the detector model $y f ( x )$ . If the input is not rejected, i.e., $y \\ne \\perp$ , then output the label y that was predicted by the detector. ",
523
+ "bbox": [
524
+ 178,
525
+ 804,
526
+ 825,
527
+ 833
528
+ ],
529
+ "page_idx": 2
530
+ },
531
+ {
532
+ "type": "text",
533
+ "text": "• Otherwise, find an input $x ^ { \\prime }$ within distance $\\epsilon / 2$ of $x$ that is not rejected, i.e., $d ( x , x ^ { \\prime } ) \\leq \\epsilon / 2$ and $f ( x ^ { \\prime } ) \\neq \\bot$ . If such an input $x ^ { \\prime }$ exists, output the label $y \\gets f ( x ^ { \\prime } )$ . Else, output a uniformly random label $y \\in [ C ]$ . ",
534
+ "bbox": [
535
+ 179,
536
+ 839,
537
+ 825,
538
+ 883
539
+ ],
540
+ "page_idx": 2
541
+ },
542
+ {
543
+ "type": "text",
544
+ "text": "An intuitive illustration for our construction, and for the proof of the theorem (see below) is in Figure 1. ",
545
+ "bbox": [
546
+ 173,
547
+ 895,
548
+ 825,
549
+ 924
550
+ ],
551
+ "page_idx": 2
552
+ },
553
+ {
554
+ "type": "image",
555
+ "img_path": "images/1703fa013d6a1ee6b00378a016c2a93fc0e3073c9d6734ac06a5bcc5e5e8f741.jpg",
556
+ "image_caption": [
557
+ "Figure 1: Illustration of the construction of a robust classifier from a robust detector in Theorem 4. The outer blue circle represents all inputs at distance at most $\\epsilon$ from the input $x$ . For a detector $f$ , the areas in green correspond to correctly classified inputs, and ratcheted gray areas correspond to rejected inputs. The detector $f$ is thus robust on $x$ up to distance $\\epsilon$ . The classifier $g$ classifies a perturbed input $\\hat { x }$ , at distance $\\epsilon / 2$ from $x$ , by finding any input within distance $\\epsilon / 2$ from $\\hat { x }$ (the red dashed circle) that is not rejected by $f$ . Such an input necessarily exists and is correctly labeled by $f$ . The classifier $g$ is thus robust on $x$ up to distance $\\epsilon / 2$ . "
558
+ ],
559
+ "image_footnote": [],
560
+ "bbox": [
561
+ 349,
562
+ 99,
563
+ 647,
564
+ 327
565
+ ],
566
+ "page_idx": 3
567
+ },
568
+ {
569
+ "type": "text",
570
+ "text": "Our construction can be viewed as an analog of minimum distance decoding in coding theory. We can view a clean data point sampled from $\\mathcal { D }$ as a codeword, and an adversarial example $\\hat { x }$ as a noisy message with a certain number of errors (where the error magnitude is measured using an arbitrary metric on $\\mathbb { R } ^ { d }$ rather than the Hamming distance that is typically used for error correcting codes). A standard result in coding theory states that if a code can detect $\\alpha$ errors, then it can correct $\\alpha / 2$ errors. This result follows from a “ball-packing” argument: if $\\alpha$ errors can be detected, then any two valid codewords must be at least at distance $\\alpha$ from each other, and therefore $\\alpha / 2$ errors can be corrected via minimum distance decoding. ",
571
+ "bbox": [
572
+ 173,
573
+ 467,
574
+ 825,
575
+ 579
576
+ ],
577
+ "page_idx": 3
578
+ },
579
+ {
580
+ "type": "text",
581
+ "text": "Proof of Theorem 4. First, note that the natural accuracy of our constructed classifier $g$ is at least as high as that of the detector $f$ , since $g$ always mimics the output of $f$ whenever $f$ does not reject an input sampled from $\\mathcal { D }$ . Thus, $R ( g ) \\dot { \\leq } R ( f ) = \\alpha$ . ",
582
+ "bbox": [
583
+ 174,
584
+ 597,
585
+ 825,
586
+ 640
587
+ ],
588
+ "page_idx": 3
589
+ },
590
+ {
591
+ "type": "text",
592
+ "text": "Now, for the sake of contradiction, consider an input $( x , y ) \\sim \\mathcal { D }$ for which the constructed classifier $g$ is not robust at distance $\\epsilon / 2$ . By construction, this means that there exists some input $\\hat { x }$ at distance $\\epsilon / 2$ from $x$ such that $\\hat { x }$ is misclassified, i.e., $g ( \\hat { x } ) = \\hat { y } \\neq y$ . We will show that the detector $f$ is not robust with detection for $x$ either (for attacks at distance up to $\\epsilon$ ). ",
593
+ "bbox": [
594
+ 173,
595
+ 646,
596
+ 825,
597
+ 703
598
+ ],
599
+ "page_idx": 3
600
+ },
601
+ {
602
+ "type": "text",
603
+ "text": "By definition of the classifier $g$ , if $g ( \\hat { x } ) = \\hat { y } \\neq y$ then either: ",
604
+ "bbox": [
605
+ 173,
606
+ 709,
607
+ 568,
608
+ 724
609
+ ],
610
+ "page_idx": 3
611
+ },
612
+ {
613
+ "type": "text",
614
+ "text": "• The detector $f$ also misclassifies $\\hat { x }$ , i.e., $f ( \\hat { x } ) = \\hat { y }$ . So $f$ is not robust with detection for $x$ at distance $\\epsilon$ . \n• There exists an input $x ^ { \\prime }$ within distance $\\epsilon / 2$ of $x$ , such that the detector $f$ misclassifies $x ^ { \\prime }$ , i.e. $\\boldsymbol { f } ( \\boldsymbol { x } ^ { \\prime } ) = \\boldsymbol { \\hat { y } }$ . Note that by the triangular inequality, $d ( x , x ^ { \\prime } ) \\leq d ( x , \\hat { x } ) + d ( \\hat { x } , x ^ { \\prime } ) \\leq \\epsilon / 2 + \\epsilon / 2 = \\epsilon ,$ and thus $f$ is not robust with detection for $x$ at distance $\\epsilon$ . \n• The detector $f$ rejects all inputs $x ^ { \\prime }$ within distance $\\epsilon / 2$ of $x$ (and thus $g$ has output $\\hat { y }$ by sampling a label at random). Since $d ( x , { \\hat { x } } ) \\leq \\epsilon / 2$ , this implies that the detector also rejects the clean input $x$ , i.e., $f ( x ) = \\bot$ , and thus $f$ is not robust with detection for $x$ . ",
615
+ "bbox": [
616
+ 217,
617
+ 739,
618
+ 825,
619
+ 924
620
+ ],
621
+ "page_idx": 3
622
+ },
623
+ {
624
+ "type": "text",
625
+ "text": "In summary, whenever the constructed classifier $g$ fails to robustly classify an input $x$ up to distance $\\epsilon / 2$ , the detector $f$ also fails to robustly classify $x$ with detection up to distance $\\epsilon$ . Taking expectations over the entire distribution $\\mathcal { D }$ concludes the proof. □ ",
626
+ "bbox": [
627
+ 176,
628
+ 103,
629
+ 823,
630
+ 146
631
+ ],
632
+ "page_idx": 4
633
+ },
634
+ {
635
+ "type": "text",
636
+ "text": "Note that the classifier $g$ constructed in Theorem 4 is computationally inefficient. Indeed, the second step of the defense consists in finding a non-rejected input within some metric ball. If the original detector $f$ is a non-convex function (e.g., a deep neural network), then this step consists in solving an intractable non-convex optimization problem. Our reduction is thus typically not suitable for building a practical robust classifier. Instead, it demonstrates the existence of an inefficient but explicit robust classifier. We discuss the implications of this result more thoroughly in Section 3. ",
637
+ "bbox": [
638
+ 174,
639
+ 162,
640
+ 825,
641
+ 246
642
+ ],
643
+ "page_idx": 4
644
+ },
645
+ {
646
+ "type": "text",
647
+ "text": "A corollary to our reduction is that many “information theoretic” results about robust classifiers can \nbe directly extended to robust detectors. For example, Tsipras et al. (2019) prove that there exists a \nformal tradeoff between a classifier’s clean accuracy and robust accuracy for certain natural tasks. \nSince their result applies to any classifier (including inefficient ones), combining their result with our \nreduction implies that a similar accuracy-robustness tradeoff exists for detectors. More precisely, \nTsipras et al. (2019) show that for certain classification tasks and suitable choices of parameters \n$\\alpha , \\beta , \\epsilon$ , any classifier $g$ which achieves risk $R ( g ) \\leq \\alpha$ must have robust risk at least $R _ { \\mathrm { a d v } } ^ { \\bar { \\epsilon } } ( g ) \\geq \\beta$ inst k at $\\ell _ { \\infty }$ -t s boumust d by o hav $\\epsilon$ . By our reduction, this implies t robust risk with detection at least $f$ withainst $R ( f ) \\leq \\alpha$ $R _ { \\mathrm { a d v - d e t } } ^ { \\epsilon / 2 } ( f ) \\geq \\beta$ \n$\\ell _ { \\infty }$ $\\epsilon / 2$ ",
648
+ "bbox": [
649
+ 173,
650
+ 253,
651
+ 825,
652
+ 396
653
+ ],
654
+ "page_idx": 4
655
+ },
656
+ {
657
+ "type": "text",
658
+ "text": "Similar arguments can be applied to show, for instance, that the increased data complexity of robust generalization from Schmidt et al. (2018), or the tradeoff between robustness to multiple perturbation types from Tramer & Boneh (2019), also apply to robust detectors. \\` ",
659
+ "bbox": [
660
+ 176,
661
+ 402,
662
+ 821,
663
+ 445
664
+ ],
665
+ "page_idx": 4
666
+ },
667
+ {
668
+ "type": "text",
669
+ "text": "Our reduction does not apply for “computational” hardness results that have been shown for robust classification. For example, Garg et al. (2020) and Bubeck et al. (2018) show (“unnatural”) distributions where learning a robust classifier is computationally hard—under standard cryptographic assumptions. We cannot use Theorem 4 to conclude that learning a robust detector is hard for these distributions, since the existence of such a detector would only imply an inefficient robust classifier which does not contradict the results of Garg et al. (2020) or Bubeck et al. (2018). ",
670
+ "bbox": [
671
+ 174,
672
+ 452,
673
+ 825,
674
+ 535
675
+ ],
676
+ "page_idx": 4
677
+ },
678
+ {
679
+ "type": "text",
680
+ "text": "2.3 ROBUST CLASSIFICATION IMPLIES INEFFICIENT ROBUST DETECTION ",
681
+ "text_level": 1,
682
+ "bbox": [
683
+ 176,
684
+ 553,
685
+ 694,
686
+ 566
687
+ ],
688
+ "page_idx": 4
689
+ },
690
+ {
691
+ "type": "text",
692
+ "text": "A similar argument as in Theorem 4 can be used in the opposite direction, to show that a robust classifier at distance $\\epsilon / 2$ implies an inefficient robust detector at distance $\\epsilon$ . Taken together, Theorem 4 and Theorem 5 show that robust detection and classification are equivalent, up to a factor 2 in the norm bound and up to computational constraints. ",
693
+ "bbox": [
694
+ 174,
695
+ 579,
696
+ 825,
697
+ 635
698
+ ],
699
+ "page_idx": 4
700
+ },
701
+ {
702
+ "type": "text",
703
+ "text": "Theorem 5 $( \\epsilon / 2$ robust-classification implies inefficient $\\epsilon$ -robust detection). Let $d ( \\cdot , \\cdot )$ be an arbi \ntrary metric. Let g be a defense than explicit (but inefficient) defense achieves robust rithat achieves risk $R _ { a d \\nu } ^ { \\epsilon / 2 } ( f ) = \\beta$ . Then, we can constructrobust risk with detection $f$ $R ( f ) \\le \\beta$ \n$R _ { a d \\nu - d e t } ^ { \\epsilon } ( f ) \\leq \\beta$ . ",
704
+ "bbox": [
705
+ 176,
706
+ 638,
707
+ 825,
708
+ 700
709
+ ],
710
+ "page_idx": 4
711
+ },
712
+ {
713
+ "type": "text",
714
+ "text": "The defense $f$ is constructed as follows on input $x$ : ",
715
+ "bbox": [
716
+ 174,
717
+ 705,
718
+ 509,
719
+ 722
720
+ ],
721
+ "page_idx": 4
722
+ },
723
+ {
724
+ "type": "text",
725
+ "text": "• Run the classifier $y g ( x )$ . ",
726
+ "bbox": [
727
+ 218,
728
+ 732,
729
+ 419,
730
+ 748
731
+ ],
732
+ "page_idx": 4
733
+ },
734
+ {
735
+ "type": "text",
736
+ "text": "• Find a perturbed input $x ^ { \\prime }$ withing distance $\\epsilon / 2$ of $x$ that is classified differently, i.e., $d ( x , x ^ { \\prime } ) \\leq \\epsilon / 2$ and $g ( x ^ { \\prime } ) \\neq y$ . If such an input $x ^ { \\prime }$ exists, reject the input and output ⊥. Else, output the class $y$ . ",
737
+ "bbox": [
738
+ 218,
739
+ 756,
740
+ 823,
741
+ 799
742
+ ],
743
+ "page_idx": 4
744
+ },
745
+ {
746
+ "type": "text",
747
+ "text": "We provide the proof of Theorem 5 in Appendix A. ",
748
+ "bbox": [
749
+ 174,
750
+ 810,
751
+ 511,
752
+ 825
753
+ ],
754
+ "page_idx": 4
755
+ },
756
+ {
757
+ "type": "text",
758
+ "text": "A main distinction between Theorem 4 and Theorem 5 is that the construction in Theorem 4 preserves clean accuracy, but the construction in Theorem 5 does not. That is, the constructed robust detector in Theorem 5 has clean accuracy that is equal to the robust classifier’s robust accuracy. ",
759
+ "bbox": [
760
+ 176,
761
+ 832,
762
+ 825,
763
+ 875
764
+ ],
765
+ "page_idx": 4
766
+ },
767
+ {
768
+ "type": "text",
769
+ "text": "The construction in Theorem 5 can be efficiently (but approximately) instantiated by a certifiably robust classifier (Wong & Kolter, 2018; Raghunathan et al., 2018). These defenses can certify that a classifier’s output is constant for all points within some distance $\\epsilon$ of the input. For an adversarial example $\\hat { x }$ for $g$ , the certification always fails and thus the constructed detector $f$ will reject $\\hat { x }$ . If $g$ is robust and the certification succeeds, the detector $f$ copies the output of $g$ . However, a certified defense may fail to certify a robust input (a false negative), and thus the detector $f$ may reject more inputs than with the “optimal” construction in Theorem 5. This reduction from a certified classifier to a detector is implicit in (Wong & Kolter, 2018, Section 3.1). ",
770
+ "bbox": [
771
+ 174,
772
+ 881,
773
+ 823,
774
+ 924
775
+ ],
776
+ "page_idx": 4
777
+ },
778
+ {
779
+ "type": "text",
780
+ "text": "",
781
+ "bbox": [
782
+ 174,
783
+ 103,
784
+ 825,
785
+ 172
786
+ ],
787
+ "page_idx": 5
788
+ },
789
+ {
790
+ "type": "text",
791
+ "text": "3 WHAT ARE DETECTION DEFENSES CLAIMING? ",
792
+ "text_level": 1,
793
+ "bbox": [
794
+ 176,
795
+ 195,
796
+ 601,
797
+ 210
798
+ ],
799
+ "page_idx": 5
800
+ },
801
+ {
802
+ "type": "text",
803
+ "text": "We now survey 14 detection defenses, and consider the robust classification performance that these defenses implicitly claim (via Theorem 4). As we will see, in 12/14 cases, the defenses’ detection results imply an inefficient classifier with far better robust accuracy than the state-of-the-art. ",
804
+ "bbox": [
805
+ 176,
806
+ 228,
807
+ 823,
808
+ 270
809
+ ],
810
+ "page_idx": 5
811
+ },
812
+ {
813
+ "type": "text",
814
+ "text": "Before presenting our experimental setup and the explicit results from the reduction, we first discuss how we believe these results should be interpreted. ",
815
+ "bbox": [
816
+ 173,
817
+ 277,
818
+ 821,
819
+ 305
820
+ ],
821
+ "page_idx": 5
822
+ },
823
+ {
824
+ "type": "text",
825
+ "text": "Interpreting our reduction. Suppose that some detector defense claims a robust accuracy that implies—via our reduction—an inefficient classifier with much higher robustness that the state-ofthe-art (e.g., the defense A described in the introduction of this paper). ",
826
+ "bbox": [
827
+ 174,
828
+ 323,
829
+ 825,
830
+ 364
831
+ ],
832
+ "page_idx": 5
833
+ },
834
+ {
835
+ "type": "text",
836
+ "text": "A first possible interpretation of our reduction is that this robust detector implies the existence of a robust classifier. This interpretation is rather weak however, since it is typically presumed that robust classification is possible, and that human perception is one concrete example of a robust classifier. The mere existence of a robust classifier is thus typically already assumed to be true. ",
837
+ "bbox": [
838
+ 174,
839
+ 371,
840
+ 823,
841
+ 428
842
+ ],
843
+ "page_idx": 5
844
+ },
845
+ {
846
+ "type": "text",
847
+ "text": "Our reduction yields a stronger result. It provides an explicit construction of an (inefficient) robust classifier from a robust detector. The question then is whether we should expect the construction of inefficient robust classifiers to be easier than the construction of efficient ones. That is, do we expect that we can leverage computational inefficiency to build more robust classifiers that the current state-of-the-art? ",
848
+ "bbox": [
849
+ 174,
850
+ 434,
851
+ 825,
852
+ 503
853
+ ],
854
+ "page_idx": 5
855
+ },
856
+ {
857
+ "type": "text",
858
+ "text": "We do not know of a positive answer to this question, and there is evidence to suggest that the answer may be negative.1 For example, the work of Schmidt et al. (2018) proves that for a synthetic classification task between Gaussian distributions, building more robust classifiers requires additional data regardless of the amount of computation power. Their results are corroborated by current state-ofthe-art robust classifiers based on adversarial training (Madry et al., 2018), which do not appear to be limited by computational constraints. On CIFAR-10 for example, adversarial training achieves $100 \\%$ robust training accuracy (Schmidt et al., 2018). Thus, it is unclear how computational inefficiency could be leveraged to build more robust classifiers using existing techniques. ",
859
+ "bbox": [
860
+ 174,
861
+ 511,
862
+ 825,
863
+ 622
864
+ ],
865
+ "page_idx": 5
866
+ },
867
+ {
868
+ "type": "text",
869
+ "text": "Candidate approaches could be to train much larger models (e.g., with an exponential number of parameters), or to perform an exhaustive architecture search to find more robust models. Yet, note that the robust classifier constructed in our reduction only uses its unbounded computational power at inference time. That is, the classifier that is built in Theorem 4 uses a trained detector model as a subroutine (which is presumed to be efficient), and then solves a non-convex optimization problem at inference time. The classifier built in our reduction is thus presumably weaker than a robust classifier that can be trained with unbounded computational power. ",
870
+ "bbox": [
871
+ 174,
872
+ 630,
873
+ 825,
874
+ 727
875
+ ],
876
+ "page_idx": 5
877
+ },
878
+ {
879
+ "type": "text",
880
+ "text": "To summarize, when a detector defense claims a certain robust accuracy, this implies the existence of a concretely instantiatable robust classifier with an inefficient inference procedure. If this inefficient classifier is much more robust than the current state-of-the-art, this does not necessarily mean that the defense’s claims is wrong. But given how challenging robust classification is proving to be, we have reason to be skeptical of such a major breakthrough (even for inefficient classifiers). To compound this, many proposed detection defenses are quite simple, and reject adversarial inputs based on some standard statistical test over a neural network’s features. It would thus be particularly surprising if such simple techniques could yield robust classifiers, given that “simple” approaches to adversarial robustness (denoising, compression, randomness, etc.) are ineffective (He et al., 2017). ",
881
+ "bbox": [
882
+ 174,
883
+ 733,
884
+ 825,
885
+ 859
886
+ ],
887
+ "page_idx": 5
888
+ },
889
+ {
890
+ "type": "text",
891
+ "text": "As a result, it is not too surprising that a number of the detector defenses that we survey have already been broken by stronger attacks (Carlini & Wagner, 2017; Tramer et al., 2020). Our reduction would \\` have already suggested that such a break was likely to happen. ",
892
+ "bbox": [
893
+ 176,
894
+ 103,
895
+ 823,
896
+ 146
897
+ ],
898
+ "page_idx": 6
899
+ },
900
+ {
901
+ "type": "text",
902
+ "text": "Experimental setup. We choose 14 detector defenses from the literature (see Table 1). Our selection of these defenses was partially motivated by a pragmatic consideration on the easiness of translating the defenses’ claims into a bound on the robust risk with detection $R _ { \\mathrm { a d v - d e t } } ^ { \\epsilon }$ . Indeed, some defenses simply report a single AUC score for the detector’s performance, from which we cannot derive a useful bound on the robust risk. We thus focus on defenses that either directly report a robust error akin to Definition 3, or that provide concrete pairs of false-positive and false-negative rates (e.g., a full ROC curve). In the latter case, we compute a “best-effort” bound on the robust risk with detection2 as: ",
903
+ "bbox": [
904
+ 173,
905
+ 161,
906
+ 825,
907
+ 272
908
+ ],
909
+ "page_idx": 6
910
+ },
911
+ {
912
+ "type": "equation",
913
+ "img_path": "images/8e622d645b3f35c117ff52694ec9773f02e8ea8238d5cfacd531d2db3bd601ae.jpg",
914
+ "text": "$$\nR _ { \\mathrm { a d v - d e t } } ^ { \\epsilon } ( f ) \\leq \\mathrm { F P R } + \\mathrm { F N R } + R ( f ) ,\n$$",
915
+ "text_format": "latex",
916
+ "bbox": [
917
+ 379,
918
+ 272,
919
+ 617,
920
+ 290
921
+ ],
922
+ "page_idx": 6
923
+ },
924
+ {
925
+ "type": "text",
926
+ "text": "where FPR and FNR are the detector’s false-positive and false-negative rates for a fixed detection threshold, and $R ( f )$ is the defense’s standard risk (i.e., the test error on natural examples). ",
927
+ "bbox": [
928
+ 171,
929
+ 301,
930
+ 821,
931
+ 329
932
+ ],
933
+ "page_idx": 6
934
+ },
935
+ {
936
+ "type": "text",
937
+ "text": "The above union bound in Equation (4) is quite pessimistic, as we may over-count examples that lead to multiple sources of errors (e.g., a natural input that is misclassified and erroneously detected). The true robustness claim made by these detector defenses might thus be stronger than what we obtain from our bound. We encourage future detection papers to report their adversarial risk with detection, $R _ { \\mathrm { a d v - d e t } } ^ { \\epsilon }$ , to facilitate direct comparisons with robust classifiers using our reduction. ",
938
+ "bbox": [
939
+ 174,
940
+ 335,
941
+ 825,
942
+ 406
943
+ ],
944
+ "page_idx": 6
945
+ },
946
+ {
947
+ "type": "text",
948
+ "text": "The 14 detector defenses use three datasets: MNIST, CIFAR-10 and ImageNet, and consider adversarial examples under the $\\ell _ { \\infty }$ or $\\ell _ { 2 }$ norms. Given a claim of robust detection at distance \u000f, we contrast it to a state-of-the-art robust classification result for distance $\\epsilon / 2$ : ",
949
+ "bbox": [
950
+ 176,
951
+ 412,
952
+ 825,
953
+ 455
954
+ ],
955
+ "page_idx": 6
956
+ },
957
+ {
958
+ "type": "text",
959
+ "text": "• On MNIST with $\\ell _ { \\infty }$ attacks, we use the adversarially-trained TRADES classifier (Zhang et al., 2019) and measure robust error with the Square attack (Andriushchenko et al., 2020). \n• On MNIST with $\\ell _ { 2 }$ attacks, we use the adversarially-trained classifier from Tramer & \\` Boneh (2019) and measure robust error with PGD (Madry et al., 2018). \n• On CIFAR-10, for both $\\ell _ { \\infty }$ and $\\ell _ { 2 }$ attacks we use the adversarially-trained classifier of Rebuffi et al. (2021) (trained without external data), and attack it using the APGD-CE attack from AutoAttack (Croce & Hein, 2020). \n• For ImageNet, for both $\\ell _ { \\infty }$ and $\\ell _ { 2 }$ attacks we use adversarially-trained classifiers and PGD attacks from Engstrom et al. (2019). ",
960
+ "bbox": [
961
+ 217,
962
+ 468,
963
+ 825,
964
+ 612
965
+ ],
966
+ "page_idx": 6
967
+ },
968
+ {
969
+ "type": "text",
970
+ "text": "We also consider two certified defenses for $\\ell _ { \\infty }$ attacks on CIFAR-10: the robust classifier of Zhang et al. (2020a), and a recent certified detector of Sheikholeslami et al. (2021). ",
971
+ "bbox": [
972
+ 174,
973
+ 626,
974
+ 823,
975
+ 654
976
+ ],
977
+ "page_idx": 6
978
+ },
979
+ {
980
+ "type": "text",
981
+ "text": "Results. As we can see from Table 1, most defenses claim a detection performance that implies a far greater robust accuracy than our current best robust classifiers. To illustrate with a concrete example, the CIFAR-10 detector of Miller et al. (2019) claims to achieve robust accuracy with detection of $7 5 \\%$ for $\\ell _ { 2 }$ attacks with $\\epsilon = 2 . 9$ . Using Theorem 4, this implies an inefficient classifier with robust accuracy of $7 5 \\%$ for $\\ell _ { 2 }$ attacks with $\\epsilon = { } ^ { 2 . 9 } / 2 = 1 . 4 5$ . Yet, the current state-of-the-art robust accuracy for such a perturbation budget is only $3 0 \\%$ (Rebuffi et al., 2021). If this detector defense’s robustness claim were correct, it would imply a remarkable breakthrough in robust classification. ",
982
+ "bbox": [
983
+ 174,
984
+ 670,
985
+ 825,
986
+ 768
987
+ ],
988
+ "page_idx": 6
989
+ },
990
+ {
991
+ "type": "text",
992
+ "text": "Why do many of these defenses claim robust accuracies that appear “too good to be true”? A primary reason is that the vast majority of the above detector defenses do not consider evaluations against adaptive attacks (Carlini et al., 2019; Athalye et al., 2018; Tramer et al., 2020). That is, \\` these defenses show that they can detect some fixed attacks, and thereafter conclude that the detector is robust against all attacks. As in the case of robust classifiers, such an evaluation is clearly insufficient! Some defenses do evaluate against adaptive adversaries, but fail to build a sufficiently strong attack to reliably approximate the worst-case robust risk. Because of the lack of a strong comparative baseline, it is not always immediately clear that these results are overly strong. ",
993
+ "bbox": [
994
+ 173,
995
+ 773,
996
+ 825,
997
+ 886
998
+ ],
999
+ "page_idx": 6
1000
+ },
1001
+ {
1002
+ "type": "table",
1003
+ "img_path": "images/35afe76e5c882d9d74a70d86ea496028c6c8d6414cd130e3ebb49471d266dd3d.jpg",
1004
+ "table_caption": [
1005
+ "Table 1: For each detector defense, we compute a (best-effort) bound on the claimed robust risk with detection $R _ { \\mathrm { a d v - d e t } } ^ { \\epsilon }$ using Equation (4), and report the complement (the robust accuracy with detection), $1 - R _ { \\mathrm { a d v - d e t } } ^ { \\epsilon }$ . For each detector’s robustness claim (at distance $\\epsilon$ ), we report the stateof-the-art robust classification accuracy for attacks at distance $\\epsilon / 2$ , denoted $1 - R _ { \\mathrm { a d v } } ^ { \\epsilon / 2 }$ . Detection defense claims that imply a higher robust classification accuracy than the current state-of-the-art are highlighted in red. "
1006
+ ],
1007
+ "table_footnote": [],
1008
+ "table_body": "<table><tr><td>Dataset</td><td>Defense</td><td>Norm</td><td>E</td><td>1-Radv-det</td><td>1- R</td></tr><tr><td rowspan=\"3\">MNIST</td><td>Grosse et al. (2017)</td><td>l8</td><td>0.5</td><td>≥ 98%</td><td>94%</td></tr><tr><td>Ma et al. (2018)</td><td>l2</td><td>4.2</td><td>≥ 99%</td><td>72%</td></tr><tr><td>Raghuram et al. (2021)</td><td>l2</td><td>8.9</td><td>≥ 74%</td><td>0%</td></tr><tr><td rowspan=\"8\">CIFAR-10</td><td>Yin et al. (2020)</td><td>l2</td><td>1.7</td><td>≥ 90%</td><td>66%</td></tr><tr><td>Feinman et al. (2017)</td><td>l2</td><td>2.7</td><td>≥ 43%</td><td>36%</td></tr><tr><td>Miller et al. (2019)</td><td>l2</td><td>2.9</td><td>≥ 75%</td><td>30%</td></tr><tr><td>Raghuram et al. (2021)</td><td>l</td><td>4.0</td><td>≥ 56%</td><td>10%</td></tr><tr><td>Ma &amp; Liu (2019)</td><td>l8</td><td>4/255</td><td>≥ 96%</td><td>85%</td></tr><tr><td>Roth et al. (2019)</td><td>l8</td><td>8/255</td><td>≥ 66%</td><td>79%</td></tr><tr><td>Lee et al. (2018)</td><td>lo</td><td>20/255</td><td>≥ 81%</td><td>59%</td></tr><tr><td>Li et al. (2019)</td><td>lo</td><td>26/255</td><td>≥ 80%</td><td>44%</td></tr><tr><td rowspan=\"5\">ImageNet</td><td>Xu et al. (2018)</td><td>l</td><td>1.0</td><td>≥ 67%</td><td>54%</td></tr><tr><td>Ma &amp; Liu (2019)</td><td>l8</td><td>2/255</td><td>≥ 68%</td><td>55%</td></tr><tr><td>Jha et al. (2019)</td><td>lo</td><td>2/255</td><td>≥ 30%</td><td>55%</td></tr><tr><td>Hendrycks &amp; Gimpel (2017)</td><td>l8</td><td>10/255</td><td>≥ 76%</td><td>30%</td></tr><tr><td>Yu et al. (2019)</td><td>l8</td><td>26/255</td><td>M 7%</td><td>5%</td></tr></table>",
1009
+ "bbox": [
1010
+ 194,
1011
+ 212,
1012
+ 797,
1013
+ 536
1014
+ ],
1015
+ "page_idx": 7
1016
+ },
1017
+ {
1018
+ "type": "text",
1019
+ "text": "For example, the recent work of Raghuram et al. (2021, ICML Long Talk) builds a detector on MNIST with a FNR of $\\leq 5 \\%$ at a FPR of $\\leq 2 0 \\%$ , for adaptive $\\ell _ { 2 }$ attacks bounded by $\\epsilon = 8 . 9$ . Yet, this perturbation bound is much larger than the average distance between an MNIST image and the nearest image from a different class! Thus, an attack within this perturbation bound can trivially reduce the detector’s accuracy to chance. On CIFAR-10, the same detector achieves $9 5 \\%$ clean accuracy, and a FNR of $\\leq 1 9 \\%$ at a FPR of $\\leq 2 0 \\%$ for adaptive $\\ell _ { 2 }$ attacks bounded by $\\epsilon = 4$ . Using Equation (4), this yields a bound on the robust accuracy with detection of $1 - R _ { \\mathrm { a d v - d e t } } ^ { \\epsilon } ( f ) \\geq$ $1 - \\bar { ( 5 \\% + 1 9 \\% + 2 0 \\% ) } = 5 6 \\%$ . In contrast, the best robust classifier we are aware of for $\\ell _ { 2 }$ attacks bounded by $\\epsilon = 2$ achieves robust accuracy of only $1 0 \\%$ (Rebuffi et al., 2021). In summary, the adaptive attack considered in this detector defense’s evaluation is highly unlikely to be good approximation of a worst-case attack, and this defense can likely be broken by stronger attacks. ",
1020
+ "bbox": [
1021
+ 173,
1022
+ 566,
1023
+ 825,
1024
+ 720
1025
+ ],
1026
+ "page_idx": 7
1027
+ },
1028
+ {
1029
+ "type": "text",
1030
+ "text": "Certifiably robust detection and classification. In Table 2, we look at the robust accuracy with detection, and standard robust accuracy achieved by certified defenses (for which the claimed robustness numbers are necessarily mathematically correct). ",
1031
+ "bbox": [
1032
+ 176,
1033
+ 742,
1034
+ 823,
1035
+ 784
1036
+ ],
1037
+ "page_idx": 7
1038
+ },
1039
+ {
1040
+ "type": "text",
1041
+ "text": "We note that our reduction is not as meaningful in the case of certified defenses, since it is highly plausible that computational inefficiency can be leveraged to build better certified classifiers. Indeed, given any robust classifier (e.g., an adversarially trained model), the classifier’s robustness can always be certified inefficiently (by enumerating over all points within an $\\epsilon$ -ball). Thus, the existence of an inefficient classifier with higher certified robustness than the state-of-the-art is to be expected. ",
1042
+ "bbox": [
1043
+ 174,
1044
+ 791,
1045
+ 823,
1046
+ 861
1047
+ ],
1048
+ "page_idx": 7
1049
+ },
1050
+ {
1051
+ "type": "text",
1052
+ "text": "Nevertheless, we find that existing results for certified classifiers and detectors perfectly match what is implied by our reduction (up to $\\pm 2 \\%$ error). For example, Zhang et al. (2020a) follow a long line of results on robust classifiers and achieve $3 9 \\%$ robust accuracy on CIFAR-10 for perturbations of $\\ell _ { \\infty }$ -norm below $^ { 4 } / 2 5 5$ . Together with Theorem 5, this implies an inefficient detector with $3 9 \\%$ robust detection accuracy for perturbations of $\\ell _ { \\infty }$ -norm below $^ { 8 } / 2 5 5$ . The recent work of Sheikholeslami et al. (2021) nearly matches that bound $3 7 \\%$ robust accuracy with detection), with a defense that has the advantage of being concretely efficient. ",
1053
+ "bbox": [
1054
+ 174,
1055
+ 867,
1056
+ 823,
1057
+ 924
1058
+ ],
1059
+ "page_idx": 7
1060
+ },
1061
+ {
1062
+ "type": "table",
1063
+ "img_path": "images/308cd424cf793a3ba0100fdea38ed9dfb95bea708f1ec26aa902357976d5420b.jpg",
1064
+ "table_caption": [
1065
+ "Table 2: Certified robust accuracy $1 - R _ { \\mathrm { a d v } } ^ { \\epsilon / 2 }$ for the defense of Zhang et al. (2020a), and certified robust accuracy with detection $1 - R _ { \\mathrm { a d v - d e t } } ^ { \\epsilon }$ dv for the defense of Sheikholeslami et al. (2021). "
1066
+ ],
1067
+ "table_footnote": [],
1068
+ "table_body": "<table><tr><td>E</td><td>1-Radv-det</td><td></td></tr><tr><td>8/255</td><td>37%</td><td>39%</td></tr><tr><td>16/255</td><td>32%</td><td>33%</td></tr></table>",
1069
+ "bbox": [
1070
+ 380,
1071
+ 156,
1072
+ 612,
1073
+ 219
1074
+ ],
1075
+ "page_idx": 8
1076
+ },
1077
+ {
1078
+ "type": "text",
1079
+ "text": "",
1080
+ "bbox": [
1081
+ 176,
1082
+ 243,
1083
+ 826,
1084
+ 285
1085
+ ],
1086
+ "page_idx": 8
1087
+ },
1088
+ {
1089
+ "type": "text",
1090
+ "text": "These results give additional credence to our thesis: with current techniques, robust classification is indeed approximately twice as hard (in terms of the perturbation bounds covered) than robust detection. ",
1091
+ "bbox": [
1092
+ 176,
1093
+ 292,
1094
+ 823,
1095
+ 334
1096
+ ],
1097
+ "page_idx": 8
1098
+ },
1099
+ {
1100
+ "type": "text",
1101
+ "text": "Extensions and open problems. The main open problem raised by our work is of course whether it could be possible to show an efficient reduction between classification and detection of adversarial examples, but this seems implausible (at least with our minimum distance decoding approach). ",
1102
+ "bbox": [
1103
+ 174,
1104
+ 349,
1105
+ 825,
1106
+ 391
1107
+ ],
1108
+ "page_idx": 8
1109
+ },
1110
+ {
1111
+ "type": "text",
1112
+ "text": "Another interesting question is whether a similar reduction can be shown for robustness to less “structured” perturbations than $\\ell _ { p }$ balls and other metric spaces. For example, there has been a line of research on defending against adversarial patches (Brown et al., 2017), using empirical (Hayes, 2018; Naseer et al., 2019; Chou et al., 2020) and certifiable techniques (Chiang et al., 2020; Zhang et al., 2020b; Xiang et al., 2021). To use our result, we would have to define some metric to measure the size of an adversarial patch’s perturbation. Yet, the size of a patch is typically defined by the number of contiguously perturbed pixels, which does not define a metric (in particular, it does not satisfy the triangular inequality which our reduction relies on). ",
1113
+ "bbox": [
1114
+ 174,
1115
+ 397,
1116
+ 825,
1117
+ 510
1118
+ ],
1119
+ "page_idx": 8
1120
+ },
1121
+ {
1122
+ "type": "text",
1123
+ "text": "Finally, similar hardness reductions might exist between other candidate approaches for building robust classifiers. For example, the question of whether (test-time) randomness can be leveraged to build more robust models is also intriguing .Empirical defenses that use randomness can be notoriously hard to evaluate (Athalye et al., 2018; Tramer et al., 2020), so a reduction similar to ours \\` might be useful in showing that we should not expect such approaches to bare fruit. ",
1124
+ "bbox": [
1125
+ 174,
1126
+ 517,
1127
+ 825,
1128
+ 587
1129
+ ],
1130
+ "page_idx": 8
1131
+ },
1132
+ {
1133
+ "type": "text",
1134
+ "text": "4 CONCLUSION ",
1135
+ "text_level": 1,
1136
+ "bbox": [
1137
+ 176,
1138
+ 606,
1139
+ 318,
1140
+ 622
1141
+ ],
1142
+ "page_idx": 8
1143
+ },
1144
+ {
1145
+ "type": "text",
1146
+ "text": "We have shown formal reductions between robust classification with, and without, a detection option. Our results show that significant progress on one of these two tasks implies similar progress on the other—unless computational inefficiency can somehow be leveraged to build more robust models. This raises the question on whether we should spend our efforts on studying both of these tasks, or focus our efforts on a single one. ",
1147
+ "bbox": [
1148
+ 174,
1149
+ 637,
1150
+ 823,
1151
+ 707
1152
+ ],
1153
+ "page_idx": 8
1154
+ },
1155
+ {
1156
+ "type": "text",
1157
+ "text": "On one hand, the two tasks represent different ways of tackling a common goal, and working on either task might result in new techniques or ideas that apply to the other task as well. On the other hand, our reductions show that unless we make progress on both tasks, work on one of the tasks can merely aim to match the robustness of our inefficient constructions, whilst improving their computational complexity. ",
1158
+ "bbox": [
1159
+ 174,
1160
+ 714,
1161
+ 825,
1162
+ 784
1163
+ ],
1164
+ "page_idx": 8
1165
+ },
1166
+ {
1167
+ "type": "text",
1168
+ "text": "We believe our reduction will serve as a useful sanity-check when assessing the claims of future detector defenses. Detector defenses’ robustness evaluations have received less stringent scrutiny than robust classifiers over the past years, perhaps in part due to a lack of strong comparative baselines. Instead of having to wait until some detector defense’s claims pass the test-of-time, we show that detection results can be directly contrasted against long-standing results for robust classification. ",
1169
+ "bbox": [
1170
+ 174,
1171
+ 791,
1172
+ 823,
1173
+ 861
1174
+ ],
1175
+ "page_idx": 8
1176
+ },
1177
+ {
1178
+ "type": "text",
1179
+ "text": "When applying this approach to past detector defenses, we find that many make robustness claims that imply significant breakthroughs in robust classification. We believe our reduction could have been useful in highlighting the suspiciously strong claims made by many of these defenses—before they were explicitly broken by stronger attacks. ",
1180
+ "bbox": [
1181
+ 174,
1182
+ 867,
1183
+ 823,
1184
+ 924
1185
+ ],
1186
+ "page_idx": 8
1187
+ },
1188
+ {
1189
+ "type": "text",
1190
+ "text": "Ethics statement. Our paper demonstrates a fundamental barrier towards detecting against adversarial examples, under the assumption that our current techniques are insufficient to achieve strong (inefficient) robust classification. We do not however explicitly break any existing defenses (our results merely strongly suggest that many existing detector defenses’ claims are suspiciously high). Our paper therefore cannot lead to any explicit harms, but aims to further our understanding of the hardness of robust classification and detection. ",
1191
+ "bbox": [
1192
+ 174,
1193
+ 103,
1194
+ 825,
1195
+ 186
1196
+ ],
1197
+ "page_idx": 9
1198
+ },
1199
+ {
1200
+ "type": "text",
1201
+ "text": "Reproducibility statement. Our paper’s contribution is mainly of theoretical nature. Section 2 is self-contained and clearly states our assumptions, results and proofs (except for the proof of Theorem 5 in Appendix A). The experiments in Section 3 use only public datasets and pre-trained models, with clearly indicates hyper-parameters for all attacks that we evaluate. ",
1202
+ "bbox": [
1203
+ 174,
1204
+ 203,
1205
+ 825,
1206
+ 258
1207
+ ],
1208
+ "page_idx": 9
1209
+ },
1210
+ {
1211
+ "type": "text",
1212
+ "text": "REFERENCES ",
1213
+ "text_level": 1,
1214
+ "bbox": [
1215
+ 174,
1216
+ 280,
1217
+ 285,
1218
+ 295
1219
+ ],
1220
+ "page_idx": 9
1221
+ },
1222
+ {
1223
+ "type": "text",
1224
+ "text": "Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, 2020. ",
1225
+ "bbox": [
1226
+ 173,
1227
+ 304,
1228
+ 823,
1229
+ 344
1230
+ ],
1231
+ "page_idx": 9
1232
+ },
1233
+ {
1234
+ "type": "text",
1235
+ "text": "Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, 2018. ",
1236
+ "bbox": [
1237
+ 174,
1238
+ 354,
1239
+ 821,
1240
+ 397
1241
+ ],
1242
+ "page_idx": 9
1243
+ },
1244
+ {
1245
+ "type": "text",
1246
+ "text": "Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndi ˇ c, Pavel Laskov, Gior- ´ gio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 387–402. Springer, 2013. ",
1247
+ "bbox": [
1248
+ 173,
1249
+ 406,
1250
+ 825,
1251
+ 464
1252
+ ],
1253
+ "page_idx": 9
1254
+ },
1255
+ {
1256
+ "type": "text",
1257
+ "text": "Tom B Brown, Dandelion Mane, Aurko Roy, Mart ´ ´ın Abadi, and Justin Gilmer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2017. ",
1258
+ "bbox": [
1259
+ 173,
1260
+ 473,
1261
+ 821,
1262
+ 502
1263
+ ],
1264
+ "page_idx": 9
1265
+ },
1266
+ {
1267
+ "type": "text",
1268
+ "text": "Sebastien Bubeck, Yin Tat Lee, Eric Price, and Ilya Razenshteyn. Adversarial examples from cryp-´ tographic pseudo-random generators. arXiv preprint arXiv:1811.06418, 2018. ",
1269
+ "bbox": [
1270
+ 173,
1271
+ 511,
1272
+ 820,
1273
+ 541
1274
+ ],
1275
+ "page_idx": 9
1276
+ },
1277
+ {
1278
+ "type": "text",
1279
+ "text": "Nicholas Carlini and David Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In AISec, pp. 3–14. ACM, 2017. ",
1280
+ "bbox": [
1281
+ 173,
1282
+ 549,
1283
+ 823,
1284
+ 579
1285
+ ],
1286
+ "page_idx": 9
1287
+ },
1288
+ {
1289
+ "type": "text",
1290
+ "text": "Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, and Aleksander Madry. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705, 2019. ",
1291
+ "bbox": [
1292
+ 176,
1293
+ 587,
1294
+ 823,
1295
+ 631
1296
+ ],
1297
+ "page_idx": 9
1298
+ },
1299
+ {
1300
+ "type": "text",
1301
+ "text": "Ping-yeh Chiang, Renkun Ni, Ahmed Abdelkader, Chen Zhu, Christoph Studer, and Tom Goldstein. Certified defenses for adversarial patches. arXiv preprint arXiv:2003.06693, 2020. ",
1302
+ "bbox": [
1303
+ 171,
1304
+ 638,
1305
+ 821,
1306
+ 669
1307
+ ],
1308
+ "page_idx": 9
1309
+ },
1310
+ {
1311
+ "type": "text",
1312
+ "text": "Edward Chou, Florian Tramer, and Giancarlo Pellegrino. Sentinet: Detecting physical attacks \\` against deep learning systems. In Workshop on Deep Learning Security, 2020. ",
1313
+ "bbox": [
1314
+ 171,
1315
+ 676,
1316
+ 821,
1317
+ 707
1318
+ ],
1319
+ "page_idx": 9
1320
+ },
1321
+ {
1322
+ "type": "text",
1323
+ "text": "Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. ",
1324
+ "bbox": [
1325
+ 169,
1326
+ 715,
1327
+ 825,
1328
+ 744
1329
+ ],
1330
+ "page_idx": 9
1331
+ },
1332
+ {
1333
+ "type": "text",
1334
+ "text": "Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020. ",
1335
+ "bbox": [
1336
+ 176,
1337
+ 752,
1338
+ 823,
1339
+ 796
1340
+ ],
1341
+ "page_idx": 9
1342
+ },
1343
+ {
1344
+ "type": "text",
1345
+ "text": "Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. URL https://github.com/MadryLab/robustness. ",
1346
+ "bbox": [
1347
+ 169,
1348
+ 805,
1349
+ 823,
1350
+ 834
1351
+ ],
1352
+ "page_idx": 9
1353
+ },
1354
+ {
1355
+ "type": "text",
1356
+ "text": "Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. ",
1357
+ "bbox": [
1358
+ 171,
1359
+ 843,
1360
+ 823,
1361
+ 872
1362
+ ],
1363
+ "page_idx": 9
1364
+ },
1365
+ {
1366
+ "type": "text",
1367
+ "text": "Sanjam Garg, Somesh Jha, Saeed Mahloujifar, and Mahmoody Mohammad. Adversarially robust learning could leverage computational hardness. In Algorithmic Learning Theory, pp. 364–385. PMLR, 2020. ",
1368
+ "bbox": [
1369
+ 176,
1370
+ 882,
1371
+ 825,
1372
+ 922
1373
+ ],
1374
+ "page_idx": 9
1375
+ },
1376
+ {
1377
+ "type": "text",
1378
+ "text": "Kathrin Grosse, Praveen Manoharan, Nicolas Papernot, Michael Backes, and Patrick McDaniel. On the (statistical) detection of adversarial examples. arXiv preprint arXiv:1702.06280, 2017. ",
1379
+ "bbox": [
1380
+ 171,
1381
+ 103,
1382
+ 823,
1383
+ 132
1384
+ ],
1385
+ "page_idx": 10
1386
+ },
1387
+ {
1388
+ "type": "text",
1389
+ "text": "Jamie Hayes. On visible adversarial perturbations & digital watermarking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 1597–1604, 2018. ",
1390
+ "bbox": [
1391
+ 171,
1392
+ 142,
1393
+ 823,
1394
+ 171
1395
+ ],
1396
+ "page_idx": 10
1397
+ },
1398
+ {
1399
+ "type": "text",
1400
+ "text": "Warren He, James Wei, Xinyun Chen, Nicholas Carlini, and Dawn Song. Adversarial example defenses: Ensembles of weak defenses are not strong. In USENIX Workshop on Offensive Technologies, 2017. ",
1401
+ "bbox": [
1402
+ 176,
1403
+ 179,
1404
+ 823,
1405
+ 223
1406
+ ],
1407
+ "page_idx": 10
1408
+ },
1409
+ {
1410
+ "type": "text",
1411
+ "text": "Dan Hendrycks and Kevin Gimpel. Early methods for detecting adversarial images. In International Conference on Learning Representations, 2017. ",
1412
+ "bbox": [
1413
+ 174,
1414
+ 232,
1415
+ 821,
1416
+ 262
1417
+ ],
1418
+ "page_idx": 10
1419
+ },
1420
+ {
1421
+ "type": "text",
1422
+ "text": "Susmit Jha, Sunny Raj, Steven Lawrence Fernandes, Sumit Kumar Jha, Somesh Jha, Gunjan Verma, Brian Jalaian, and Ananthram Swami. Attribution-driven causal analysis for detection of adversarial examples. arXiv preprint arXiv:1903.05821, 2019. ",
1423
+ "bbox": [
1424
+ 173,
1425
+ 271,
1426
+ 823,
1427
+ 314
1428
+ ],
1429
+ "page_idx": 10
1430
+ },
1431
+ {
1432
+ "type": "text",
1433
+ "text": "Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems, 2018. ",
1434
+ "bbox": [
1435
+ 173,
1436
+ 323,
1437
+ 823,
1438
+ 366
1439
+ ],
1440
+ "page_idx": 10
1441
+ },
1442
+ {
1443
+ "type": "text",
1444
+ "text": "Yingzhen Li, John Bradshaw, and Yash Sharma. Are generative classifiers more robust to adversarial attacks? In International Conference on Machine Learning, pp. 3804–3814. PMLR, 2019. ",
1445
+ "bbox": [
1446
+ 173,
1447
+ 375,
1448
+ 823,
1449
+ 405
1450
+ ],
1451
+ "page_idx": 10
1452
+ },
1453
+ {
1454
+ "type": "text",
1455
+ "text": "Shiqing Ma and Yingqi Liu. Nic: Detecting adversarial samples with neural network invariant checking. In Proceedings of the 26th Network and Distributed System Security Symposium (NDSS 2019), 2019. ",
1456
+ "bbox": [
1457
+ 174,
1458
+ 414,
1459
+ 823,
1460
+ 457
1461
+ ],
1462
+ "page_idx": 10
1463
+ },
1464
+ {
1465
+ "type": "text",
1466
+ "text": "Xingjun Ma, Bo Li, Yisen Wang, Sarah M Erfani, Sudanthi Wijewickrema, Grant Schoenebeck, Dawn Song, Michael E Houle, and James Bailey. Characterizing adversarial subspaces using local intrinsic dimensionality. In International Conference on Learning Representations, 2018. ",
1467
+ "bbox": [
1468
+ 176,
1469
+ 465,
1470
+ 823,
1471
+ 510
1472
+ ],
1473
+ "page_idx": 10
1474
+ },
1475
+ {
1476
+ "type": "text",
1477
+ "text": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. ",
1478
+ "bbox": [
1479
+ 174,
1480
+ 517,
1481
+ 823,
1482
+ 560
1483
+ ],
1484
+ "page_idx": 10
1485
+ },
1486
+ {
1487
+ "type": "text",
1488
+ "text": "David Miller, Yujia Wang, and George Kesidis. When not to classify: Anomaly detection of attacks (ada) on dnn classifiers at test time. Neural computation, 31(8):1624–1670, 2019. ",
1489
+ "bbox": [
1490
+ 173,
1491
+ 570,
1492
+ 821,
1493
+ 599
1494
+ ],
1495
+ "page_idx": 10
1496
+ },
1497
+ {
1498
+ "type": "text",
1499
+ "text": "Muzammal Naseer, Salman Khan, and Fatih Porikli. Local gradients smoothing: Defense against localized adversarial attacks. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1300–1307. IEEE, 2019. ",
1500
+ "bbox": [
1501
+ 173,
1502
+ 608,
1503
+ 826,
1504
+ 652
1505
+ ],
1506
+ "page_idx": 10
1507
+ },
1508
+ {
1509
+ "type": "text",
1510
+ "text": "Tianyu Pang, Huishuai Zhang, Di He, Yinpeng Dong, Hang Su, Wei Chen, Jun Zhu, and Tie-Yan Liu. Adversarial training with rectified rejection. arXiv preprint arXiv:2105.14785, 2021. ",
1511
+ "bbox": [
1512
+ 174,
1513
+ 661,
1514
+ 823,
1515
+ 690
1516
+ ],
1517
+ "page_idx": 10
1518
+ },
1519
+ {
1520
+ "type": "text",
1521
+ "text": "Aditi Raghunathan, Jacob Steinhardt, and Percy Liang. Certified defenses against adversarial examples. In International Conference on Learning Representations, 2018. ",
1522
+ "bbox": [
1523
+ 173,
1524
+ 699,
1525
+ 823,
1526
+ 729
1527
+ ],
1528
+ "page_idx": 10
1529
+ },
1530
+ {
1531
+ "type": "text",
1532
+ "text": "Jayaram Raghuram, Varun Chandrasekaran, Somesh Jha, and Suman Banerjee. A general framework for detecting anomalous inputs to dnn classifiers. In International Conference on Machine Learning, 2021. ",
1533
+ "bbox": [
1534
+ 173,
1535
+ 738,
1536
+ 823,
1537
+ 781
1538
+ ],
1539
+ "page_idx": 10
1540
+ },
1541
+ {
1542
+ "type": "text",
1543
+ "text": "Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. ",
1544
+ "bbox": [
1545
+ 173,
1546
+ 790,
1547
+ 825,
1548
+ 833
1549
+ ],
1550
+ "page_idx": 10
1551
+ },
1552
+ {
1553
+ "type": "text",
1554
+ "text": "Kevin Roth, Yannic Kilcher, and Thomas Hofmann. The odds are odd: A statistical test for detecting adversarial examples. In International Conference on Machine Learning, 2019. ",
1555
+ "bbox": [
1556
+ 169,
1557
+ 843,
1558
+ 823,
1559
+ 872
1560
+ ],
1561
+ "page_idx": 10
1562
+ },
1563
+ {
1564
+ "type": "text",
1565
+ "text": "Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarially robust generalization requires more data. In Advances In Neural Information Processing Systems, pp. 5019–5031, 2018. ",
1566
+ "bbox": [
1567
+ 174,
1568
+ 881,
1569
+ 825,
1570
+ 924
1571
+ ],
1572
+ "page_idx": 10
1573
+ },
1574
+ {
1575
+ "type": "text",
1576
+ "text": "Fatemeh Sheikholeslami, Ali Lotfi, and J Zico Kolter. Provably robust classification of adversarial examples with detection. In International Conference on Learning Representations, 2021. ",
1577
+ "bbox": [
1578
+ 171,
1579
+ 103,
1580
+ 823,
1581
+ 132
1582
+ ],
1583
+ "page_idx": 11
1584
+ },
1585
+ {
1586
+ "type": "text",
1587
+ "text": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. ",
1588
+ "bbox": [
1589
+ 176,
1590
+ 140,
1591
+ 823,
1592
+ 184
1593
+ ],
1594
+ "page_idx": 11
1595
+ },
1596
+ {
1597
+ "type": "text",
1598
+ "text": "Florian Tramer and Dan Boneh. Adversarial training and robustness for multiple perturbations. In \\` Advances In Neural Information Processing Systems, 2019. ",
1599
+ "bbox": [
1600
+ 171,
1601
+ 193,
1602
+ 825,
1603
+ 222
1604
+ ],
1605
+ "page_idx": 11
1606
+ },
1607
+ {
1608
+ "type": "text",
1609
+ "text": "Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to \\` adversarial example defenses. In Conference on Neural Information Processing Systems, 2020. ",
1610
+ "bbox": [
1611
+ 174,
1612
+ 229,
1613
+ 823,
1614
+ 260
1615
+ ],
1616
+ "page_idx": 11
1617
+ },
1618
+ {
1619
+ "type": "text",
1620
+ "text": "Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In International Conference on Learning Representations, 2019. ",
1621
+ "bbox": [
1622
+ 173,
1623
+ 267,
1624
+ 825,
1625
+ 310
1626
+ ],
1627
+ "page_idx": 11
1628
+ },
1629
+ {
1630
+ "type": "text",
1631
+ "text": "Eric Wong and Zico Kolter. Provable defenses against adversarial examples via the convex outer adversarial polytope. In International Conference on Machine Learning, pp. 5283–5292, 2018. ",
1632
+ "bbox": [
1633
+ 173,
1634
+ 319,
1635
+ 821,
1636
+ 349
1637
+ ],
1638
+ "page_idx": 11
1639
+ },
1640
+ {
1641
+ "type": "text",
1642
+ "text": "Chong Xiang, Arjun Nitin Bhagoji, Vikash Sehwag, and Prateek Mittal. Patchguard: A provably robust defense against adversarial patches via small receptive fields and masking. In 30th {USENIX} Security Symposium ({USENIX} Security 21), 2021. ",
1643
+ "bbox": [
1644
+ 173,
1645
+ 356,
1646
+ 825,
1647
+ 401
1648
+ ],
1649
+ "page_idx": 11
1650
+ },
1651
+ {
1652
+ "type": "text",
1653
+ "text": "Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In Network and Distributed System Security Symposium, 2018. ",
1654
+ "bbox": [
1655
+ 173,
1656
+ 409,
1657
+ 821,
1658
+ 438
1659
+ ],
1660
+ "page_idx": 11
1661
+ },
1662
+ {
1663
+ "type": "text",
1664
+ "text": "Xuwang Yin, Soheil Kolouri, and Gustavo K Rohde. Gat: Generative adversarial training for adversarial example detection and robust classification. In International Conference on Learning Representations, 2020. ",
1665
+ "bbox": [
1666
+ 174,
1667
+ 445,
1668
+ 825,
1669
+ 488
1670
+ ],
1671
+ "page_idx": 11
1672
+ },
1673
+ {
1674
+ "type": "text",
1675
+ "text": "Tao Yu, Shengyuan Hu, Chuan Guo, Wei-Lun Chao, and Kilian Q Weinberger. A new defense against adversarial images: Turning a weakness into a strength. In Advances in Neural Information Processing Systems, 2019. ",
1676
+ "bbox": [
1677
+ 174,
1678
+ 497,
1679
+ 825,
1680
+ 541
1681
+ ],
1682
+ "page_idx": 11
1683
+ },
1684
+ {
1685
+ "type": "text",
1686
+ "text": "Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P. Xing, Laurent El Ghaoui, and Michael I. Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning, 2019. ",
1687
+ "bbox": [
1688
+ 173,
1689
+ 549,
1690
+ 825,
1691
+ 592
1692
+ ],
1693
+ "page_idx": 11
1694
+ },
1695
+ {
1696
+ "type": "text",
1697
+ "text": "Huan Zhang, Hongge Chen, Chaowei Xiao, Bo Li, Duane Boning, and Cho-Jui Hsieh. Towards stable and efficient training of verifiably robust neural networks. In International Conference on Learning Representations, 2020a. ",
1698
+ "bbox": [
1699
+ 173,
1700
+ 601,
1701
+ 821,
1702
+ 643
1703
+ ],
1704
+ "page_idx": 11
1705
+ },
1706
+ {
1707
+ "type": "text",
1708
+ "text": "Zhanyuan Zhang, Benson Yuan, Michael McCoyd, and David Wagner. Clipped bagnet: Defending against sticker attacks with clipped bag-of-features. In 2020 IEEE Security and Privacy Workshops (SPW), pp. 55–61. IEEE, 2020b. ",
1709
+ "bbox": [
1710
+ 176,
1711
+ 652,
1712
+ 823,
1713
+ 695
1714
+ ],
1715
+ "page_idx": 11
1716
+ },
1717
+ {
1718
+ "type": "text",
1719
+ "text": "A PROOF OF THEOREM 5. ",
1720
+ "text_level": 1,
1721
+ "bbox": [
1722
+ 176,
1723
+ 102,
1724
+ 406,
1725
+ 118
1726
+ ],
1727
+ "page_idx": 12
1728
+ },
1729
+ {
1730
+ "type": "text",
1731
+ "text": "We recall Theorem 5: ",
1732
+ "bbox": [
1733
+ 174,
1734
+ 133,
1735
+ 318,
1736
+ 148
1737
+ ],
1738
+ "page_idx": 12
1739
+ },
1740
+ {
1741
+ "type": "text",
1742
+ "text": "Theorem 5 $( \\epsilon / 2$ robust-classification implies inefficient $\\epsilon$ -robust detection). Let $d ( \\cdot , \\cdot )$ be an arbi \ntrary metric. Let g be a defense than explicit (but inefficient) defense achieves robust rithat achieves risk $R _ { a d \\nu } ^ { \\epsilon / 2 } ( f ) = \\beta$ . Then, we can constructrobust risk with detection $f$ $R ( f ) \\leq \\beta$ \n$R _ { a d \\nu - d e t } ^ { \\epsilon } ( f ) \\leq \\beta$ . ",
1743
+ "bbox": [
1744
+ 173,
1745
+ 151,
1746
+ 825,
1747
+ 213
1748
+ ],
1749
+ "page_idx": 12
1750
+ },
1751
+ {
1752
+ "type": "text",
1753
+ "text": "The defense $f$ is constructed as follows on input $x$ : ",
1754
+ "bbox": [
1755
+ 176,
1756
+ 218,
1757
+ 509,
1758
+ 234
1759
+ ],
1760
+ "page_idx": 12
1761
+ },
1762
+ {
1763
+ "type": "text",
1764
+ "text": "• Run the classifier $y g ( x )$ . ",
1765
+ "bbox": [
1766
+ 218,
1767
+ 244,
1768
+ 419,
1769
+ 261
1770
+ ],
1771
+ "page_idx": 12
1772
+ },
1773
+ {
1774
+ "type": "text",
1775
+ "text": "• Find a perturbed input $x ^ { \\prime }$ withing distance $\\epsilon / 2$ of $x$ that is classified differently, i.e., $\\bar { d } ( x , x ^ { \\prime } ) \\bar { \\ } \\leq \\ \\epsilon / 2$ and $g ( x ^ { \\prime } ) \\neq y$ . If such an input $x ^ { \\prime }$ exists, reject the input and output ⊥. Else, output the class $y$ . ",
1776
+ "bbox": [
1777
+ 217,
1778
+ 268,
1779
+ 823,
1780
+ 311
1781
+ ],
1782
+ "page_idx": 12
1783
+ },
1784
+ {
1785
+ "type": "text",
1786
+ "text": "Proof of Theorem 5. Note that for any input $( x , y )$ for which the classifier $g$ is robust at distance $\\epsilon / 2$ , no input $x ^ { \\prime }$ above exists and so $f ( x ) = y$ . Thus, the risk of $f$ is at most the robust risk of $g$ , so $R ( f ) \\leq \\beta$ . ",
1787
+ "bbox": [
1788
+ 174,
1789
+ 327,
1790
+ 823,
1791
+ 369
1792
+ ],
1793
+ "page_idx": 12
1794
+ },
1795
+ {
1796
+ "type": "text",
1797
+ "text": "Now, consider an input $( x , y ) \\sim \\mathcal { D }$ for which $f$ is not robust with detection at distance $\\epsilon$ . That is, either $f ( x ) \\neq y$ , or there exists an input $\\hat { x }$ at distance $d ( x , { \\hat { x } } ) \\leq \\epsilon$ such that $f ( \\hat { x } ) = \\hat { y } \\not \\in \\{ y , \\bot \\}$ . We will show that the defense $g$ is not robust for $x$ either (for attacks at distance up to $\\epsilon / 2$ .) ",
1798
+ "bbox": [
1799
+ 174,
1800
+ 376,
1801
+ 823,
1802
+ 419
1803
+ ],
1804
+ "page_idx": 12
1805
+ },
1806
+ {
1807
+ "type": "text",
1808
+ "text": "If $f ( x ) \\neq y$ , then by the same argument as above it cannot be the case that $g$ is robust at distance $\\epsilon / 2$ for $x$ . ",
1809
+ "bbox": [
1810
+ 174,
1811
+ 425,
1812
+ 823,
1813
+ 454
1814
+ ],
1815
+ "page_idx": 12
1816
+ },
1817
+ {
1818
+ "type": "text",
1819
+ "text": "So let us consider the case where $f ( \\hat { x } ) = \\hat { y } \\not \\in \\{ y , \\bot \\}$ . By the definition of $f$ , this means that for all $x ^ { \\prime }$ at distance at most $\\epsilon / 2$ from $\\hat { x }$ , we have $g ( x ^ { \\prime } ) = \\hat { y }$ . But, note that there exists a point $x ^ { * }$ that is at distance at most $\\epsilon / 2$ from both $\\hat { x }$ and $x$ . Since we must have $g ( x ^ { * } ) = \\hat { y }$ , we conclude that $g$ is not robust at distance $\\epsilon / 2$ for $x$ . ",
1820
+ "bbox": [
1821
+ 174,
1822
+ 459,
1823
+ 825,
1824
+ 516
1825
+ ],
1826
+ "page_idx": 12
1827
+ },
1828
+ {
1829
+ "type": "text",
1830
+ "text": "Taking expectations over the distribution $\\mathcal { D }$ concludes the proof. ",
1831
+ "bbox": [
1832
+ 174,
1833
+ 523,
1834
+ 596,
1835
+ 539
1836
+ ],
1837
+ "page_idx": 12
1838
+ }
1839
+ ]
parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_layout.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c01d2ae97475f8c6d143c666ab42ccc1f41b8a2a6548b5252910cb488ec89328
3
+ size 429948
parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_middle.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_model.json ADDED
The diff for this file is too large to render. See raw diff
 
parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3f7bb9d2b43dd2afc1c77feefca4845be616717aeb87237acb041ef9768d34d8
3
+ size 284090
parse/dev/QJb1-8NH2Ux/QJb1-8NH2Ux_span.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eadd73d6d6302188af8c87f250d8588d7476cd1af562008c2564f1d4228db07d
3
+ size 441379
parse/dev/QJb1-8NH2Ux/images/1703fa013d6a1ee6b00378a016c2a93fc0e3073c9d6734ac06a5bcc5e5e8f741.jpg ADDED

Git LFS Details

  • SHA256: 23215f05494e542d16606230a91cb27a9cafe1fce4160f0c533e93d1c1cf67d1
  • Pointer size: 130 Bytes
  • Size of remote file: 44.3 kB
parse/dev/QJb1-8NH2Ux/images/280c3197d6ce35502a31b2d1bc26122b4c360c6fa70eabe111b6e26a291e0bec.jpg ADDED

Git LFS Details

  • SHA256: eb1ad0506982b243d29147daec0d99670344342969ef6cfff94ffd9a039e3554
  • Pointer size: 129 Bytes
  • Size of remote file: 7.83 kB
parse/dev/QJb1-8NH2Ux/images/308cd424cf793a3ba0100fdea38ed9dfb95bea708f1ec26aa902357976d5420b.jpg ADDED

Git LFS Details

  • SHA256: 17c0f7900190b2d95166907b5a32a63bbb1b5810e9d29cb2c345640e48ec66c3
  • Pointer size: 129 Bytes
  • Size of remote file: 9.45 kB
parse/dev/QJb1-8NH2Ux/images/35afe76e5c882d9d74a70d86ea496028c6c8d6414cd130e3ebb49471d266dd3d.jpg ADDED

Git LFS Details

  • SHA256: df408380ba20d4c4b346993606c020db5b6f8352a3a86d3fd799da010070da40
  • Pointer size: 131 Bytes
  • Size of remote file: 108 kB
parse/dev/QJb1-8NH2Ux/images/44c657694cd74a56c1870443ece67a7d806cd61222c843330a77a072ca61eca5.jpg ADDED

Git LFS Details

  • SHA256: cd34b7e7c32594d898a6e04d497d49990bd422876cea1839f81c57fdea3e08fb
  • Pointer size: 129 Bytes
  • Size of remote file: 4.12 kB
parse/dev/QJb1-8NH2Ux/images/8e622d645b3f35c117ff52694ec9773f02e8ea8238d5cfacd531d2db3bd601ae.jpg ADDED

Git LFS Details

  • SHA256: 54484229687a52e515f2b8bd6cf38239c76a42311bc15c76aa7b42f447b3a795
  • Pointer size: 129 Bytes
  • Size of remote file: 4.63 kB
parse/dev/QJb1-8NH2Ux/images/bf3313c2eebf9b088d68940d3fcdf78503af5aea78a5151d35b998f8ac2a42af.jpg ADDED

Git LFS Details

  • SHA256: b2a03aed464d55ff30ac97e4641072a53649c7f39db975cef74e81ea7ea7811c
  • Pointer size: 129 Bytes
  • Size of remote file: 6.36 kB
parse/dev/QNBzcgY0f4e/QNBzcgY0f4e.md ADDED
@@ -0,0 +1,526 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Easy incremental learning methods to consider for commercial fine-tuning applications
2
+
3
+ Anonymous Author(s)
4
+ Affiliation
5
+ Address
6
+ email
7
+
8
+ # Abstract
9
+
10
+ Fine-tuning deep learning models for commercial use cases is growing exponentially as more and more companies are adopting AI to enhance their core products and services, as well as automate their diurnal processes and activities. However, not many countries like the U.S. and those in Europe follow quality data collection methods for AI vision or NLP related automation applications. Thus, on many of these kinds of data, existing state-of-the-art pre-trained deep learning models fail to perform accurately, and when fine-tuning is done on these models, issues like catastrophic forgetting or being less specific in predictions as expected occur. Hence, in this paper, simplified incremental learning methods are introduced to be considered in existing fine-tuning infrastructures of pre-trained models (such as those available in huggingface.com) to help mitigate the aforementioned issues for commercial applications. The methods introduced are: 1) Fisher Shut-off, 2) Fractional Data Retention and 3) Border Control. Results show that when applying these methods on vanilla pre-trained models, the models are in fact able to add more to their knowledge without hurting much on what they had learned previously.
11
+
12
+ # 16 1 Introduction
13
+
14
+ 17 Many companies and organizations today are adopting AI in automation, automating their daily
15
+ 18 processes and activities, as well as offering them in their core products and services. Automation
16
+ 19 has traditionally been in the industry for many years, as a means for which economics of scale could
17
+ 20 be acheived so as to remain competitive in the market. Now with AI, more and more intelligence is
18
+ 21 being brought into automation, and in countries like India, organizations are beginning to adopt AI
19
+ 22 for this particular purpose.
20
+ 23 With recent advancements in AI vision and NLP models such as the GPT-3, Jurassic-1, and so on,
21
+ 24 organizations today are using AI for 1) Document Reading and Understanding, 2) Online Proctoring,
22
+ 25 3) Chatbots, 4) Intelligent Information Parsing and other application related process automations.
23
+ 26 Given these use cases, AI solutions need to be specific to their processes, but yet be an addition to
24
+ 27 their generally known formats. This in a sense, is more like making use of a human employee who
25
+ 28 has some kind of general education on various tasks or processes but still is required to learn the
26
+ 29 companies counterparts well and in detail before he/she is allowed to execute them. These processes
27
+ 30 can include between, reading customer emails for entering relevant information about their product
28
+ 31 requirements onto a structured database, to understanding various types of printed documents for
29
+ 32 information parsing, and to identifying newer objects for either document filtering or malicious
30
+ 33 activity detection.
31
+ 34 For natural language related tasks, powerful models like the GPT-3 are now being widely used, but
32
+ 35 they require good prompt engineering skills to get the best out of them. Also, given that they are
33
+ 36 probabilistic models, the generated outputs can sometimes falter away from what is expected, and
34
+ 37 this can become a problem when selling it to customers, because even the slightest faltering may not
35
+ 38 be acceptable to them at all. Hence, to reduce this, more and more examples have to be provided in
36
+ 39 the prompt, and this can come at a high cost not suitable for low cost of living countries like India.
37
+ 40 The other workaround is to fine-tune the model on the new datasets, but this has epoch limitations
38
+ 41 on how deeply it can fit on the new dataset without hurting the body of general knowledge it gained
39
+ 42 earlier. Also, fine-tuning models like the GPT-3 comes at a very high cost now-a-days, and is no
40
+ 43 more an option. This leaves the automation builders to use huggingface.com transformers instead.
41
+ 44 In vision, although state-of-the-art pre-trained deep learning models are able to achieve human level
42
+ 45 performance on a variety of inputs, they can only perform so in upto close to high quality inputs. If
43
+ 46 the quality goes lower, they fail terribly. Not all organizations have a good quality data collection
44
+ 47 process involved for applying automation, and this is ubiquitously the case in many parts of the world.
45
+ 48 So it becomes quite difficult to sell AI as a human-level performer, and at this point AI becomes of
46
+ 49 lesser use than it could potentially be.
47
+ 50 Another approach typically used to resolve such problems is to employ transfer learning, which
48
+ 51 typically involves replacing the last layers of the model with a new model to get the specific outputs
49
+ 52 required. Some examples done in research are Too, et al. (2019), Dif & Elberrichi (2020), Alshalali
50
+ 53 & Joysula (2018), Jung, et al. (2015), Qian, et al. (2021) and Vrbanciˇ c & Podgorelec (2020). While ˇ
51
+ 54 this may not seem to be a problem with vision based tasks, it is definitely a problem with natural
52
+ 55 language based tasks. This is because the final layers of the natural language models have all the vital
53
+ 56 information of language structure that help with the language generative process. When this is to
54
+ 57 be changed, catastrophic forgetting can happen. Catastrophic forgetting is a phenomenon in which
55
+ 58 previously learned knowledge is lost partly by the application of new data for training. Also, with
56
+ 59 vision based tasks, when the requirement is to just improve the performance on lower quality data,
57
+ 60 transfer learning may not be the appropriate approach. Fine-tuning for these must involve the final
58
+ 61 layers of the model which could inevitably lead to catastrophic forgetting on the higher quality inputs.
59
+ 62 This brings the only solution towards incremental learning. This type of learning is all about
60
+ 63 learning on newer datasets without having the side-effects catastrophic forgetting, and there has been
61
+ 64 substantial amount of research done in this area. Luo, et al. (2020) summarizes all the work that has
62
+ 65 happened in this area so far. There are several approaches to implementing incremental learning on
63
+ 66 pre-trained models, some of which will be discussed in the forthcoming sections. In this paper, a
64
+ 67 few of these approaches will be simplified for commercial applications along with novel intuitive
65
+ 68 additions to further help the learning process. The paper introduces: 1) Fisher Shut-off which is a
66
+ 69 simplification of the work done by Kirkpatrick, et al. (2017), 2) Fractional Data Retention which
67
+ 70 adopts ideas from Castro, et al. (2018), and 3) Border Control which is an extension to the idea
68
+ 71 outlined by Ren, et al. (2018) on reweighting examples by employing a method similar to Adaboost.
69
+ 72 The last one is the novel addition as it formulates a different approach to retaining salient examples
70
+ 73 for incremental learning. It is based on the work by Ruping (2001) on incremental learning with
71
+ 74 SVMs. But since SVMs are too complex in the context on neural networks, a similar but simplified
72
+ 75 approach is proposed.
73
+ 76 The purpose of this work is to initiate the development of a new infrastructure for commercial
74
+ 77 fine-tuning of pre-trained models with simplified incremental learning methods.
75
+ 78 The rest of this paper proceeds as follows: Section 2 will provide a brief discussion on incremental
76
+ 79 learning methods developed so far, followed by the proposal of simplified incremental learning
77
+ 80 methods in Section 3. Section 4 will show sample results of the proposed methods on a vanilla
78
+ 81 pre-trained model using a toy dataset. A toy dataset is used for the only purpose of providing
79
+ 82 visualizations on the performance of the proposed methods. Nevertheless, these methods can be
80
+ 83 extended on to real world datasets. The paper then concludes in Section 5 discussing steps forward
81
+ 84 for implementation.
82
+
83
+ # 85 2 Incremental Learning
84
+
85
+ 86 This section is a summary of the review published by Luo, et al. (2020). In this review, four different
86
+ 87 types of strategies for incremental learning are highlighted, and every work published in this area
87
+ 88 uses either one or more such strategies. Some examples are Castro, et al. (2018) and He, et al. (2020).
88
+ 89 The four strategies are:
89
+
90
+ • Architectural • Regularization • Rehearsal • Pseudo-Rehearsal
91
+
92
+ 94 The following subsections will disccuss these briefly.
93
+
94
+ # 2.1 Architectural Strategy
95
+
96
+ 96 This strategy is similar to boosting techniques where multiple models are trained. But when used in the context of incremental learning, each model is trained on a different task separately. Then another 98 meta-model that effectively selects which model to use for inference is trained. The work done by 99 Poliker, et al. (2001) resembles this in many ways. In this work, multiple classifiers are trained with 0 different training sets, and then a Adaboost style of ensemble learning is employed to combine the model outputs.
97
+
98
+ 102 Another interesting work is by Rusu, et al. (2016) on Progressive Neural Networks (PNN). In this
99
+ 103 work, a neural network is trained sequentially on different tasks or training sets. However, each time,
100
+ 104 new neurons are added in each layer with new weights, and the weights of the previously learned
101
+ 105 neural network are frozen. Then, to prevent catastrophic forgetting, the outputs of each layer of the
102
+ 106 previous neural network on the earlier training set are used in addition to the new task or training
103
+ 107 set, when training the new layer neurons. The results on this type of incremental learning were quite
104
+ 108 encouraging that it set a new direction in the research of dynamically expanding networks that could
105
+ 109 make better use the neural networks capacity than the PNN. In fact, it will be seen later that the Fisher
106
+ 110 Shut-off method proposed in this paper inherently employs the idea of PNNs.
107
+
108
+ # 2.2 Regularization Strategy
109
+
110
+ 112 In this strategy, as the name suggest, a regularization term is added in the loss function that measures
111
+ 113 the importance of old knowledge when learning on a new training set. The representive work done in
112
+ 114 this is Kirkpatrick, et al. (2017), whereby they introduce the concept of Elastic Weight Consolidation
113
+ 115 (EWC) by means of a Fisher Information Matrix. The EWC brings about the regularization term in
114
+ 116 the loss function as
115
+
116
+ $$
117
+ R ( w ) = \sum _ { i } \frac { \lambda } { 2 } F _ { i } ( w _ { i } - w _ { i , o l d } ) ^ { 2 }
118
+ $$
119
+
120
+ 117 where $F _ { i }$ is the Fisher Information Matrix which suggests the importance of the $i$ -th weight trained
121
+ 118 on the old (or previous) training set. Here, as one could speculate, the term Fisher Shut-off proposed
122
+ 119 in this paper actually derives itself from the Fisher Information Matrix, meaning that this matrix is
123
+ 120 used as the basis for shutting off the training of certain weights when training on a new set.
124
+ 121 Another popular type of regularization strategy is Knowledge Distillation introduced by Hinton, et al.
125
+ 122 (2015). In this method, knowledge from an ensemble of models trained on different tasks (or training
126
+ 123 sets) separately are distilled into a smaller model that can be deployed much easily for inference.
127
+ 124 There are many huggingface.com transformers that are a product of such knowledge distillation.
128
+ 125 The distillation ensures that the smaller model holds all the knowledge of the ensemble, and that it
129
+ 126 can infer as good as it. Distillation is done by setting soft-targets on the smaller network from all the
130
+ 127 earlier training sets of the ensemble. The soft-targets are the output logits from the ensemble models
131
+ 128 on their respective trained datasets.
132
+
133
+ # 129 2.3 Rehearsal and Pseudo-Rehearsal Strategies
134
+
135
+ 130 Rehearsal strategies in incremental learning make use of the earlier training sets when training a
136
+ 131 model on new tasks or training sets. This by far is the simplest of all incremental learning strategies
137
+ 132 that ensures catastrophic forgetting is prevented. The only issue is that when this strategy is used for
138
+ 133 deep learning models trained on large datasets, the training on new datasets could become extremely
139
+ 134 slow and even time consuming before any fruitful results are achieved. Hence, newer research work in
140
+ 135 this area formulate methods for retaining only the most important data points to prevent catastrophic
141
+ 136 forgetting. The work done by Castro, et al. (2018) is an example of this. In this work, selection and
142
+ 137 removal mechanisms on data are introduced for assimilation into a memory network.
143
+ 138 Talking about memory networks, the Pseudo-Rehearsal strategy involves training an additional data
144
+ 139 generator to generate the samples, the neural network was trained on earlier. Hence, newer research
145
+ 140 in this area involve GANs for data generation. Examples are Odena, et al. (2017) and Wu, et al.
146
+ 141 (2018).
147
+
148
+ # 142 3 Proposed Incremental Learning Methods
149
+
150
+ 143 Commercial applications always require simplistic implementations of advanced methods no matter
151
+ 144 how complex they may be. Therefore, it is for this purpose alone this paper proposes some simplified
152
+ 145 methods for implementing incremental learning. As metioned earlier in Section 1, these methods are:
153
+ 146 1) Fisher Shut-off, 2) Fractional Data Retention, and 3) Border Control. This section covers them in
154
+ 147 detail.
155
+
156
+ # 148 3.1 Fisher Shut-off
157
+
158
+ 149 As mentioned in the previous section, the term Fisher Shut-off derives itself from the Fisher Infor
159
+ 150 mation Matrix which weighs the importance of weights trained on previous datasets. Hence, in this
160
+ 151 sub-section, a brief overview of the details behind this matrix is covered with the help of Aich (2021).
161
+ 152 Let $\mathcal { D }$ represent a dataset coming from a stream of data for incremental learning. Then $p ( w | \mathcal { D } )$
162
+ 153 represents the model trained on data $\mathcal { D }$ . This means that to train a model on a new dataset, the
163
+ 154 following posterior must satisfy:
164
+
165
+ $$
166
+ p ( w | \mathcal D _ { n e w } ) = \frac { p ( \mathcal D _ { n e w } | w ) p ( w | \mathcal D _ { o l d } ) } { p ( \mathcal D _ { n e w } ) }
167
+ $$
168
+
169
+ 155 Note here that $p ( w | \mathcal { D } _ { o l d } )$ is written in place of $p ( w )$ because when $\mathcal { D } _ { n e w }$ is applied to the model, the
170
+ 156 weights $w$ have already been trained with $\mathcal { D } _ { o l d }$ . Hence, given the model, $p ( w | \mathcal { D } _ { o l d } )$ , the log-likelihood
171
+ 157 loss on $\mathcal { D } _ { n e w }$ becomes,
172
+
173
+ $$
174
+ \begin{array} { r l } & { \mathcal { L } _ { \mathcal { D } _ { n e w } } ( w ) = l o g ( p ( w | \mathcal { D } _ { n e w } ) ) } \\ & { \qquad = l o g ( p ( \mathcal { D } _ { n e w } | w ) ) + l o g ( p ( w | \mathcal { D } _ { o l d } ) ) - l o g ( p ( \mathcal { D } _ { n e w } ) ) } \\ & { \qquad \approx l o g ( p ( \mathcal { D } _ { n e w } | w ) ) + l o g ( p ( w | \mathcal { D } _ { o l d } ) ) } \end{array}
175
+ $$
176
+
177
+ 158 Here, the $l o g ( p ( \mathcal { D } _ { n e w } | w ) )$ equals the cross-entropy loss of the model on $\mathcal { D } _ { n e w }$ while $l o g ( p ( w | \mathcal { D } _ { o l d } ) )$
178
+ 159 is loss of the model on $\mathcal { D } _ { o l d }$ . To ensure that catastrophic forgetting does not occur on $\mathcal { D } _ { o l d }$ in its
179
+ 160 absence while training on $\mathcal { D } _ { n e w }$ , the loss on $\mathcal { D } _ { o l d }$ will have to be approximated using $w$ alone. To do
180
+ 161 this, the Taylor’s expansion on $l o g ( p ( w | \mathcal { D } _ { o l d } ) )$ is taken as,
181
+
182
+ $$
183
+ \begin{array} { r l } & { \mathcal { L } _ { \mathcal { D } _ { o l d } } ( w ) \approx \mathcal { L } ( w ) \big | _ { \mathcal { D } _ { o l d } } + \left( \frac { \partial \mathcal { L } ( w ) } { \partial w } \Big | _ { \mathcal { D } _ { o l d } } \right) + \frac { 1 } { 2 } ( w - w \big | _ { \mathcal { D } _ { o l d } } ) ^ { T } \left( \frac { \partial ^ { 2 } \mathcal { L } ( w ) } { \partial ^ { 2 } w } \Big | _ { \mathcal { D } _ { o l d } } \right) ( w - w \big | _ { \mathcal { D } _ { o l d } } ) } \\ & { \qquad \approx \mathcal { L } ( w ) \big | _ { \mathcal { D } _ { o l d } } + \frac { 1 } { 2 } ( w - w \big | _ { \mathcal { D } _ { o l d } } ) ^ { T } \left( \frac { \partial ^ { 2 } \mathcal { L } ( w ) } { \partial ^ { 2 } w } \Big | _ { \mathcal { D } _ { o l d } } \right) ( w - w \big | _ { \mathcal { D } _ { o l d } } ) } \end{array}
184
+ $$
185
+
186
+ since technically 162 $\begin{array} { r } { \frac { \partial \mathcal { L } ( w ) } { \partial w } \bigg | _ { \mathcal { D } _ { o l d } } = 0 } \end{array}$ , if the model is trained well on $\mathcal { D } _ { o l d }$ . Then, noting that the last term 163 in (4) is equivalent to a regularization term, this term alone could be considered as the loss on $\mathcal { D } _ { o l d }$
187
+
188
+ 164 for preventing catastrophic forgetting. In doing so, the Fisher Information Matrix will equal to the
189
+ 165 Hessian, $\frac { \partial ^ { 2 } \mathcal { L } ( w ) } { \partial ^ { 2 } w } \bigg | _ { \mathscr { D } _ { o l d } } .$ This Hessian, $\mathcal { H }$ , can be simply computed by the model gradients $\frac { \bar { \partial \mathcal { L } } ( w ) } { \partial w } \bigg | _ { \mathscr { D } _ { o l d } }$
190
+ 166 assuming that not all gradients are zero, as,
191
+
192
+ $$
193
+ \mathcal { H } = \frac { \partial \mathcal { L } ( w ) } { \partial w } \bigg | _ { \mathcal { D } _ { o l d } } \cdot \frac { \partial \mathcal { L } ( w ) } { \partial w } \bigg | _ { \mathcal { D } _ { o l d } } ^ { T }
194
+ $$
195
+
196
+ 167 Doing so, and keeping only the diagonal terms, would imply that the model gradients are more
197
+ 168 than enough to weigh the important weights of the model trained on $\mathcal { D } _ { o l d }$ . Replacing (5) in (4) and
198
+ 169 substituting in (3) would give the loss on $\mathcal { D } _ { n e w }$ as,
199
+
200
+ $$
201
+ \mathcal { L } _ { \mathcal { D } _ { n e w } } ( w ) \approx l o g ( p ( \mathcal { D } _ { n e w } | w ) ) + \frac { 1 } { 2 } ( w - w \big | _ { \mathcal { D } _ { o l d } } ) ^ { T } \left( \frac { \partial \mathcal { L } ( w ) } { \partial w } \bigg | _ { \mathcal { D } _ { o l d } } \cdot \frac { \partial \mathcal { L } ( w ) } { \partial w } \bigg | _ { \mathcal { D } _ { o l d } } ^ { T } \right) ( w - w \big | _ { \mathcal { D } _ { o l d } } )
202
+ $$
203
+
204
+ 170 which to an extent implies that if the model gradients on $\mathcal { D } _ { o l d }$ are absolutely zero, they get trained on
205
+ 171 $\mathcal { D } _ { n e w }$ without regularization, while those that are not, get regularized towards $w \big | _ { \mathscr { D } _ { o l d } }$ .
206
+ 172 This is what the proposed Fisher Shut-off exploits. In Fisher Shut-off, all weights of the model
207
+ 173 trained on $\mathcal { D } _ { o l d }$ that do not have absolute zero gradients get shut-off for training on $\mathcal { D } _ { n e w }$ , while the
208
+ 174 remaining that do take part. Also, since in practice $R e L U$ functions are commonly used in deep
209
+ 175 learning models as the activation functions of the neurons, shutting off these weights becomes as
210
+ 176 simple as setting a condition. Figure 1 shows a sample performance of Fisher Shut-off on a regression
211
+ 177 model trained sequentially on mutually exclusive batches of data. These batches could represent the
212
+ 178 different tasks or training sets.
213
+ 179 However, when it comes to classification, simple shut-off does not work completely. This is because,
214
+ 180 while in regression problems datasets could inherently employ some kind of piece-wise nonlinear fit
215
+ 181 in their distributions, the same cannot always be guaranteed in classification. Thus, in classification,
216
+ 182 the shut-off weights must also take part in training. And, as per (1), there is a learning constant
217
+ 183 required in the regularization to ensure that the right balances between $\mathcal { D } _ { n e w }$ and $\mathcal { D } _ { o l d }$ are met on
218
+ 184 these weights. This paper provides a novel learning constant determination for this regularization.
219
+ 185 This is detailed in Appendix A.
220
+ 86 Also in regression problems, if datasets have batch distributions that are quite far apart from each
221
+ 87 previous batch, then Fisher-Shutoff may not fully work too. Appendix B shows some of these
222
+ 188 examples
223
+
224
+ ![](images/ec3ca6e1de828ff9f19c334a4f580525e37d7f6bbb827e170bad1c560561bda1.jpg)
225
+ Figure 1: Fisher Shut-off on a regression model on six mutually exclusive batches of data. Blue dots represent the overall dataset, while green square dots are the batch or task data. The red line is the model’s output after each batch is fed to it. Fisher Shut-off is used from Batch $\# 2$ onwards.
226
+
227
+ # 3.2 Fractional Data Retention
228
+
229
+ This is a very simply proposal. The idea is to retain only a fraction of the data trained on the neural network on the earlier tasks or training sets. There is nothing more to this. However, banking on the ideas of selection highlighted in Castro, et al. (2018), whereby data is selected based on their proximity to cluster centers, to be more representative of the classes, this paper uses this as the baseline idea behind its proposal on Fractional Data Retention. Thus in Fractional Data Retention, a fraction of the data within the data cluster is retained and appended in every stage of incremental learning.
230
+
231
+ # 197 3.3 Border Control
232
+
233
+ The most important requirement when incrementally learning classes is to ensure that the decision boundaries of the earlier training tasks are protected as much as possible when training on new sets. If data points are used for this purpose, it would seem that, those that lie closest to the decision boundaries after training would be the most important ones to retain, for any succeeding incremental learning tasks. Thus, the Border Control method proposed in this paper exploits this. Ruping (2001) used SVMs to identify these data points as the support vectors that helped define the overall decision boundaries. But with deep learning models or vanilla neural networks, SVM is quite complex and therefore in order to be able retain data points closest to the decision boundaries, a different selection mechanism is required. This selection mechanism could instead be based on selecting data points on how large the absolute errors in sigmoidal outputs are for the applied dataset, as the data points closest to the decision boundaries have this inherent property.
234
+
235
+ 209 Furthermore, since real world data can be quite complex, it would be necessary to not only select data
236
+ 210 points based on how large their errors in sigmoidal outputs are, but also those points that are far away
237
+ 211 from them. This is because, given the context of incremental learning where there is a high chance
238
+ 212 that newer training sets may have data points that could potentially set newer decision boundaries in
239
+ 213 those fartherest regions, these data points would help protect those.
240
+ 214 Hence in Border Control, the top- $\mathtt { k }$ data points that have the largest absolute errors in the sigmoidal
241
+ 215 outputs and their respective top- $\mathtt { k }$ fartherest data points are retained in every task or training set
242
+ 216 for further incremental learning. These points are appended to the newer training sets before further
243
+ 217 training is applied.
244
+
245
+ # 218 4 Sample Results
246
+
247
+ 219 The proposed methods are tested on a toy dataset, as mentioned in Section 1, only to provide some
248
+ 220 visuals on how the incremental learning progresses using the proposed methods. Figure 2 shows this
249
+ 221 dataset. A vanilla deep neural network of size, 1000-1000-1000-1000-3, is used for incrementally
250
+ 222 learning batches of data from this toy dataset. The activation functions for all layers are $R e L U$ except
251
+ 223 for the output which is a sof tmax. All weights are uniformly but randomly initialized with a single
252
+ 224 random seed to make the results comparable. The weights are also scaled by a $\frac { 2 } { \sqrt { n } }$ factor to ensure
253
+ 225 that minimal overfitting occurs during training. Here $n$ is the layer fan-in.
254
+ 226 To visualize incremental learning on the proposed methods, the dataset is divided into 6 batches
255
+ 227 with mututally exclusive data points. This gives roughly between 100 to 200 data points in each
256
+ 228 batch, a size that is commonly used when training neural networks of this size. Figure 3 shows this.
257
+ 229 In this figure, it can be clearly seen that the batch distributions on the class data for incremental
258
+ 230 learning do not always form a piece-wise nonlinear fit, and therefore, plain shut-off of weights cannot
259
+ 231 fully retain knowledge learned earlier. Also, among these distributions, some allowed incremental
260
+ 232 learning to happen easily, while others did not, and the distribution shown in Figure 3 is one such.
261
+ 233 Table 1 summarizes the results of the proposed methods on this particular distribution. For other batch
262
+ 234 distributions, similar results could be achieved. Note here that quite some ML-Ops were required to
263
+ 235 achieve the results in Table 1. This was especially the case for those that employed Fisher Shut-off,
264
+ 236 since this method has a regularization constant that requires adapting on each batch. Furthermore,
265
+ 237 training on each batch was stopped once $1 0 0 . 0 \%$ accuracy was obtained on the batch. This left quite
266
+ 238 some data points to lie very close to the boundary lines or in some cases just right on them. Thus,
267
+ 239 the neural network was very vulnerable to catastrophic forgetting when succeeding batch trainings
268
+ 240 occurred as part of incremental learning.
269
+ 241 However, taking a look at Table 1, it can be seen that when Fisher Shut-off is applied, additional
270
+ 242 leverage against catastrophic forgetting occurs on each incremental batch, than when it is not used.
271
+ 243 And, among the three methods proposed in this paper, the Border Control method shows much
272
+ 244 stronger performance. In Figure 4, sample decision boundaries learned when each incremental batch
273
+ 245 is applied to the neural network using Fisher Shut-off and Border Control together is shown. A topk
274
+ 246 value of 5 is used for the Border Control. Also, note in Figure 4 that the red circles mark the border
275
+ 247 points accumulated on each batch. It can be seen that they clearly assume the data points closest
276
+ 248 to the decision boundaries, as well as those far away from it. All with respect to their batches. For
277
+ 249 the far away data points, their purpose can be clearly seen between batches #1 and $\# 2$ , where the
278
+ 250 fartherest points of class 2 in Batch #1 helped protect the decision boundaries from the data points
279
+ 251 of class 0 in Batch $\# 2$ . This means that more complex datasets can be accommodated by simply
280
+ 252 applying Border Control. More examples are shown in Appendix C.
281
+
282
+ ![](images/ed5b8993eaca01d170d93842c74b1f8afc29195bd92a7249a0f7e6d54ce3a8c4.jpg)
283
+ Figure 2: The toy dataset having three nonlinearly arranged classes.
284
+
285
+ ![](images/b02e14d6cf3d7912dafbbcf85b828e0529858490504fb69b8d48e1be08459fa6.jpg)
286
+ Figure 3: Batches on the toy dataset.
287
+
288
+ Table 1: Performance of proposed methods on the dataset of Figure 2
289
+
290
+ <table><tr><td rowspan="2">Method</td><td colspan="6"> Sample accuracy on accumulated dataset after Batch1</td></tr><tr><td>#1</td><td>#2</td><td>#3</td><td>#4</td><td>#5</td><td>#6</td></tr><tr><td>No Incremental Learning</td><td>100.0%</td><td>90.98%2</td><td>92.76%</td><td>94.89%</td><td>98.32%</td><td>96.11%</td></tr><tr><td>Fisher Shut-off (FS)</td><td>100.0%</td><td>99.74%</td><td>98.19%</td><td>98.88%</td><td>98.96%</td><td>96.89%</td></tr><tr><td>Frac. Data Ret. (FDR)[10%]</td><td>100.0%</td><td>97.94%</td><td>98.39%</td><td>98.89%</td><td>98.71%</td><td>98.67%</td></tr><tr><td>FDR[20%]</td><td>100.0%</td><td>98.71%</td><td>98.59%</td><td>99.36%</td><td>98.97%</td><td>98.78%</td></tr><tr><td>Border Ctrl. (BC)[t opk = 5]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>99.84%</td><td>100.0%</td><td>100.0%</td></tr><tr><td>BC[topk = 10]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>99.84%</td><td>100.0%</td><td>100.0%</td></tr><tr><td>FS + FDR[10%]</td><td>100.0%</td><td>99.74%</td><td>98.79%</td><td>99.52%</td><td>99.23%</td><td>98.78%</td></tr><tr><td>FS + FDR[20%]</td><td>100.0%</td><td>100.0%</td><td>99.19%</td><td>99.52%</td><td>99.48%</td><td>99.11%</td></tr><tr><td>FS +BC[topk = 5]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>99.84%</td><td>100.0%</td><td>100.0%</td></tr><tr><td>FS + BC[topk = 10]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td></tr></table>
291
+
292
+ ![](images/5bccdefef0dd25c13f1ecc98bcf95de7a924e776d8da83295052a7d16b995c74.jpg)
293
+ Figure 4: Incremental learning using Fisher Shut-off and Border Control together $[ t o p k = 5 ]$ ]. Black circles mark the batch data, while the red circles mark the accumulated border points.
294
+
295
+ Also, to add on further to this, for the most difficult incremental learning applications such as learning new classes as highlighted in Castro, et al. (2018) and He, et al. (2020), Border Control can help leverage the many issues associated with it like class imbalance, concept drift and so on.
296
+
297
+ # 5 Conclusion
298
+
299
+ 257 To summarize the work in this paper, three simplified methods for implementing incremental learning
300
+ 258 for commercial fine-tuning of pre-trained models was proposed. Results showed that while Border
301
+ 259 Control performed the best, Fisher Shut-off was able to leverage the performances. However, dataset
302
+ 260 used in this paper was a toy dataset and not one of the benchmark datasets typically used for
303
+ 261 incremental learning. Hence, testing these methods on the benchmark datasets is a potential next step
304
+ 262 forward. Then, preparing the prerequisites for each model available, like say in huggingface.com,
305
+ 263 for incremental learning must be done so that automation companies or any other AI organization
306
+ 264 can make use of them. From the methods proposed in this paper, the prerequisites would be: 1)
307
+ 265 the Shut-off matrix for the neural network weights, and 2) the border points for each of the learned
308
+ 266 classes. Additionally, an ML-Ops infrastructure can be provided to optimize the performances of
309
+ 267 the models that employ the Fisher Shut-off method. Metrics like the Backward Transfer (BWT) and
310
+ 268 Forward Transfer (FWT) proposed in Lopez-Paz & Ranzato (2017) can be used for this purpose.
311
+
312
+ # 269 References
313
+
314
+ 270 [1] Aich, A. (2021) Elastic weight consolidation (EWC): Nuts and bolts. arXiv preprint arXiv:2105.04093.
315
+ 271 [2] Alshalali, T. & Joysula, D. (2018) Fine-Tuning of Pre-Trained Deep Learning Models with Extreme Learning
316
+ 272 Machine. IEEE International Conference on Computational Science and Computational Intelligence (CSCI), pp.
317
+ 273 469-473.
318
+ 274 [3] Castro, F., Marín-Jiménez, M.J., Guil, N., Schmid, C. & Alahari, K. (2018) End-to-end incremental learning.
319
+ 275 Proceedings of the European Conference on Computer Vision (ECCV), pp. 233-248.
320
+ 276 [4] Dif, N. & Elberrichi, Z. (2020) A New Intra Fine-Tuning Method. International Journal of Service Science,
321
+ 277 Management, Engineering, and Technology, 11(2), pp. 16-40.
322
+ 278 [5] He, J., Mao, R., Shao, Z. & Zhu, F. (2020) Incremental learning in online scenario. IEEE/CVF Conference
323
+ 279 on Computer Vision and Pattern Recognition, pp. 13926-13935.
324
+ 280 [6] Hinton, G., Vinyals, O. & Dean, J. (2015) Distilling the Knowledge in a Neural Network. arXiv preprint
325
+ 281 arXiv:1503.02531, 2(7).
326
+ 282 [7] Jung, H., Lee, S., Yim, J., Park, S. & Kim, J. (2015) Joint fine-tuning in deep neural networks for facial
327
+ 283 expression recognition. Proceedings of the IEEE International Conference on Computer Vision, pp. 2983-2991.
328
+ 284 [8] Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J.,
329
+ 285 Ramalho, T., Grabska-Barwinska, A. & Hassabis, D. (2017) Overcoming catastrophic forgetting in neural
330
+ 286 networks. Proceedings of the National Academy of Sciences, 114(13), pp. 3521-3526.
331
+ 287 [9] Lopez-Paz, D. & Ranzato, M.A. (2017) Gradient episodic memory for continual learning. Proceedings of
332
+ 288 Neural Information Processing Systems (NIPS), pp. 6467-6476.
333
+ 289 [10] Luo, Y., Yin, L., Bai, W. & Mao, K. (2020) An Appraisal of Incremental Learning Methods. Entropy,
334
+ 290 22(11), pp. 1190-1216.
335
+ 291 [11] Odena, A., Olah, C. & Shlens, J. (2017) Conditional image synthesis with auxiliary classifier GANs.
336
+ 292 International Conference on Machine Learning (ICML), pp. 2642–2651.
337
+ 293 [12] Polikar, R., Udpa, L., Udpa, S. & Honavar, V. (2001) Learn++: An Incremental Learning Algorithm for
338
+ 294 Supervised Neural Networks. IEEE Transactions on Systems, Man, and Cybernetics, part C (applications and
339
+ 295 reviews), 31(4), pp. 497-508.
340
+ 296 [13] Qian, X., Zhang, C., Yella, J., Huang, Y., Huang, M.C. & Bom, S. (2021) Soft sensing model visualization:
341
+ 297 Fine-tuning neural network from what model learned. IEEE International Conference on Big Data (Big Data),
342
+ 298 pp. 1900-1908.
343
+ 299 [14] Ren, M., Zeng, W., Yang, B. & Urtasun, R. (2018) Learning to reweight examples for robust deep learning.
344
+ 300 International Conference on Machine Learning (ICML), pp. 4334-4343.
345
+ 301 [15] Ruping, S. (2001) Incremental learning with support vector machines. IEEE International Conference on
346
+ 302 Data Mining, pp. 641-642.
347
+ 303 [16] Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R. &
348
+ 304 Hadsell, R. (2016) Progressive neural networks. arXiv preprint arXiv:1606.04671.
349
+ 305 [17] Too, E., Yujian, L., Njuki, S. & Yingchun, L. (2019) A comparative study of fine-tuning deep learning
350
+ 306 models for plant disease. Computers and Electronics in Agriculture, 161, pp. 272-279.
351
+ 307 [18] Vrbanciˇ c, G. & Podgorelec, V. (2020) ˇ Transfer learning with adaptive fine-tuning. IEEE Access, Volume 8,
352
+ 308 pp. 196197-196211.
353
+ 309 [19] Wu, Y., Chen, Y.P., Wang, L.J., Ye, Y.C., Liu, Z.C., Guo, Y.D., Zhang, Z.Y. & Fu, Y. (2018) Incremental
354
+ 310 Classifier Learning with Generative Adversarial Networks. arXiv preprint arXiv:1802.00853.
355
+
356
+ The checklist follows the references. Please read the checklist guidelines carefully for information on how to answer these questions. For each question, change the default [TODO] to [Yes] , [No] , or [N/A] . You are strongly encouraged to include a justification to your answer, either by referencing the appropriate section of your paper or providing a brief inline description. For example:
357
+
358
+ • Did you include the license to the code and datasets? [Yes] See Section ??.
359
+ • Did you include the license to the code and datasets? [No] The code and the data are proprietary.
360
+ • Did you include the license to the code and datasets? [N/A]
361
+
362
+ Please do not modify the questions and only use the provided macros for your answers. Note that the Checklist section does not count towards the page limit. In your paper, please delete this instructions block and only keep the Checklist section heading above along with the questions/answers below.
363
+
364
+ 1. For all authors...
365
+
366
+ (a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
367
+ (b) Did you describe the limitations of your work? [Yes] See Sections 4 and 5
368
+ (c) Did you discuss any potential negative societal impacts of your work? [N/A]
369
+ (d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
370
+
371
+ 2. If you are including theoretical results...
372
+
373
+ (a) Did you state the full set of assumptions of all theoretical results? [Yes] See Section 3 and Appendix A
374
+ (b) Did you include complete proofs of all theoretical results? [Yes] See Section 3 and Appendix A
375
+
376
+ 3. If you ran experiments...
377
+
378
+ (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] As a supplemental material
379
+ (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [No] But it’s in the supplemental material
380
+ (c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [No] But additional examples are shown in Appendix C
381
+ (d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [N/A]
382
+
383
+ 4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
384
+
385
+ (a) If your work uses existing assets, did you cite the creators? [N/A]
386
+ (b) Did you mention the license of the assets? [N/A]
387
+ (c) Did you include any new assets either in the supplemental material or as a URL? [N/A]
388
+ (d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
389
+ (e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A]
390
+
391
+ 5. If you used crowdsourcing or conducted research with human subjects...
392
+
393
+ (a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
394
+ (b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
395
+ (c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
396
+
397
+ 362 Here, the derivation of regularization constant for Fisher Shut-off method is detailed. To start with,
398
+ 363 let the neural network be defined as,
399
+
400
+ $$
401
+ y = w ^ { T } \phi ( x , \omega )
402
+ $$
403
+
404
+ 364 Here, $w$ is the weights of the output layer, $\phi ( \cdot )$ is the output of the preceding layer, $x$ is the input and
405
+ 365 $\omega$ represents the rest of the weights of the neural network. Throughout the derivation, we will be
406
+ 366 dealing with only the output layer, and so the $\phi ( x , \omega )$ will be written in short form as $\Phi$ from here on.
407
+ 367 Let $e _ { k }$ denoted the error of fitting in the $k$ -th iteration, and $g _ { k }$ denote the error gradient. This would
408
+ 368 mean that $g _ { k } = \Phi _ { k } e _ { k }$ .
409
+ 369 Then, given the regularization term in (6), let $\tilde { w }$ denote difference in weights, between the new
410
+ 370 training and the previous training. This would give the weight updation policy as,
411
+
412
+ $$
413
+ w _ { k + 1 } = w _ { k } - \eta g _ { k } - \beta \tilde { w } _ { k }
414
+ $$
415
+
416
+ If 371 $e _ { k } = w _ { k } ^ { T } \Phi _ { k } - Y$ , then the error $e _ { k + 1 }$ after the weight updation would equal,
417
+
418
+ $$
419
+ \begin{array} { r l } & { e _ { k + 1 } = w _ { k + 1 } ^ { T } \Phi _ { k + 1 } - Y } \\ & { \qquad = ( w _ { k } - \eta g _ { k } - \beta \tilde { w } _ { k } ) ^ { T } \Phi _ { k + 1 } - Y } \\ & { \qquad = w _ { k } ^ { T } \Phi _ { k + 1 } - \eta g _ { k } ^ { T } \Phi _ { k + 1 } - \beta \tilde { w } _ { k } ^ { T } \Phi _ { k + 1 } - Y } \end{array}
420
+ $$
421
+
422
+ 372 Asumming for simplicity sake that $\Phi _ { k + 1 } \approx \Phi _ { k } + \delta$ , then (9) can continue as,
423
+
424
+ $$
425
+ \begin{array} { c } { { e _ { k + 1 } \approx w _ { k } ^ { T } \Phi _ { k } - \eta g _ { k } ^ { T } \Phi _ { k } - \beta \tilde { w } _ { k } ^ { T } \Phi _ { k } - Y + \Delta } } \\ { { { } } } \\ { { \approx e _ { k } - \eta g _ { k } ^ { T } \Phi _ { k } - \beta \tilde { w } _ { k } ^ { T } \Phi _ { k } } } \end{array}
426
+ $$
427
+
428
+ 373 Taking the square norm of $e _ { k + 1 }$ in (10), would equate this to,
429
+
430
+ $$
431
+ \begin{array} { r l } & { | | e _ { k + 1 } | | ^ { 2 } = | | e _ { k } | | ^ { 2 } + \eta ^ { 2 } | | g _ { k } ^ { T } \Phi _ { k } | | ^ { 2 } + \beta ^ { 2 } | | \tilde { w } _ { k } ^ { T } \Phi _ { k } | | ^ { 2 } } \\ & { ~ - ~ 2 \eta e _ { k } ^ { T } \Phi _ { k } ^ { T } g _ { k } - 2 \beta e _ { k } ^ { T } \Phi _ { k } ^ { T } \tilde { w } _ { k } + 2 \eta \beta \tilde { w } _ { k } ^ { T } \Phi _ { k } \Phi _ { k } ^ { T } g _ { k } } \end{array}
432
+ $$
433
+
434
+ 374 We require that $| | e _ { k + 1 } | | ^ { 2 } < | | e _ { k } | | ^ { 2 }$ at all times, so that regularization does not affect the fit at any
435
+ 375 point during the training. Applying this condition in (11) would give,
436
+
437
+ $$
438
+ \eta ^ { 2 } | | g _ { k } ^ { T } \Phi _ { k } | | ^ { 2 } + \beta ^ { 2 } | | \tilde { w } _ { k } ^ { T } \Phi _ { k } | | ^ { 2 } - 2 \eta e _ { k } ^ { T } \Phi _ { k } ^ { T } g _ { k } - 2 \beta e _ { k } ^ { T } \Phi _ { k } ^ { T } \tilde { w } _ { k } + 2 \eta \beta \tilde { w } _ { k } ^ { T } \Phi _ { k } \Phi _ { k } ^ { T } g _ { k } < 0
439
+ $$
440
+
441
+ 376 Then, taking the partial derivatives of (12) w.r.t $\eta$ and $\beta$ would give the following equations to be
442
+ 377 satisfied:
443
+
444
+ $$
445
+ \eta | | g _ { k } ^ { T } \Phi _ { k } | | ^ { 2 } - e _ { k } ^ { T } \Phi _ { k } ^ { T } g _ { k } + \beta \tilde { w } _ { k } ^ { T } \Phi _ { k } \Phi _ { k } ^ { T } g _ { k } = 0
446
+ $$
447
+
448
+ $$
449
+ \beta | | \tilde { w } _ { k } ^ { T } \Phi _ { k } | | ^ { 2 } - e _ { k } ^ { T } \Phi _ { k } ^ { T } \tilde { w } _ { k } + \eta \tilde { w } _ { k } ^ { T } \Phi _ { k } \Phi _ { k } ^ { T } g _ { k } = 0
450
+ $$
451
+
452
+ 378 Solving, (13) and (14) can result in negative $\eta$ and $\beta$ , which is not acceptable, and so to simplify the
453
+ 379 solution, we neglect the $\beta$ -term in (13). Doing so we get,
454
+
455
+ $$
456
+ \begin{array} { r l } & { \eta = \frac { e _ { k } ^ { T } \Phi _ { k } ^ { T } g _ { k } } { \vert \vert g _ { k } ^ { T } \Phi _ { k } \vert \vert ^ { 2 } } , } \\ & { } \\ & { \beta = \frac { e _ { k } ^ { T } \Phi _ { k } ^ { T } \tilde { w } _ { k } - \eta \tilde { w } _ { k } ^ { T } \Phi _ { k } \Phi _ { k } ^ { T } g _ { k } } { \vert \vert \tilde { w } _ { k } ^ { T } \Phi _ { k } \vert \vert ^ { 2 } } } \end{array}
457
+ $$
458
+
459
+ 380 Then, substituting for $g _ { k }$ and openning up the norms, we get,
460
+
461
+ $$
462
+ \eta = \frac { e _ { k } ^ { T } \Phi _ { k } ^ { T } \Phi _ { k } e _ { k } } { e _ { k } ^ { T } ( \Phi _ { k } ^ { T } \Phi _ { k } ) ( \Phi _ { k } ^ { T } \Phi _ { k } ) e _ { k } } ,
463
+ $$
464
+
465
+ $$
466
+ \beta = \frac { e _ { k } ^ { T } \Phi _ { k } ^ { T } \tilde { w } _ { k } - \eta \tilde { w } _ { k } ^ { T } ( \Phi _ { k } \Phi _ { k } ^ { T } ) \Phi _ { k } e _ { k } } { \tilde { w } _ { k } ^ { T } ( \Phi _ { k } \Phi _ { k } ^ { T } ) \tilde { w } _ { k } }
467
+ $$
468
+
469
+ 381 Simplifying (16) gives,
470
+
471
+ $$
472
+ \eta = \frac { e _ { k } ^ { T } e _ { k } } { e _ { k } ^ { T } \Phi _ { k } ^ { T } \Phi _ { k } e _ { k } } ,
473
+ $$
474
+
475
+ $$
476
+ \beta = ( 1 - \eta ) \frac { \tilde { w } _ { k } ^ { T } \Phi _ { k } e _ { k } } { \tilde { w } _ { k } ^ { T } \tilde { w } _ { k } }
477
+ $$
478
+
479
+ 382 Equation (17) gives the raw form for both $\eta$ and $\beta$ to be regulated. However, this will be further
480
+ 383 simplified for computating purposes, but will be used as a basis.
481
+ 384 Since the errors $e _ { k }$ get smaller as the neural network fits the data, using them in learning constants
482
+ 385 will only slow down the fits. A common way to overcome this is by replacing $e _ { k }$ with all ones.
483
+ 386 Similarly, for the $\tilde { w } _ { k }$ , all weights that are to be regularized are replaced with ones. If we denote
484
+ 387 the weights to be regularized as $w _ { r }$ , and there are $m$ patterns in the dataset with $n$ weights to be
485
+ 388 regularized, the $\eta$ and $\beta$ computations become,
486
+
487
+ $$
488
+ \eta = \frac { 1 } { | | \Phi _ { k } | | ^ { 2 } } ,
489
+ $$
490
+
491
+ $$
492
+ \beta = \frac { \alpha } { m n } \sum _ { i : w \in w _ { r } } \Phi _ { i , k }
493
+ $$
494
+
495
+ 389 Here, $\alpha$ represents the $( 1 - \eta )$ -term in (17). This constant will not neccessarily take the computed $\eta$
496
+ 390 when being regulated. Instead, this constant will have to be adapted each time for every incremental
497
+ 391 batch applied to the neural network.
498
+ 392 The reason why the computed $\eta$ is not used for the $\alpha$ adaptation is because this $\eta$ can sometimes
499
+ 393 become too small in the adaptation, that the $1 - \eta$ would always tend towards 1. When this was
500
+ 394 empirically tested on the toy dataset, the regularization was found at times to have gone too strong
501
+ 395 that the fit never happened. ML-Ops on the $\alpha$ found that this constant is not always 1, and can be
502
+ 396 anywhere between 0 and 1, or higher in some cases.
503
+
504
+ 398 Additional examples on the regression problem with Fisher Shut-off. Fisher Shut-off could not be used completely, and regularization had to take over for some batches. Figures 5 and 6 show this.
505
+
506
+ ![](images/72e9762490cf8ee10f248a8d3a0d85b6b28ea32a596f13c1b194dc4898ba90c7.jpg)
507
+ Figure 5: Complete Fisher Shut-off is used in Batches #2 and #6. Batches #3, #4 and #5 are regularized.
508
+
509
+ ![](images/b70e35df8ae79c3da760bbd94c1191f3f81fda2aacf6961298510e013e24a3ed.jpg)
510
+ Figure 6: Complete Fisher Shut-off is used in Batches $\# 2$ and $\# 6$ . Batches $\# 3$ and $\# 4$ are regularized. Batch #5 is fine-tuned
511
+
512
+ 401 In this appendix, additional examples on the toy dataset classification is shown. Figures 7 and 8 show 402 the batch distributions considered. Among these, Figure 8 has more cases in which the farthest points 403 in Border Control can play a vital role in retaining previously learned knowledge. Tables 2 and 3 summarize their performances.
513
+
514
+ ![](images/df00e79da7646169cbe201b37f70a3bcbb78337b977ceaff9e318f5cd16ecc4a.jpg)
515
+ Figure 7: Another batch distribution on the toy dataset.
516
+
517
+ Table 2: Performance of proposed methods on the dataset of Figure 7
518
+
519
+ <table><tr><td rowspan="2">Method</td><td colspan="6">Sample accuracy on accumulated dataset after Batch</td></tr><tr><td>#1</td><td>#2</td><td>#3</td><td>#4</td><td>#5</td><td>#6</td></tr><tr><td>No Incremental Learning</td><td>100.0%</td><td>73.14%</td><td>92.07%</td><td>97.56%</td><td>88.33%</td><td>91.67%</td></tr><tr><td>Fisher Shut-off (FS)</td><td>100.0%</td><td>99.43%</td><td>98.26%</td><td>99.54%</td><td>92.85%</td><td>97.78%</td></tr><tr><td>Frac. Data Ret. (FDR)[10%]</td><td>100.0%</td><td>97.43%</td><td>96.13%</td><td>98.93%</td><td>98.75%</td><td>96.78%</td></tr><tr><td>FDR[20%]</td><td>100.0%</td><td>98.86%</td><td>98.84%</td><td>99.69%</td><td>99.75%</td><td>98.89%</td></tr><tr><td>Border Ctrl. (BC)[topk = 5]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>99.78%</td></tr><tr><td>BC[topk = 10]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td></tr><tr><td>FS + FDR[10%]</td><td>100.0%</td><td>99.71%</td><td>98.84%</td><td>99.54%</td><td>98.75%</td><td>98.33%</td></tr><tr><td>FS + FDR[20%]</td><td>100.0%</td><td>100.0%</td><td>99.23%</td><td>99.85%</td><td>99.87%</td><td>98.89%</td></tr><tr><td>FS +BC[topk = 5]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td></tr><tr><td>FS + BC[topk = 10]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td></tr></table>
520
+
521
+ ![](images/198dee612d5cdc63ee02fcb729ec0a2fd31f55ddf925d97353d37ff12ee507bd.jpg)
522
+ Figure 8: Yet another batch distribution on the toy dataset.
523
+
524
+ Table 3: Performance of proposed methods on the dataset of Figure 8
525
+
526
+ <table><tr><td rowspan="2">Method</td><td colspan="6">Sample accuracy on accumulated dataset after Batch</td></tr><tr><td>#1</td><td>#2</td><td>#3</td><td>#4</td><td>#5</td><td>#6</td></tr><tr><td>No Incremental Learning</td><td>100.0%</td><td>89.34%</td><td>93.80%</td><td>95.60%</td><td>97.81%</td><td>84.78%</td></tr><tr><td>Fisher Shut-off (FS)</td><td>100.0%</td><td>95.36%</td><td>99.59%</td><td>99.55%</td><td>99.36%</td><td>94.56%</td></tr><tr><td>Frac. Data Ret. (FDR)[10%]</td><td>100.0%</td><td>95.08%</td><td>96.07%</td><td>97.42%</td><td>99.61%</td><td>98.67%</td></tr><tr><td>FDR[20%]</td><td>100.0%</td><td>98.36%</td><td>98.97%</td><td>99.85%</td><td>100.0%</td><td>99.67%</td></tr><tr><td>Border Ctrl. (BC)[top k = 5]</td><td>100.0%</td><td>100.0%</td><td>99.79%</td><td>99.85%</td><td>100.0%</td><td>100.0%</td></tr><tr><td>BC[topk = 10]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>99.74%</td><td>100.0%</td></tr><tr><td>FS + FDR[10%]</td><td>100.0%</td><td>95.36%</td><td>99.79%</td><td>98.48%</td><td>99.61%</td><td>98.89%</td></tr><tr><td>FS + FDR[20%]</td><td>100.0%</td><td>98.36%</td><td>99.79%</td><td>99.69%</td><td>100.0%</td><td>99.67%</td></tr><tr><td>FS+BC[topk = 5]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>99.74%</td><td>100.0%</td></tr><tr><td>FS + BC[topk = 10]</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td><td>100.0%</td></tr></table>