Add DeepFense-CodecFake-W2V2-MLP submission for LRLspoof (1-SRR=7.2898)

#31
by korallll - opened
submissions/deepfense-codecfake-w2v2-mlp.yaml ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: DeepFense-CodecFake-W2V2-MLP
4
+ slug: deepfense-codecfake-w2v2-mlp
5
+ description: 'DeepFense (Kheir et al., 2026) CodecFake recipe: a fully fine-tuned wav2vec 2.0 XLS-R
6
+ 300M front-end (last transformer layer, features_only) followed by an MLP back-end (Linear 1024->128,
7
+ LayerNorm, ReLU, mean pooling over time) and a 2-way linear CrossEntropy head. Trained on the Codecfake
8
+ corpus (Xie et al., AISHELL-3/VCTK-based codec re-synthesis) without data augmentation, seed 42; the
9
+ released best_model.pth (epoch 2) as is. Input as the DeepFense eval pipeline: raw float waveform
10
+ (no normalisation), first 64000 samples (4 s), shorter clips repeat-padded. score = logit(bonafide)
11
+ - logit(spoof); FP32.'
12
+ code: https://github.com/Yaselley/deepfense-framework/tree/3f2fe1146c48ea4bacc480dd96d806b6c72c09cf
13
+ checkpoint: https://huggingface.co/DeepFense/CodecFake_Wav2Vec2_MLP_NoAug_Seed42/blob/c08bd2125c83e6c8fdc64566bfbfda18dff4dfb4/best_model.pth
14
+ params_millions: 317.522306
15
+ paper:
16
+ arxiv_id: '2604.08450'
17
+ url: https://arxiv.org/abs/2604.08450
18
+ bibtex: "@article{kheir2026deepfense,\n title={DeepFense: A Unified, Modular, and Extensible Framework\
19
+ \ for Robust Deepfake Audio Detection},\n author={Kheir, Yassine El and Das, Arnab and Xiao, Yixuan\
20
+ \ and Wang, Xin and Kallel, Feidi and Erdogan, Enes Erdem and Vu, Ngoc Thang and Polzehl, Tim and\
21
+ \ Moeller, Sebastian},\n journal={arXiv preprint arXiv:2604.08450},\n year={2026}\n}\n"
22
+ dataset:
23
+ id: lab260/LRLspoof
24
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
25
+ split: test
26
+ scores:
27
+ srr_complement: 7.289776094969291
28
+ n_trials: 1304455
29
+ n_skipped: 286
30
+ calibration:
31
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
32
+ threshold: -5.76036528
33
+ artifact:
34
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DeepFense-CodecFake-W2V2-MLP/resolve/6f2a68c8eda6f0048ea40ee80379c42c262e3a6c/.eval_results/lab260/LRLspoof/scores.txt
35
+ scores_sha256: 87a70142bdd0d087102f3af25b4098360f6a0aa2c6e7cd1069b78c131903a5dc
36
+ bench_version: speech-spoof-bench==0.4.1
37
+ reproduction:
38
+ reproduced_by: SpeechAntiSpoofingBenchmarks
39
+ reproduced_at: '2026-11-09'
40
+ reproduced_bench_version: speech-spoof-bench==0.4.1
41
+ match: inference
42
+ submitter:
43
+ hf_username: korallll
44
+ contact: k.n.borodin@mtuci.ru
45
+ submitted_at: '2026-11-09'
46
+ notes: 'DeepFense-CodecFake-W2V2-MLP = HF DeepFense/CodecFake_Wav2Vec2_MLP_NoAug_Seed42 @ c08bd2125c83e6c8fdc64566bfbfda18dff4dfb4,
47
+ best_model.pth (sha256 1d4bf44afd0f7e887266c8dc49875b0ddb60a07fb16279fa2d56598519ea3c1a; a training
48
+ checkpoint with model_state + Adam optimizer_state, epoch 2, step 46298). The HF repo states no weight
49
+ license, so the weights are linked, not re-hosted; the wrapper loads only model_state, saved unchanged
50
+ as model_state.safetensors (sha256 5c97bd4ae3a2075aa8df309bde8b7dd7f9ce540a50df969d4ff0ba7a74a560a5,
51
+ 435 fp32 tensors). Code: Yaselley/deepfense-framework @ 3f2fe11 (Apache-2.0) + the checkpoint''s config.yaml
52
+ (StandardDetector: wav2vec2 fairseq frontend on xlsr2_300m.pt, MLP backend projection [128], layer norm,
53
+ ReLU, mean pooling; CrossEntropy n_classes 2; label_map bonafide=1, spoof=0). XLS-R 300M is the fairseq
54
+ Wav2Vec2Model built from the xlsr2_300m.pt model config; all 435 checkpoint tensors load strictly (the
55
+ 7 pre-training quantizer/project_q/final_proj tensors are unused on the features_only path). Eval preprocessing
56
+ = StandardDataset + pad: soundfile float mono, first 64000 samples, np.tile repeat-pad, no normalisation;
57
+ no resampling; FP32. Score = logit1 - logit0 (higher = bona fide). The authors'' test.py EER uses get_score
58
+ = logit1 only (a different, non-monotone-equivalent score); see implementation-notes.md. Verified against
59
+ the authors'' ModularDetector + fairseq_local loader on 12 clips: max |d logits| 4.8e-6. Training: Codecfake,
60
+ no augmentation, seed 42. Self-reported EER (DeepFense deepfense-results Space): Codecfake 10.71, In-the-Wild
61
+ 8.10, ASVspoof2019 LA 4.15, ASVspoof2021 DF 6.10, ASVspoof2021 LA 15.46, ASVspoof5 29.93, ODSS 11.51,
62
+ HABLA 7.82, ADD22 track 3 7.48.'