---
license: cc-by-nc-4.0
language:
- en
pretty_name: RobustGenBench
size_categories:
- 100K<n<1M
task_categories:
- image-classification
task_ids:
- multi-class-image-classification
tags:
- adversarial-robustness
- common-corruptions
- robustness
- benchmark
- transfer-attacks
- vision-language-models
- autoattack
configs:
- config_name: caltech101
  default: true
  data_files:
  - split: train
    path: data/caltech101/train-*.parquet
  - split: validation
    path: data/caltech101/validation-*.parquet
  - split: test
    path: data/caltech101/test-*.parquet
  - split: common
    path: data/caltech101/common-*.parquet
  - split: random_linf_30
    path: data/caltech101/random_linf_30-*.parquet
  - split: clip_vitb16_l1_75
    path: data/caltech101/clip_vitb16_l1_75-*.parquet
  - split: clip_vitb16_l1_300
    path: data/caltech101/clip_vitb16_l1_300-*.parquet
  - split: clip_vitb16_l2_2
    path: data/caltech101/clip_vitb16_l2_2-*.parquet
  - split: clip_vitb16_l2_8
    path: data/caltech101/clip_vitb16_l2_8-*.parquet
  - split: clip_vitb16_linf_4
    path: data/caltech101/clip_vitb16_linf_4-*.parquet
  - split: clip_vitb16_linf_8
    path: data/caltech101/clip_vitb16_linf_8-*.parquet
  - split: clip_vitb16_linf_30
    path: data/caltech101/clip_vitb16_linf_30-*.parquet
  - split: clip_vith14_l1_75
    path: data/caltech101/clip_vith14_l1_75-*.parquet
  - split: clip_vith14_l1_300
    path: data/caltech101/clip_vith14_l1_300-*.parquet
  - split: clip_vith14_l2_2
    path: data/caltech101/clip_vith14_l2_2-*.parquet
  - split: clip_vith14_l2_8
    path: data/caltech101/clip_vith14_l2_8-*.parquet
  - split: clip_vith14_linf_4
    path: data/caltech101/clip_vith14_linf_4-*.parquet
  - split: clip_vith14_linf_8
    path: data/caltech101/clip_vith14_linf_8-*.parquet
  - split: clip_vith14_linf_30
    path: data/caltech101/clip_vith14_linf_30-*.parquet
  - split: metaclip_vith14_l1_75
    path: data/caltech101/metaclip_vith14_l1_75-*.parquet
  - split: metaclip_vith14_l1_300
    path: data/caltech101/metaclip_vith14_l1_300-*.parquet
  - split: metaclip_vith14_l2_2
    path: data/caltech101/metaclip_vith14_l2_2-*.parquet
  - split: metaclip_vith14_l2_8
    path: data/caltech101/metaclip_vith14_l2_8-*.parquet
  - split: metaclip_vith14_linf_4
    path: data/caltech101/metaclip_vith14_linf_4-*.parquet
  - split: metaclip_vith14_linf_8
    path: data/caltech101/metaclip_vith14_linf_8-*.parquet
  - split: metaclip_vith14_linf_30
    path: data/caltech101/metaclip_vith14_linf_30-*.parquet
  - split: siglip2_so400m_l1_75
    path: data/caltech101/siglip2_so400m_l1_75-*.parquet
  - split: siglip2_so400m_l1_300
    path: data/caltech101/siglip2_so400m_l1_300-*.parquet
  - split: siglip2_so400m_l2_2
    path: data/caltech101/siglip2_so400m_l2_2-*.parquet
  - split: siglip2_so400m_l2_8
    path: data/caltech101/siglip2_so400m_l2_8-*.parquet
  - split: siglip2_so400m_linf_4
    path: data/caltech101/siglip2_so400m_linf_4-*.parquet
  - split: siglip2_so400m_linf_8
    path: data/caltech101/siglip2_so400m_linf_8-*.parquet
  - split: siglip2_so400m_linf_30
    path: data/caltech101/siglip2_so400m_linf_30-*.parquet
  - split: siglip2_base_linf_30
    path: data/caltech101/siglip2_base_linf_30-*.parquet
- config_name: fgvc-aircraft-2013b
  data_files:
  - split: train
    path: data/fgvc-aircraft-2013b/train-*.parquet
  - split: validation
    path: data/fgvc-aircraft-2013b/validation-*.parquet
  - split: test
    path: data/fgvc-aircraft-2013b/test-*.parquet
  - split: common
    path: data/fgvc-aircraft-2013b/common-*.parquet
  - split: random_linf_30
    path: data/fgvc-aircraft-2013b/random_linf_30-*.parquet
  - split: clip_vitb16_l1_75
    path: data/fgvc-aircraft-2013b/clip_vitb16_l1_75-*.parquet
  - split: clip_vitb16_l1_300
    path: data/fgvc-aircraft-2013b/clip_vitb16_l1_300-*.parquet
  - split: clip_vitb16_l2_2
    path: data/fgvc-aircraft-2013b/clip_vitb16_l2_2-*.parquet
  - split: clip_vitb16_l2_8
    path: data/fgvc-aircraft-2013b/clip_vitb16_l2_8-*.parquet
  - split: clip_vitb16_linf_4
    path: data/fgvc-aircraft-2013b/clip_vitb16_linf_4-*.parquet
  - split: clip_vitb16_linf_8
    path: data/fgvc-aircraft-2013b/clip_vitb16_linf_8-*.parquet
  - split: clip_vitb16_linf_30
    path: data/fgvc-aircraft-2013b/clip_vitb16_linf_30-*.parquet
  - split: clip_vith14_l1_75
    path: data/fgvc-aircraft-2013b/clip_vith14_l1_75-*.parquet
  - split: clip_vith14_l1_300
    path: data/fgvc-aircraft-2013b/clip_vith14_l1_300-*.parquet
  - split: clip_vith14_l2_2
    path: data/fgvc-aircraft-2013b/clip_vith14_l2_2-*.parquet
  - split: clip_vith14_l2_8
    path: data/fgvc-aircraft-2013b/clip_vith14_l2_8-*.parquet
  - split: clip_vith14_linf_4
    path: data/fgvc-aircraft-2013b/clip_vith14_linf_4-*.parquet
  - split: clip_vith14_linf_8
    path: data/fgvc-aircraft-2013b/clip_vith14_linf_8-*.parquet
  - split: clip_vith14_linf_30
    path: data/fgvc-aircraft-2013b/clip_vith14_linf_30-*.parquet
  - split: metaclip_vith14_l1_75
    path: data/fgvc-aircraft-2013b/metaclip_vith14_l1_75-*.parquet
  - split: metaclip_vith14_l1_300
    path: data/fgvc-aircraft-2013b/metaclip_vith14_l1_300-*.parquet
  - split: metaclip_vith14_l2_2
    path: data/fgvc-aircraft-2013b/metaclip_vith14_l2_2-*.parquet
  - split: metaclip_vith14_l2_8
    path: data/fgvc-aircraft-2013b/metaclip_vith14_l2_8-*.parquet
  - split: metaclip_vith14_linf_4
    path: data/fgvc-aircraft-2013b/metaclip_vith14_linf_4-*.parquet
  - split: metaclip_vith14_linf_8
    path: data/fgvc-aircraft-2013b/metaclip_vith14_linf_8-*.parquet
  - split: metaclip_vith14_linf_30
    path: data/fgvc-aircraft-2013b/metaclip_vith14_linf_30-*.parquet
  - split: siglip2_base_linf_30
    path: data/fgvc-aircraft-2013b/siglip2_base_linf_30-*.parquet
- config_name: flowers-102
  data_files:
  - split: train
    path: data/flowers-102/train-*.parquet
  - split: validation
    path: data/flowers-102/validation-*.parquet
  - split: test
    path: data/flowers-102/test-*.parquet
  - split: common
    path: data/flowers-102/common-*.parquet
  - split: random_linf_30
    path: data/flowers-102/random_linf_30-*.parquet
  - split: clip_vitb16_l1_75
    path: data/flowers-102/clip_vitb16_l1_75-*.parquet
  - split: clip_vitb16_l1_300
    path: data/flowers-102/clip_vitb16_l1_300-*.parquet
  - split: clip_vitb16_l2_2
    path: data/flowers-102/clip_vitb16_l2_2-*.parquet
  - split: clip_vitb16_l2_8
    path: data/flowers-102/clip_vitb16_l2_8-*.parquet
  - split: clip_vitb16_linf_4
    path: data/flowers-102/clip_vitb16_linf_4-*.parquet
  - split: clip_vitb16_linf_8
    path: data/flowers-102/clip_vitb16_linf_8-*.parquet
  - split: clip_vitb16_linf_30
    path: data/flowers-102/clip_vitb16_linf_30-*.parquet
  - split: clip_vith14_l1_75
    path: data/flowers-102/clip_vith14_l1_75-*.parquet
  - split: clip_vith14_l1_300
    path: data/flowers-102/clip_vith14_l1_300-*.parquet
  - split: clip_vith14_l2_2
    path: data/flowers-102/clip_vith14_l2_2-*.parquet
  - split: clip_vith14_l2_8
    path: data/flowers-102/clip_vith14_l2_8-*.parquet
  - split: clip_vith14_linf_4
    path: data/flowers-102/clip_vith14_linf_4-*.parquet
  - split: clip_vith14_linf_8
    path: data/flowers-102/clip_vith14_linf_8-*.parquet
  - split: clip_vith14_linf_30
    path: data/flowers-102/clip_vith14_linf_30-*.parquet
  - split: metaclip_vith14_l1_75
    path: data/flowers-102/metaclip_vith14_l1_75-*.parquet
  - split: metaclip_vith14_l1_300
    path: data/flowers-102/metaclip_vith14_l1_300-*.parquet
  - split: metaclip_vith14_l2_2
    path: data/flowers-102/metaclip_vith14_l2_2-*.parquet
  - split: metaclip_vith14_l2_8
    path: data/flowers-102/metaclip_vith14_l2_8-*.parquet
  - split: metaclip_vith14_linf_4
    path: data/flowers-102/metaclip_vith14_linf_4-*.parquet
  - split: metaclip_vith14_linf_8
    path: data/flowers-102/metaclip_vith14_linf_8-*.parquet
  - split: metaclip_vith14_linf_30
    path: data/flowers-102/metaclip_vith14_linf_30-*.parquet
  - split: siglip2_so400m_l1_75
    path: data/flowers-102/siglip2_so400m_l1_75-*.parquet
  - split: siglip2_so400m_l1_300
    path: data/flowers-102/siglip2_so400m_l1_300-*.parquet
  - split: siglip2_so400m_l2_2
    path: data/flowers-102/siglip2_so400m_l2_2-*.parquet
  - split: siglip2_so400m_l2_8
    path: data/flowers-102/siglip2_so400m_l2_8-*.parquet
  - split: siglip2_so400m_linf_4
    path: data/flowers-102/siglip2_so400m_linf_4-*.parquet
  - split: siglip2_so400m_linf_8
    path: data/flowers-102/siglip2_so400m_linf_8-*.parquet
  - split: siglip2_so400m_linf_30
    path: data/flowers-102/siglip2_so400m_linf_30-*.parquet
  - split: siglip2_base_linf_30
    path: data/flowers-102/siglip2_base_linf_30-*.parquet
- config_name: oxford-iiit-pet
  data_files:
  - split: train
    path: data/oxford-iiit-pet/train-*.parquet
  - split: validation
    path: data/oxford-iiit-pet/validation-*.parquet
  - split: test
    path: data/oxford-iiit-pet/test-*.parquet
  - split: common
    path: data/oxford-iiit-pet/common-*.parquet
  - split: random_linf_30
    path: data/oxford-iiit-pet/random_linf_30-*.parquet
  - split: clip_vitb16_l1_75
    path: data/oxford-iiit-pet/clip_vitb16_l1_75-*.parquet
  - split: clip_vitb16_l1_300
    path: data/oxford-iiit-pet/clip_vitb16_l1_300-*.parquet
  - split: clip_vitb16_l2_2
    path: data/oxford-iiit-pet/clip_vitb16_l2_2-*.parquet
  - split: clip_vitb16_l2_8
    path: data/oxford-iiit-pet/clip_vitb16_l2_8-*.parquet
  - split: clip_vitb16_linf_4
    path: data/oxford-iiit-pet/clip_vitb16_linf_4-*.parquet
  - split: clip_vitb16_linf_8
    path: data/oxford-iiit-pet/clip_vitb16_linf_8-*.parquet
  - split: clip_vitb16_linf_30
    path: data/oxford-iiit-pet/clip_vitb16_linf_30-*.parquet
  - split: clip_vith14_l1_75
    path: data/oxford-iiit-pet/clip_vith14_l1_75-*.parquet
  - split: clip_vith14_l1_300
    path: data/oxford-iiit-pet/clip_vith14_l1_300-*.parquet
  - split: clip_vith14_l2_2
    path: data/oxford-iiit-pet/clip_vith14_l2_2-*.parquet
  - split: clip_vith14_l2_8
    path: data/oxford-iiit-pet/clip_vith14_l2_8-*.parquet
  - split: clip_vith14_linf_4
    path: data/oxford-iiit-pet/clip_vith14_linf_4-*.parquet
  - split: clip_vith14_linf_8
    path: data/oxford-iiit-pet/clip_vith14_linf_8-*.parquet
  - split: clip_vith14_linf_30
    path: data/oxford-iiit-pet/clip_vith14_linf_30-*.parquet
  - split: metaclip_vith14_l1_75
    path: data/oxford-iiit-pet/metaclip_vith14_l1_75-*.parquet
  - split: metaclip_vith14_l1_300
    path: data/oxford-iiit-pet/metaclip_vith14_l1_300-*.parquet
  - split: metaclip_vith14_l2_2
    path: data/oxford-iiit-pet/metaclip_vith14_l2_2-*.parquet
  - split: metaclip_vith14_l2_8
    path: data/oxford-iiit-pet/metaclip_vith14_l2_8-*.parquet
  - split: metaclip_vith14_linf_4
    path: data/oxford-iiit-pet/metaclip_vith14_linf_4-*.parquet
  - split: metaclip_vith14_linf_8
    path: data/oxford-iiit-pet/metaclip_vith14_linf_8-*.parquet
  - split: metaclip_vith14_linf_30
    path: data/oxford-iiit-pet/metaclip_vith14_linf_30-*.parquet
  - split: siglip2_base_linf_30
    path: data/oxford-iiit-pet/siglip2_base_linf_30-*.parquet
- config_name: stanford_cars
  data_files:
  - split: train
    path: data/stanford_cars/train-*.parquet
  - split: validation
    path: data/stanford_cars/validation-*.parquet
  - split: test
    path: data/stanford_cars/test-*.parquet
  - split: common
    path: data/stanford_cars/common-*.parquet
  - split: random_linf_30
    path: data/stanford_cars/random_linf_30-*.parquet
  - split: clip_vitb16_l1_75
    path: data/stanford_cars/clip_vitb16_l1_75-*.parquet
  - split: clip_vitb16_l1_300
    path: data/stanford_cars/clip_vitb16_l1_300-*.parquet
  - split: clip_vitb16_l2_2
    path: data/stanford_cars/clip_vitb16_l2_2-*.parquet
  - split: clip_vitb16_l2_8
    path: data/stanford_cars/clip_vitb16_l2_8-*.parquet
  - split: clip_vitb16_linf_4
    path: data/stanford_cars/clip_vitb16_linf_4-*.parquet
  - split: clip_vitb16_linf_8
    path: data/stanford_cars/clip_vitb16_linf_8-*.parquet
  - split: clip_vitb16_linf_30
    path: data/stanford_cars/clip_vitb16_linf_30-*.parquet
  - split: clip_vith14_l1_75
    path: data/stanford_cars/clip_vith14_l1_75-*.parquet
  - split: clip_vith14_l1_300
    path: data/stanford_cars/clip_vith14_l1_300-*.parquet
  - split: clip_vith14_l2_2
    path: data/stanford_cars/clip_vith14_l2_2-*.parquet
  - split: clip_vith14_l2_8
    path: data/stanford_cars/clip_vith14_l2_8-*.parquet
  - split: clip_vith14_linf_4
    path: data/stanford_cars/clip_vith14_linf_4-*.parquet
  - split: clip_vith14_linf_8
    path: data/stanford_cars/clip_vith14_linf_8-*.parquet
  - split: clip_vith14_linf_30
    path: data/stanford_cars/clip_vith14_linf_30-*.parquet
  - split: metaclip_vith14_l1_75
    path: data/stanford_cars/metaclip_vith14_l1_75-*.parquet
  - split: metaclip_vith14_l1_300
    path: data/stanford_cars/metaclip_vith14_l1_300-*.parquet
  - split: metaclip_vith14_l2_2
    path: data/stanford_cars/metaclip_vith14_l2_2-*.parquet
  - split: metaclip_vith14_l2_8
    path: data/stanford_cars/metaclip_vith14_l2_8-*.parquet
  - split: metaclip_vith14_linf_4
    path: data/stanford_cars/metaclip_vith14_linf_4-*.parquet
  - split: metaclip_vith14_linf_8
    path: data/stanford_cars/metaclip_vith14_linf_8-*.parquet
  - split: metaclip_vith14_linf_30
    path: data/stanford_cars/metaclip_vith14_linf_30-*.parquet
  - split: siglip2_base_linf_30
    path: data/stanford_cars/siglip2_base_linf_30-*.parquet
- config_name: uc-merced-land-use-dataset
  data_files:
  - split: train
    path: data/uc-merced-land-use-dataset/train-*.parquet
  - split: validation
    path: data/uc-merced-land-use-dataset/validation-*.parquet
  - split: test
    path: data/uc-merced-land-use-dataset/test-*.parquet
  - split: common
    path: data/uc-merced-land-use-dataset/common-*.parquet
  - split: random_linf_30
    path: data/uc-merced-land-use-dataset/random_linf_30-*.parquet
  - split: clip_vitb16_l1_75
    path: data/uc-merced-land-use-dataset/clip_vitb16_l1_75-*.parquet
  - split: clip_vitb16_l1_300
    path: data/uc-merced-land-use-dataset/clip_vitb16_l1_300-*.parquet
  - split: clip_vitb16_l2_2
    path: data/uc-merced-land-use-dataset/clip_vitb16_l2_2-*.parquet
  - split: clip_vitb16_l2_8
    path: data/uc-merced-land-use-dataset/clip_vitb16_l2_8-*.parquet
  - split: clip_vitb16_linf_4
    path: data/uc-merced-land-use-dataset/clip_vitb16_linf_4-*.parquet
  - split: clip_vitb16_linf_8
    path: data/uc-merced-land-use-dataset/clip_vitb16_linf_8-*.parquet
  - split: clip_vitb16_linf_30
    path: data/uc-merced-land-use-dataset/clip_vitb16_linf_30-*.parquet
  - split: clip_vith14_l1_75
    path: data/uc-merced-land-use-dataset/clip_vith14_l1_75-*.parquet
  - split: clip_vith14_l1_300
    path: data/uc-merced-land-use-dataset/clip_vith14_l1_300-*.parquet
  - split: clip_vith14_l2_2
    path: data/uc-merced-land-use-dataset/clip_vith14_l2_2-*.parquet
  - split: clip_vith14_l2_8
    path: data/uc-merced-land-use-dataset/clip_vith14_l2_8-*.parquet
  - split: clip_vith14_linf_4
    path: data/uc-merced-land-use-dataset/clip_vith14_linf_4-*.parquet
  - split: clip_vith14_linf_8
    path: data/uc-merced-land-use-dataset/clip_vith14_linf_8-*.parquet
  - split: clip_vith14_linf_30
    path: data/uc-merced-land-use-dataset/clip_vith14_linf_30-*.parquet
  - split: metaclip_vith14_l1_75
    path: data/uc-merced-land-use-dataset/metaclip_vith14_l1_75-*.parquet
  - split: metaclip_vith14_l1_300
    path: data/uc-merced-land-use-dataset/metaclip_vith14_l1_300-*.parquet
  - split: metaclip_vith14_l2_2
    path: data/uc-merced-land-use-dataset/metaclip_vith14_l2_2-*.parquet
  - split: metaclip_vith14_l2_8
    path: data/uc-merced-land-use-dataset/metaclip_vith14_l2_8-*.parquet
  - split: metaclip_vith14_linf_4
    path: data/uc-merced-land-use-dataset/metaclip_vith14_linf_4-*.parquet
  - split: metaclip_vith14_linf_8
    path: data/uc-merced-land-use-dataset/metaclip_vith14_linf_8-*.parquet
  - split: metaclip_vith14_linf_30
    path: data/uc-merced-land-use-dataset/metaclip_vith14_linf_30-*.parquet
  - split: siglip2_so400m_l1_75
    path: data/uc-merced-land-use-dataset/siglip2_so400m_l1_75-*.parquet
  - split: siglip2_so400m_l1_300
    path: data/uc-merced-land-use-dataset/siglip2_so400m_l1_300-*.parquet
  - split: siglip2_so400m_l2_2
    path: data/uc-merced-land-use-dataset/siglip2_so400m_l2_2-*.parquet
  - split: siglip2_so400m_l2_8
    path: data/uc-merced-land-use-dataset/siglip2_so400m_l2_8-*.parquet
  - split: siglip2_so400m_linf_4
    path: data/uc-merced-land-use-dataset/siglip2_so400m_linf_4-*.parquet
  - split: siglip2_so400m_linf_8
    path: data/uc-merced-land-use-dataset/siglip2_so400m_linf_8-*.parquet
  - split: siglip2_so400m_linf_30
    path: data/uc-merced-land-use-dataset/siglip2_so400m_linf_30-*.parquet
  - split: siglip2_base_linf_30
    path: data/uc-merced-land-use-dataset/siglip2_base_linf_30-*.parquet
---

# RobustGenBench

**A standardized benchmark for robust generalization of image classifiers across common and adversarial ($\ell_1, \ell_2, \ell_\infty$) perturbations.**

📄 Paper · 💻 [Code](https://github.com/legolasflagstaff/robust_training) · 📜 [Citation](#citation)

RobustGenBench evaluates how well a classifier keeps its accuracy on perturbation types it was not trained against. It provides six downstream classification datasets with fixed splits at 224×224 resolution, together with perturbed copies of every test set:

- **Common corruptions** (11 types, severity 3): model-independent, the same images for every model.
- **Random uniform noise** ($\ell_\infty \le 30/255$): a control for the adversarial perturbations.
- **Transfer adversarial perturbations** (136,180 images): crafted with AutoAttack on 5 open-weight vision-language surrogates, across 7 budgets in $\ell_1$, $\ell_2$, $\ell_\infty$.

Because the adversarial images are pre-computed, any model, including API-only systems such as vision-enabled LLMs, can be evaluated **by inference alone**, without GPU access. For white-box evaluation, attack your own model on the clean `test` split.

In total: 5,420 clean test images and 147,020 perturbed test images.

## Quick start

```python
from datasets import load_dataset

# One config per dataset; one split per perturbation setting.
clean = load_dataset("legolasflagstaff/RobustGenBench", "uc-merced-land-use-dataset", split="test")
adv   = load_dataset("legolasflagstaff/RobustGenBench", "uc-merced-land-use-dataset", split="clip_vith14_linf_30")

row = adv[0]
row["image"]                                  # PIL.Image, 224×224 RGB
adv.features["label"].int2str(row["label"])   # class name
row["image_id"]                               # aligns with the clean image of the same id
```

Evaluate a classifier on every perturbation setting of one dataset:

```python
from datasets import load_dataset, get_dataset_split_names

repo, cfg = "legolasflagstaff/RobustGenBench", "flowers-102"
for split in get_dataset_split_names(repo, cfg):
    if split in ("train", "validation"):
        continue
    ds = load_dataset(repo, cfg, split=split)
    acc = sum(predict(x["image"]) == x["label"] for x in ds) / len(ds)   # your model here
    print(f"{split:24s} {acc:.3f}")
```

Splits are aligned by `image_id`: row *i* of every perturbed split is a perturbed version of row *i* of `test`, with the same label.

## Datasets

| Config | Dataset | Domain | Granularity | Classes | Train | Validation | Test |
|---|---|---|---|---|---|---|---|
| `caltech101` | Caltech-101 | objects | coarse | 101 | 5,899 | 1,042 | 1,000 |
| `fgvc-aircraft-2013b` | FGVC-Aircraft (2013b) | aircraft | fine | 100 | 3,334 | 3,333 | 1,000 |
| `flowers-102` | Oxford Flowers-102 | flowers | fine | 102 | 1,020 | 1,020 | 1,000 |
| `oxford-iiit-pet` | Oxford-IIIT Pet | pets | fine | 37 | 3,128 | 552 | 1,000 |
| `stanford_cars` | Stanford Cars | cars | fine | 196 | 6,922 | 1,222 | 1,000 |
| `uc-merced-land-use-dataset` | UC Merced Land Use | satellite | coarse | 21 | 1,428 | 252 | 420 |

Test sets are standardized to at most 1,000 images by uniform sampling, preserving all classes. All images are resized to 224×224×3 and stored as lossless PNG.

## Splits

Each config (dataset) contains the following splits.

| Split | Content |
|---|---|
| `train`, `validation` | Clean images, for fine-tuning. |
| `test` | Clean test images, the reference for all perturbed splits. |
| `common` | Each test image with one of 11 common corruptions at severity 3 (column `corruption`). |
| `random_linf_30` | Uniform random noise with $\lVert\delta\rVert_\infty \le 30/255$. |
| `<surrogate>_<norm>_<eps>` | Transfer adversarial images, e.g. `clip_vitb16_linf_30`, `metaclip_vith14_l1_300`. |

### Common corruptions

Shot noise, impulse noise, defocus blur, motion blur, zoom blur, snow, brightness, contrast, elastic transform, pixelate, JPEG compression, all at severity 3 of the ImageNet-C scale. Test image *i* receives corruption *i* mod 11; the assignment is deterministic, so every model is evaluated on identical (image, corruption) pairs.

### Adversarial perturbations

Perturbations are crafted with the standard AutoAttack ensemble (APGD-CE, APGD-T, FAB-T, Square) against zero-shot surrogate classifiers. Each surrogate predicts the class whose text embedding (`"a photo of a {class}"`) is closest to the image embedding. AutoAttack only perturbs images the surrogate classifies correctly; other images are copied unchanged (`is_modified = false`).

| Norm | Budgets $\epsilon$ | Split suffix |
|---|---|---|
| $\ell_1$ | 75, 300 | `l1_75`, `l1_300` |
| $\ell_2$ | 2, 8 | `l2_2`, `l2_8` |
| $\ell_\infty$ | 4/255, 8/255, 30/255 | `linf_4`, `linf_8`, `linf_30` |

Budgets are expressed on pixel values in $[0, 1]$; the `epsilon` column stores the exact value (e.g. 0.1176 for 30/255).

| Surrogate prefix | Model | Checkpoint |
|---|---|---|
| `clip_vitb16` | CLIP ViT-B/16 | open_clip `ViT-B-16` / `laion2b_s34b_b88k` |
| `clip_vith14` | CLIP ViT-H/14 | open_clip `ViT-H-14` / `laion2b_s32b_b79k` |
| `metaclip_vith14` | MetaCLIP ViT-H/14 | open_clip `ViT-H-14-quickgelu` / `metaclip_fullcc` |
| `siglip2_so400m` | SigLIP2 SO400M | `google/siglip2-so400m-patch14-384` |
| `siglip2_base` | SigLIP2 base | `google/siglip2-base-patch16-224` |

### Coverage

SigLIP-family surrogates classify the fine-grained datasets poorly, and a surrogate that misclassifies the clean image provides no attack signal. SigLIP2 SO400M is therefore provided on the three coarse-grained datasets only. SigLIP2 base is provided at $\ell_\infty$ 30/255 only; on the fine-grained datasets it leaves most images unmodified (see `is_modified`).

| Split | Caltech-101 | FGVC-Aircraft (2013b) | Oxford Flowers-102 | Oxford-IIIT Pet | Stanford Cars | UC Merced Land Use |
|---|---|---|---|---|---|---|
| `clip_vitb16_l1_75` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vitb16_l1_300` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vitb16_l2_2` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vitb16_l2_8` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vitb16_linf_4` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vitb16_linf_8` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vitb16_linf_30` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vith14_l1_75` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vith14_l1_300` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vith14_l2_2` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vith14_l2_8` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vith14_linf_4` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vith14_linf_8` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `clip_vith14_linf_30` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `metaclip_vith14_l1_75` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `metaclip_vith14_l1_300` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `metaclip_vith14_l2_2` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `metaclip_vith14_l2_8` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `metaclip_vith14_linf_4` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `metaclip_vith14_linf_8` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `metaclip_vith14_linf_30` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| `siglip2_so400m_l1_75` | ✓ | – | ✓ | – | – | ✓ |
| `siglip2_so400m_l1_300` | ✓ | – | ✓ | – | – | ✓ |
| `siglip2_so400m_l2_2` | ✓ | – | ✓ | – | – | ✓ |
| `siglip2_so400m_l2_8` | ✓ | – | ✓ | – | – | ✓ |
| `siglip2_so400m_linf_4` | ✓ | – | ✓ | – | – | ✓ |
| `siglip2_so400m_linf_8` | ✓ | – | ✓ | – | – | ✓ |
| `siglip2_so400m_linf_30` | ✓ | – | ✓ | – | – | ✓ |
| `siglip2_base_linf_30` | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |

## Data fields

| Field | Type | Description |
|---|---|---|
| `image` | image | 224×224 RGB PNG. |
| `label` | class label | Ground-truth class (use `features["label"].names` for names). |
| `image_id` | int32 | Index of the source image within its split; aligns perturbed splits with `test`. |
| `perturbation` | string | `clean`, `common_corruption_severity3`, `random_uniform`, or `transfer_autoattack`. |
| `surrogate` | string | Surrogate used to craft the attack (null otherwise). |
| `norm` | string | `l1`, `l2`, or `linf` (null for clean and common). |
| `epsilon` | float32 | Perturbation budget (null for clean and common). |
| `corruption` | string | Common corruption applied (`common` split only). |
| `is_modified` | bool | Whether the image differs from its clean counterpart. |

## Evaluation protocol

RobustGenBench reports accuracy per (dataset, perturbation setting), aggregated with:

- **Sum score**: sum of accuracies over all (dataset, setting) pairs. Pool-independent; use it to compare new models with published results.
- **Borda score**: rank-based consistency across pairs, only meaningful within a fixed pool of models.

The paper uses two sets of settings:

- **Fine-tuned vision models (5 settings):** `test`, `common`, and white-box AutoAttack on the evaluated model at $\ell_1 = 75$, $\ell_2 = 2$, $\ell_\infty = 4/255$, crafted from the clean `test` split. Maximum sum score: 30.
- **Black-box models, e.g. vision-enabled LLMs (9 settings):** `test`, `common`, and the 7 transfer budgets, taking the strongest surrogate per dataset and setting.

## Intended use and limitations

**Intended use:** held-out evaluation of the robustness of image classifiers, fine-tuned or zero-shot, including API-only models. Do not train on the test or perturbed splits.

**Limitations**
- Transfer attacks give an upper bound on robust accuracy. They are not a substitute for white-box or adaptive attacks when model access is available.
- All surrogates are attention-based, so transfer to convolutional targets is weaker.
- Budgets are not matched across norms; compare norms only at the intensities provided.
- Common corruptions are evaluated at a single severity.
- Coverage is limited to six source taxonomies; results do not imply robustness on open-world imagery.

**Biases.** The source datasets inherit the biases of their collection, for example Western car models (Stanford Cars) and breeds common in the UK (Oxford-IIIT Pet). Some images may incidentally show people or license plates; no additional anonymization was performed.

## License

The perturbed images, splits, and metadata produced for RobustGenBench are released under **[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)**. This license does not override the terms of the source datasets: every image is a derivative of an upstream image whose rights remain with its owners. Users must comply with both licenses; where they differ, the more restrictive applies.

| Source | Homepage | Upstream terms |
|---|---|---|
| Caltech-101 | [https://data.caltech.edu/records/mzrjq-6wc02](https://data.caltech.edu/records/mzrjq-6wc02) | CC BY 4.0 |
| FGVC-Aircraft (2013b) | [https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/](https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/) | non-commercial research use only |
| Oxford Flowers-102 | [https://www.robots.ox.ac.uk/~vgg/data/flowers/102/](https://www.robots.ox.ac.uk/~vgg/data/flowers/102/) | see upstream terms |
| Oxford-IIIT Pet | [https://www.robots.ox.ac.uk/~vgg/data/pets/](https://www.robots.ox.ac.uk/~vgg/data/pets/) | CC BY-SA 4.0 |
| Stanford Cars | [https://ai.stanford.edu/~jkrause/cars/car_dataset.html](https://ai.stanford.edu/~jkrause/cars/car_dataset.html) | see upstream terms (research use) |
| UC Merced Land Use | [http://weegee.vision.ucmerced.edu/datasets/landuse.html](http://weegee.vision.ucmerced.edu/datasets/landuse.html) | USGS public-domain imagery |

## Citation

```bibtex
@inproceedings{heuillet2026robustgenbench,
  title     = {RobustGenBench: A Benchmark for Robust Generalization to Adversarial and Common Perturbations, with Applications to Vision and Vision-Enabled Large Language Models},
  author    = {Heuillet, Maxime and Ngnaw{\'e}, Jonas and Pequignot, Yann and Bhagwatkar, Rishika and Larouche, Alexandre and Gagn{\'e}, Christian and Rish, Irina and Ahmad, Ola and Durand, Audrey},
  booktitle = {TODO},
  year      = {2026}
}
```

Please also cite the six source datasets.

## Versions

- **v2.0** (current): Parquet layout with Dataset Viewer support; restricted to the surrogates reported in the paper.
- **v1.0**: original `.tar.zst` archives, kept for reproducibility under the git tag [`v1.0-archives`](https://huggingface.co/datasets/legolasflagstaff/RobustGenBench/tree/v1.0-archives).