RobustGenBench

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

📄 Paper · 💻 Code · 🌐 Project page · 📜 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

from datasets import load_dataset

repo = "legolasflagstaff/RobustGenBench"

# Clean and model-independent splits: config "<dataset>".
clean  = load_dataset(repo, "uc-merced-land-use-dataset", split="test")
common = load_dataset(repo, "uc-merced-land-use-dataset", split="common")

# Transfer attacks: config "<dataset>-<surrogate>", one split per budget.
adv = load_dataset(repo, "uc-merced-land-use-dataset-clip_vith14", split="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:

from datasets import load_dataset, get_dataset_config_names, get_dataset_split_names

repo, name = "legolasflagstaff/RobustGenBench", "flowers-102"
settings = [(name, s) for s in ("test", "common", "random_linf_30")]
for cfg in get_dataset_config_names(repo):
    if cfg.startswith(name + "-"):                                   # one config per surrogate
        settings += [(cfg, s) for s in get_dataset_split_names(repo, cfg)]

for cfg, split in settings:
    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"{cfg:34s} {split:15s} {acc:.3f}")

All perturbed splits are aligned with test 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.

Configs and splits

Config	Split	Content
<dataset>	train, validation	Clean images, for fine-tuning.
<dataset>	test	Clean test images, the reference for all perturbed splits.
<dataset>	common	Each test image with one of 11 common corruptions at severity 3 (column corruption).
<dataset>	random_linf_30	Uniform random noise with $\lVert\delta\rVert_\infty \le 30/255$.
<dataset>-<surrogate>	<norm>_<eps>	Transfer adversarial images, e.g. config flowers-102-clip_vitb16, split linf_30.

The Parquet files themselves live under data/<dataset>/.

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. The test set is split into 11 contiguous blocks of ⌊n/11⌋ images, one per corruption in the order above, and the remaining images wrap around to the first corruption (column corruption). The images are fixed, 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 $\epsilon$	Splits
$\ell_1$	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).

Config suffix	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).

Surrogate	Split	Caltech-101	FGVC-Aircraft (2013b)	Oxford Flowers-102	Oxford-IIIT Pet	Stanford Cars	UC Merced Land Use
CLIP ViT-B/16	l1_75	✓	✓	✓	✓	✓	✓
CLIP ViT-B/16	l1_300	✓	✓	✓	✓	✓	✓
CLIP ViT-B/16	l2_2	✓	✓	✓	✓	✓	✓
CLIP ViT-B/16	l2_8	✓	✓	✓	✓	✓	✓
CLIP ViT-B/16	linf_4	✓	✓	✓	✓	✓	✓
CLIP ViT-B/16	linf_8	✓	✓	✓	✓	✓	✓
CLIP ViT-B/16	linf_30	✓	✓	✓	✓	✓	✓
CLIP ViT-H/14	l1_75	✓	✓	✓	✓	✓	✓
CLIP ViT-H/14	l1_300	✓	✓	✓	✓	✓	✓
CLIP ViT-H/14	l2_2	✓	✓	✓	✓	✓	✓
CLIP ViT-H/14	l2_8	✓	✓	✓	✓	✓	✓
CLIP ViT-H/14	linf_4	✓	✓	✓	✓	✓	✓
CLIP ViT-H/14	linf_8	✓	✓	✓	✓	✓	✓
CLIP ViT-H/14	linf_30	✓	✓	✓	✓	✓	✓
MetaCLIP ViT-H/14	l1_75	✓	✓	✓	✓	✓	✓
MetaCLIP ViT-H/14	l1_300	✓	✓	✓	✓	✓	✓
MetaCLIP ViT-H/14	l2_2	✓	✓	✓	✓	✓	✓
MetaCLIP ViT-H/14	l2_8	✓	✓	✓	✓	✓	✓
MetaCLIP ViT-H/14	linf_4	✓	✓	✓	✓	✓	✓
MetaCLIP ViT-H/14	linf_8	✓	✓	✓	✓	✓	✓
MetaCLIP ViT-H/14	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. 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	CC BY 4.0
FGVC-Aircraft (2013b)	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/	see upstream terms
Oxford-IIIT Pet	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	see upstream terms (research use)
UC Merced Land Use	http://weegee.vision.ucmerced.edu/datasets/landuse.html	USGS public-domain imagery

Citation

@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 = {The Fortieth Annual Conference on Neural Information Processing Systems Evaluations and Datasets Track},
  url       = {https://openreview.net/forum?id=KDsuVST8by},
  year      = {2026}
}

Please also cite the six source datasets.

Versions

• v2.1 (current): the common split contains the common-corruption images of the original processed archives.
• v2.0: 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.