README.md

metadata

viewer: false
tags:
  - adversarial-robustness
  - image-classification
  - robustness-benchmark

RobustGenBench

A benchmark for evaluating the adversarial robustness of zero-shot image classifiers across six fine-grained / domain-specific datasets and a range of threat models.

A small stratified sample (2 images per class) is also available for quick inspection: 👉 https://huggingface.co/datasets/legolasflagstaff/RobustGenBench-sample

Structure

caltech101_processed.tar.zst
fgvc-aircraft-2013b_processed.tar.zst
flowers-102_processed.tar.zst
oxford-iiit-pet_processed.tar.zst
stanford_cars_processed.tar.zst
uc-merced-land-use-dataset_processed.tar.zst

class_names/
  <dataset>.json                 ← integer-label → class-name mappings

adversarial/
  common/common_severity3/<dataset>__common_severity3_processed.tar.zst
  random/linf_eps30_random_uniform/<dataset>__random_linf_eps30_random_uniform_processed.tar.zst
  zeroshot_clip_vitb16_laion2b/<threat_model>/<dataset>__<threat_model>_processed.tar.zst
  zeroshot_clip_vith14_laion2b/<threat_model>/<dataset>__<threat_model>_processed.tar.zst
  zeroshot_metaclip_vith14_fullcc2_5b/<threat_model>/<dataset>__<threat_model>_processed.tar.zst
  zeroshot_siglip2_base_patch16_224/<threat_model>/<dataset>__<threat_model>_processed.tar.zst
  zeroshot_siglip2_so400m_patch14_384/<threat_model>/<dataset>__<threat_model>_processed.tar.zst
  zeroshot_siglip2_so400m_patch16_naflex/<threat_model>/<dataset>__<threat_model>_processed.tar.zst
  zeroshot_siglip2_so400m_patch16_naflex_patchify/<threat_model>/<dataset>__<threat_model>_processed.tar.zst

Each archive is a .tar.zst containing flat-numbered PNGs and a labels.csv.

Clean archives

<dataset>_processed.tar.zst
├── train/
│   ├── 00000.png
│   ├── 00001.png
│   ├── ...
│   └── labels.csv
├── val/
│   ├── 00000.png
│   ├── ...
│   └── labels.csv
├── test/
│   ├── 00000.png
│   ├── ...
│   └── labels.csv
└── metadata.json                ← split counts and number of classes N

Adversarial archives

<dataset>__<threat_model>_processed.tar.zst
└── test/
    ├── 00000.png
    ├── ...
    └── labels.csv

Filenames are aligned across all archives: test/00000.png in every adversarial archive corresponds to the same source image (and same label) used to generate the perturbation. labels.csv provides the filename,label mapping with integer class indices; resolve to class names via class_names/<dataset>.json.

Datasets

Dataset	Classes	Test size
Caltech101	101	1000
FGVC-Aircraft 2013b	100	1000
Oxford Flowers 102	102	1000
Oxford-IIIT Pet	37	1000
Stanford Cars	196	1000
UC Merced Land Use	21	420

Threat models

The adversarial/ tree is organized by surrogate model used to craft the attack, then by threat model.

Untargeted attacks (AutoAttack standard, per surrogate):

• linf_eps8_autoattack_standard — L∞, ε = 8/255
• linf_eps30_autoattack_standard — L∞, ε = 30/255
• l2_eps2_autoattack_standard — L2, ε = 2
• l2_eps8_autoattack_standard — L2, ε = 8
• l1_eps75_autoattack_standard — L1, ε = 75
• l1_eps300_autoattack_standard — L1, ε = 300

Surrogate-agnostic baselines:

• common/common_severity3 — common corruption suite at severity 3
• random/linf_eps30_random_uniform — random uniform L∞ noise at ε = 30/255

Loading

import tarfile, io, csv
import zstandard as zstd
from PIL import Image
from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="legolasflagstaff/RobustGenBench",
    repo_type="dataset",
    filename="uc-merced-land-use-dataset_processed.tar.zst",
)

with open(archive, "rb") as f:
    buf = io.BytesIO(zstd.ZstdDecompressor().stream_reader(f).read())
with tarfile.open(fileobj=buf, mode="r:") as tar:
    images = {}
    for m in tar.getmembers():
        if m.name.startswith("test/") and m.name.endswith(".png"):
            images[m.name] = Image.open(io.BytesIO(tar.extractfile(m).read())).convert("RGB")
    labels_f = tar.extractfile(tar.getmember("test/labels.csv"))
    labels = list(csv.DictReader(io.TextIOWrapper(labels_f)))

print(f"Loaded {len(images)} images, {len(labels)} label rows")

Citation

If you use RobustGenBench in your work, please cite:

@inproceedings{robustgenbench2025,
  title  = {RobustGenBench: ...},
  author = {...},
  year   = {2025},
}

License

[Specify license here — e.g. CC-BY-4.0, or per-dataset license inheritance.]