nimaeb/xai-attack-detection-cifar10

XAI Attack Detection — CIFAR-10 PGD

This private research dataset contains balanced, paired clean and adversarial images for studying whether an attack can be detected from a classifier explanation map.

Dataset construction

The source is the CIFAR-10 test split. A fine-tuned OpenCLIP ViT-B/16 classifies each clean image. Clean-correct examples are attacked with untargeted L-infinity PGD using epsilon 8/255, step size 2/255, 10 steps, and deterministic random starts.

Clean/adversarial pairs are serialized as lossless PNG and reloaded. A pair is eligible only when the clean image remains correct, the attacked image remains misclassified, and the serialized perturbation remains inside the epsilon bound. The released corpus uses the same pair count for all ten classes and source-disjoint 70/15/15 splits.

Exact counts, software versions, model checksum, attack-success rates, and discarded counts are recorded in the accompanying manifest and generation report.

Fields

Each row contains a stable pair ID, CIFAR source index, clean/adversarial images, source class, clean/adversarial predictions and confidence, PGD parameters and seed, perturbation norms, model revision/checksum, code revision, and schema version.

Intended use and limitations

The dataset is intended for controlled adversarial-robustness and XAI research. It is not a general-purpose attack detector benchmark. It includes one source dataset, one classifier family, and one attack configuration, and it retains only successful post-serialization attacks. Balancing discards valid examples from higher-success classes.

License and access

This artifact is private and currently has no reuse license. Do not redistribute it. CIFAR-10 source terms and a derived-dataset license must be reviewed before any public release.