IC-448CLIP ViT-L/14 achieves 0% accuracy under 2/255 and 4/255 L-infinity adversarial perturbations across all 15 evaluation datasets

Shizhan Gong, Haoyu LEI, Qi Dou, Farzan Farnia

SourceBoosting the visual interpretability of CLIP via adversarial fine-tuning

The paper evaluates the adversarial robustness of the original CLIP ViT-L/14 using the first two attacks of AutoAttack (APGD with cross-entropy and APGD with targeted DLR loss). Under both 2/255 and 4/255 L-infinity perturbation budgets, the original CLIP achieves exactly 0.00% accuracy on every one of the 15 zero-shot classification datasets tested, including ImageNet, CIFAR-10, STL-10, and ImageNet-R. This complete vulnerability contrasts sharply with the AFT-fine-tuned variants, which retain 44-46% average accuracy under 2/255 perturbations.

Evidence
correlational
Key metric
0.00% accuracy on all 15 datasets (ImageNet, CIFAR-10, STL-10, CIFAR-100, Cars, Caltech, OxfordPets, Flowers, DTD, EuroSAT, FGVC, PCAM, ImageNet-R, ImageNet-S) under both 2/255 and 4/255 L-infinity AutoAttack
Caveat
The evaluation uses only 1000 samples per dataset for robustness testing, and the attacks are limited to the first two AutoAttack components.
Model
CLIP / CLIP-ViT (LC)
Concepts
Failure mode
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval], CIFAR-10 [eval], STL-10 [eval], CIFAR-100 [eval], ImageNet-R / ImageNet-Rendition [eval], ImageNet-Sketch [eval]
Methods
AutoAttack [eval]
Related work
FARE [compared-to]
Related findings
IC-447, IC-449, IC-450
Extraction
automatic-extraction