IC-449CLIP ViT-B/16 Grad-CAM explanations are highly sensitive to input noise, with SSIM dropping from 91.18% to 70.58% as noise standard deviation increases from 1/255 to 9/255

Shizhan Gong, Haoyu LEI, Qi Dou, Farzan Farnia

SourceBoosting the visual interpretability of CLIP via adversarial fine-tuning

The paper measures the stability of Grad-CAM saliency maps by injecting small random Gaussian noise into the input and computing the SSIM between the noisy and clean Grad-CAM maps. For the original CLIP ViT-B/16, even a tiny noise level of 1/255 reduces SSIM to 91.18%, and at 9/255 it drops to 70.58%. In contrast, the AFT-fine-tuned CLIP maintains SSIM above 98.55% across all noise levels. The authors attribute the original CLIP's fragility to the non-smoothness in its decision boundary introduced by false negatives during contrastive training.

Evidence
correlational
Key metric
SSIM: CLIP 91.18 / 82.98 / 77.56 / 73.62 / 70.58 (sigma = 1/255 to 9/255); CLIP w/ AFT 99.99 / 99.96 / 99.79 / 99.37 / 98.55
Caveat
The test uses only Grad-CAM (not simple gradients, which the authors note are 'extremely noisy' in the original CLIP and thus excluded from this analysis).
Model
CLIP / CLIP-ViT (LC)
Concepts
Failure mode, Explanation faithfulness
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval]
Methods
Grad-CAM [primary], SSIM [eval]
Related findings
IC-447, IC-448, IC-450
Extraction
automatic-extraction