IC-686Adversarial perturbations alter CLIP ViT-B/32's token representations most strongly starting around layer 10

Haozhe Chen, Junfeng Yang, Carl Vondrick, Chengzhi Mao

SourceINViTE: INterpret and Control Vision-Language Models with Text Explanations

In an appendix experiment, the paper applies an adversarial perturbation to a forest image that causes CLIP to misclassify it as beach. By interpreting each token at every layer with ViTE before and after the attack, they measure the percentage of tokens that shift from interpreting as 'forest' to 'beach.' The percentage of affected tokens increases sharply starting around layer 10, indicating the adversarial effect is concentrated in the later layers of the 13-layer transformer.

Evidence
correlational
Key metric
adversarial attack impacts model most starting around layer 10
Caveat
Single-image qualitative analysis in appendix; no aggregate statistics or multiple examples reported.
Model
CLIP / CLIP-ViT (LC)
Concepts
Depth-dependent structure
Datasets
UC Merced Land Use [eval]
Related findings
IC-684, IC-685
Extraction
automatic-extraction