AI Paper Atlas

Research map / Adversarial robustness and security / Certified and empirical robustness

You Don't Protect if You Don't Expect: Breaking the Key Assumption behind CLIP's Test-Time Defenses

Ruize Zhang, Yu Li, Zhang Wan, Juan Cao, Jie Zhang, Sheng Tang

ICML 2026 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training with feature consistency

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add CLIP-MAD on top of full-gradient attackTTC defense · PGD-10 eps=4/255, 13 datasetsaverage robust accuracy (lower means stronger attack): 11.0 -> 4.7Improved
add CLIP-MAD on top of full-gradient attackDOC defense · PGD-10 eps=4/255, 13 datasetsaverage robust accuracy (lower means stronger attack): 19.5 -> 9.4Improved
add CLIP-MAD on top of static attackR-TPT defense · PGD-10 eps=4/255, 13 datasetsaverage robust accuracy (lower means stronger attack): 37.5 -> 0.7Improved
add CLIP-MAD on top of full-gradient attackTAPT-VLI defense · PGD-10 eps=4/255, 13 datasetsaverage robust accuracy (lower means stronger attack): 0.3 vs 0.4No clear effect
use CLIP-MAD instead of static attackCLIPure-Cos defense · PGD-10 eps=1/255, 13 datasetsaverage robust accuracy (lower means stronger attack): 6.6 -> 4.8Improved
use CLIP-MAD instead of static attackCLIPure-Diff defense · PGD-10 eps=1/255, 13 datasetsaverage robust accuracy (lower means stronger attack): 7.3 -> 6.0Improved
use full-gradient attack instead of static attackCLIPure-Cos defense · PGD-10 eps=1/255, 13 datasetsaverage robust accuracy (lower means stronger attack): 6.6 -> 78.0Hurt
use CLIP-MAD instead of static attackR-TPT defense · APGD restart=2 eps=4/255, ImageNetaverage robust accuracy (lower means stronger attack): 37.6 -> 0.6Improved
use CLIP-MAD instead of full-gradient attackTTC defense · PGD restart=3 EOT-10 eps=4/255, CIFAR10average robust accuracy (lower means stronger attack): 12.0 -> 3.9Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy