Sen Nie

New defense method boosts CLIP's adversarial robustness

Researchers have developed a new test-time defense method called Contrastive Spectral Rectification (CSR) to improve the adversarial robustness of vision-language models like CLIP. This method addresses the vulnerabilit…