Contrastive Spectral Rectification: Test-Time Defense towards Zero-shot Adversarial Robustness of CLIP

Vision-language models (VLMs) such as CLIP have demonstrated remarkable zero-shot generalization, yet remain highly vulnerable to adversarial examples (AEs). While test-time defenses are promising, existing methods fail to provide sufficient robustness against strong attacks and are often hampered by high inference latency and task-specific applicability. To address these limitations, we start by investigating the intrinsic properties of AEs, which reveals that AEs exhibit severe feature inconsistency under progressive frequency attenuation. We further attribute this to the model's inherent spectral bias. Leveraging this insight, we propose an efficient test-time defense named Contrastive Spectral Rectification (CSR). CSR optimizes a rectification perturbation to realign the input with the natural manifold under a spectral-guided contrastive objective, which is applied input-adaptively. Extensive experiments across 16 classification benchmarks demonstrate that CSR outperforms the SOTA by an average of 18.1% against strong APGD with modest inference overhead. Furthermore, CSR exhibits broad applicability across diverse visual tasks. Code is available at https://github.com/Summu77/CSR.

(1) Modern AI tools all share a few "eyes" — vision models like CLIP — and attackers can sprinkle invisible noise on an image to silently fool every tool that uses them. (2) We noticed this noise hides in an image's fine details, while its true meaning lives in the blurry outline; so our method first checks whether the model still recognises a blurred copy, and, if not, gently pulls the suspicious image back toward that trustworthy blurry version in a few quick steps. (3) The fix takes milliseconds, needs no retraining, and restores accuracy across image recognition, captioning, and visual question answering.