AI RESEARCH
AGC: Adaptive Geodesic Correction for Adversarial Robustness on Vision-Language Models
arXiv CS.CV
•
ArXi:2605.15584v1 Announce Type: new Vision-language models like CLIP have nstrated remarkable zero-shot transfer capabilities. However, their susceptibility to imperceptible adversarial perturbations remains a critical security concern. While test-time defenses offer a pragmatic solution for deployed models, existing approaches typically rely on gradient-based optimization during inference, incurring significant computational overhead.