Boosting the Local Invariance for Better Adversarial Transferability
Adversarial perturbations often fail to transfer well between different models because they lack local invariance. This research introduces Local Invariance Boosting (LI-Boost), a method that optimizes attacks by considering perturbations across translated versions of the image. By enhancing local invariance, LI-Boost significantly improves the success rate of adversarial attacks across various architectures, including CNNs, ViTs, and Vision-Language Models.