Skip to main content

Toward Interpreting and Boosting Adversarial Transferability via Transferability Error Minimization

By
Yunfei Long; Zilin Tian; Liguo Zhang; Ye Tian; Jianguo Sun; Huosheng Xu

Deep neural networks (DNNs) have been demonstrated to be vulnerable to meticulously crafted adversarial examples. Due to the transferability, adversarial examples can attack across different models and help expose weaknesses of DNNs. Despite existing transfer-based adversarial attacks having achieved great potential, they rely on empirical observations and lack theoretical interpretability. To address this limitation, we introduce Transferability Error, defined as the gap in prediction-change risk between an adversarial example and the optimal example in the adversarial region, and derive an upper bound that links this error to the maximum prediction discrepancy of the surrogate model. Motivated by this bound, we propose a novel attack method, Transferability Error Minimization (TEM), which integrates the standard loss with a differentiable proxy that reduces the loss discrepancy between local maxima and minima of the surrogate model. In addition to directly boosts the adversarial transferability, this regularizer implicitly flattens the loss landscape and avoids gradient conflicts with the standard loss. Extensive experiments on the ImageNet-compatible and CIFAR-10 datasets demonstrate that TEM achieves state-of-the-art transferability, significantly outperforming various baselines by up to 5% in average attack success rates against diverse normally trained models, adversarially trained models, as well as against several defense mechanisms. Source code will be released at https://github.com/unclelongheu/Transferability-Error

Read on IEEE Xplore