Logit Margin Matters: Improving Transferable Targeted Adversarial Attack by Logit Calibration

Previous works have extensively studied the transferability of adversarial samples in untargeted black-box scenarios. However, it still remains challenging to craft targeted examples with higher than non-targeted ones. Recent studies reveal that traditional Cross-Entropy (CE) loss function is insufficient learn transferable due issue vanishing gradient. In this work, we provide a comprehensive investigation CE and find logit margin between classes will quickly obtain saturation CE, which largely limits transferability. Therefore, paper, devote goal continually increasing along optimization deal propose two simple effective calibration methods, are achieved by downscaling logits temperature factor an adaptive margin, respectively. Both them can effectively encourage produce larger lead Besides, show minimizing cosine distance classifier weights target class further improve transferability, benefited from via L2-normalization. Experiments conducted on ImageNet dataset validate effectiveness proposed outperform state-of-the-art methods attacks. The source code available at Link.