AMGAA: Attention-Guided Multi-Target Generative Adversarial Attack for Vision Transformers

Vision Transformers (ViTs) have achieved strong performance in computer vision, but their adversarial robustness remains underexplored. Existing ViT-oriented attacks mainly rely on iterative optimization, leading to high generation cost and limited transferability. Moreover, most generative attacks target a single class, making them inefficient for multi-target scenarios. To address these issues, we propose Attention-Guided Multi-Target Generative Adversarial Attack (AMGAA). AMGAA leverages ViT self-attention to guide target feature fusion and adaptively selects important source patches for perturbation generation. It jointly optimizes adversarial, attention constraint, and total variation losses to improve targeted attack success while preserving visual naturalness. Experiments on CIFAR-10 and ImageNet show that AMGAA achieves average attack success rates (ASRs) of 43.2% and 39.0% in ImageNet single-target transfer attacks and CIFAR-10 multi-target attacks, respectively. Compared with the generative attack methods evaluated in our experiments, AMGAA improves ASR by 5.7 percentage points in the multi-target setting and by 8.1 percentage points in unknown-class generalization. AMGAA also obtains a low LPIPS of 0.018, indicating good visual imperceptibility. Ablation studies confirm the effectiveness of its key components.