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

Area of Science:

• Computer Vision
• Deep Learning Security
• Adversarial Machine Learning

Background:

• Vision Transformers (ViTs) excel in computer vision but lack robust adversarial defense.
• Current ViT attacks are costly, have limited transferability, and are inefficient for multi-target scenarios.

Purpose of the Study:

• To develop an efficient and effective adversarial attack method for Vision Transformers.
• To address the limitations of existing single-target and iterative attacks.

Main Methods:

• Propose Attention-Guided Multi-Target Generative Adversarial Attack (AMGAA).
• Leverage ViT self-attention for feature fusion and adaptive perturbation generation.
• Employ joint optimization of adversarial, attention constraint, and total variation losses.

Main Results:

• AMGAA achieves 43.2% ASR on ImageNet single-target transfer attacks and 39.0% on CIFAR-10 multi-target attacks.
• Improves ASR by 5.7% in multi-target and 8.1% in unknown-class generalization compared to existing methods.
• Achieves low visual imperceptibility with an LPIPS score of 0.018.

Conclusions:

• AMGAA offers a more effective and efficient adversarial attack for Vision Transformers.
• The method demonstrates strong performance in both single- and multi-target attack settings.
• Key components of AMGAA are validated through ablation studies, confirming their effectiveness.