Scholarlink

Recently, large-scale transformer-based models have achieved remarkable breakthroughs in computer vision and natural language processing. However, Vision Transformers remain highly vulnerable to adversarial examples, where minor imperceptible perturbations can lead to high-confidence misclassifications. To systematically investigate these vulnerabilities and develop more effective security evaluation methodologies, this paper introduces a novel spatial-domain adversarial attack called Diversified Block-wise Data Transform and Attention Fusion (DBT-AF). DBT-AF employs diversified block-level data transformations and bidirectional cross-layer attention fusion to disrupt local overfitting while integrating shallow details with deep semantic cues, thereby enhancing attack transferability across different model architectures. Extensive evaluations on the ILSVRC 2012 dataset demonstrate that DBT-AF achieves average attack success rates of 94.05% on eight ViT models, 85.15% on four CNN models, and 65.77% on four adversarially trained CNN models, all while maintaining high visual imperceptibility. These findings provide critical insights into the security vulnerabilities of transformer based architectures and contribute to the development of more robust adversarial defense mechanisms.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Boosting Transferability of Adversarial Attacks On Vision Transformer

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shengyan Huo,
  • Zhonghua Yang,
  • Xixiang Lyu,
  • Jing Ma

摘要

Recently, large-scale transformer-based models have achieved remarkable breakthroughs in computer vision and natural language processing. However, Vision Transformers remain highly vulnerable to adversarial examples, where minor imperceptible perturbations can lead to high-confidence misclassifications. To systematically investigate these vulnerabilities and develop more effective security evaluation methodologies, this paper introduces a novel spatial-domain adversarial attack called Diversified Block-wise Data Transform and Attention Fusion (DBT-AF). DBT-AF employs diversified block-level data transformations and bidirectional cross-layer attention fusion to disrupt local overfitting while integrating shallow details with deep semantic cues, thereby enhancing attack transferability across different model architectures. Extensive evaluations on the ILSVRC 2012 dataset demonstrate that DBT-AF achieves average attack success rates of 94.05% on eight ViT models, 85.15% on four CNN models, and 65.77% on four adversarially trained CNN models, all while maintaining high visual imperceptibility. These findings provide critical insights into the security vulnerabilities of transformer based architectures and contribute to the development of more robust adversarial defense mechanisms.