Scholarlink

In recent years, deep neural networks (DNNs) have made significant advancements in various vision tasks, including image classification and object detection. However, these models remain vulnerable to adversarial examples subtle perturbations that can significantly mislead predictions. In response to this, we propose a novel adversarial attack method called Local Masking and Multi-stage Momentum Optimization (LMMMO). LMMMO integrates a self-adaptive local masking mechanism to focus perturbations on decision-critical regions, improving both attack efficiency and perceptual stealth. Additionally, LMMMO uses a multi-stage momentum optimization strategy, adjusting step sizes dynamically and refining gradients through a coarse-to-fine approach to ensure stable convergence. Compared to existing methods, LMMMO achieves competitive high attack success rates (e.g., 0.989 average ASR) while demonstrating the lowest average L2 distortion (0.234). It also shows favorable perceptual similarity (LPIPS) compared to several iterative baselines like PGD and MIFGSM, offering a promising balance between attack performance and imperceptibility.

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

An Adversarial Attack Method Based on Local Masking and Multi-stage Momentum Optimization

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Liqiang Lin,
  • Yu Lin,
  • Yunyu Kang,
  • Shuwu Chen,
  • Xiaolong Liu

摘要

In recent years, deep neural networks (DNNs) have made significant advancements in various vision tasks, including image classification and object detection. However, these models remain vulnerable to adversarial examples subtle perturbations that can significantly mislead predictions. In response to this, we propose a novel adversarial attack method called Local Masking and Multi-stage Momentum Optimization (LMMMO). LMMMO integrates a self-adaptive local masking mechanism to focus perturbations on decision-critical regions, improving both attack efficiency and perceptual stealth. Additionally, LMMMO uses a multi-stage momentum optimization strategy, adjusting step sizes dynamically and refining gradients through a coarse-to-fine approach to ensure stable convergence. Compared to existing methods, LMMMO achieves competitive high attack success rates (e.g., 0.989 average ASR) while demonstrating the lowest average L2 distortion (0.234). It also shows favorable perceptual similarity (LPIPS) compared to several iterative baselines like PGD and MIFGSM, offering a promising balance between attack performance and imperceptibility.