Scholarlink

To address the issue of weak transferability of adversarial attacks based on a single surrogate model, this paper proposes a Multi-Model Output Fusion-based Average Gradient Mapping adversarial attack method (MMOF-AGM). By weighting and combining the outputs of multiple models, it alleviates the issue of poor generalization that arises from over-reliance on the gradient of a single model. The average gradient mapping strategy preserves the relative magnitudes of gradients both before and after mapping, ensuring consistent updates. Experimental results demonstrate that the proposed approach outperforms traditional gradient-based attacks in both white-box and black-box settings, improving the effectiveness of adversarial examples across different radar signal modulation classification models.

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

Multi-model Output Fusion-Based Average Gradient Mapping Adversarial Attack

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sicheng Zhang,
  • Mengchao Wang,
  • Zhida Bao,
  • Yandie Yang,
  • Qiao Tian

摘要

To address the issue of weak transferability of adversarial attacks based on a single surrogate model, this paper proposes a Multi-Model Output Fusion-based Average Gradient Mapping adversarial attack method (MMOF-AGM). By weighting and combining the outputs of multiple models, it alleviates the issue of poor generalization that arises from over-reliance on the gradient of a single model. The average gradient mapping strategy preserves the relative magnitudes of gradients both before and after mapping, ensuring consistent updates. Experimental results demonstrate that the proposed approach outperforms traditional gradient-based attacks in both white-box and black-box settings, improving the effectiveness of adversarial examples across different radar signal modulation classification models.