Scholarlink

<p>Fast Adversarial Training (FAT) enhances model robustness against adversarial attacks with minimal computational overhead. However, standard FAT often leads to catastrophic overfitting and low robustness due to the weak perturbation initialization. To address this, we propose a refined FAT framework, Precision-Aware Adversarial Initialization (PAAI), that stabilizes FAT by generating initialization priors that are high-loss and semantically meaningful. PAAI combines a multi-temporal perturbation buffer, semantic consistency based on the gradient alignment metric, and alignment-constrained FGSM refinement, which is used to reject degenerate updates. With adaptive scheduling of exploration and alignment weight, PAAI guarantees robust, single-step training without catastrophic overfitting. Experiments on CIFAR-10, CIFAR-100, and Tiny ImageNet show that PAAI is superior to baseline FAT methods with an accuracy of 50.31% against AutoAttack on CIFAR-10, where the improvement of PAAI over traditional fast adversarial training is 13.13%. Our approach achieves a balance between efficiency, robustness, and stability in training, which makes a new state-of-the-art for FAT methods. The source code is available at <a href="https://github.com/anjumiqbalse/PAAI">https://github.com/anjumiqbalse/PAAI</a>.</p>

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

Enhancing fast adversarial training: precision-aware initialization with historical perturbation guidance

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Anjum Iqbal,
  • Weiqiang Kong,
  • Yasir Iqbal

摘要

Fast Adversarial Training (FAT) enhances model robustness against adversarial attacks with minimal computational overhead. However, standard FAT often leads to catastrophic overfitting and low robustness due to the weak perturbation initialization. To address this, we propose a refined FAT framework, Precision-Aware Adversarial Initialization (PAAI), that stabilizes FAT by generating initialization priors that are high-loss and semantically meaningful. PAAI combines a multi-temporal perturbation buffer, semantic consistency based on the gradient alignment metric, and alignment-constrained FGSM refinement, which is used to reject degenerate updates. With adaptive scheduling of exploration and alignment weight, PAAI guarantees robust, single-step training without catastrophic overfitting. Experiments on CIFAR-10, CIFAR-100, and Tiny ImageNet show that PAAI is superior to baseline FAT methods with an accuracy of 50.31% against AutoAttack on CIFAR-10, where the improvement of PAAI over traditional fast adversarial training is 13.13%. Our approach achieves a balance between efficiency, robustness, and stability in training, which makes a new state-of-the-art for FAT methods. The source code is available at https://github.com/anjumiqbalse/PAAI.