Scholarlink

<p>Deep neural networks (DNNs) have achieved remarkable success across a wide range of vision tasks, yet they remain highly vulnerable to adversarial attacks. Moreover, adversarial perturbations often exhibit strong cross-model transferability in black-box settings. However, existing transfer-based attacks face a fundamental trade-off: achieving high attack success rates typically compromises perceptual imperceptibility, whereas visually stealthy perturbations often suffer from degraded transferability. To address this challenge, we propose TRIAD-Attack, a diffusion-based adversarial framework that jointly enhances transferability and imperceptibility. TRIAD-Attack integrates complementary mechanisms at both the optimization and spatial levels. Specifically, we aggregate gradients from an ensemble of surrogate models to alleviate single-surrogate bias and introduce a rescaled variance-reduced gradient (r-SVRG) strategy to mitigate inter-model gradient inconsistency, thereby stabilizing the optimization trajectory and improving cross-model transferability. In addition, we incorporate a multi-granularity semantic guidance mechanism into the reverse diffusion process, which adaptively suppresses perturbations in semantically insignificant regions at both block and pixel levels. This semantic-aware control effectively preserves visual quality without sacrificing attack effectiveness. Extensive experiments on CNNs, ViTs, and VisionMamba validate the superiority of the proposed approach. On average, TRIAD-Attack achieves 2.7% and 3.4% higher attack success rates on CNNs and ViTs, respectively, compared to state-of-the-art methods, while consistently yielding better performance across multiple image quality metrics. Code is available at <a href="https://github.com/AdvML-Group/TRIAD-Attack">https://github.com/AdvML-Group/TRIAD-Attack</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 the transferability and imperceptibility of adversarial attacks via rescaled variance-reduced diffusion

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Junwei Xie,
  • Wenwen Zhang,
  • Wen Yang,
  • Guodong Liu,
  • Di Ming

摘要

Deep neural networks (DNNs) have achieved remarkable success across a wide range of vision tasks, yet they remain highly vulnerable to adversarial attacks. Moreover, adversarial perturbations often exhibit strong cross-model transferability in black-box settings. However, existing transfer-based attacks face a fundamental trade-off: achieving high attack success rates typically compromises perceptual imperceptibility, whereas visually stealthy perturbations often suffer from degraded transferability. To address this challenge, we propose TRIAD-Attack, a diffusion-based adversarial framework that jointly enhances transferability and imperceptibility. TRIAD-Attack integrates complementary mechanisms at both the optimization and spatial levels. Specifically, we aggregate gradients from an ensemble of surrogate models to alleviate single-surrogate bias and introduce a rescaled variance-reduced gradient (r-SVRG) strategy to mitigate inter-model gradient inconsistency, thereby stabilizing the optimization trajectory and improving cross-model transferability. In addition, we incorporate a multi-granularity semantic guidance mechanism into the reverse diffusion process, which adaptively suppresses perturbations in semantically insignificant regions at both block and pixel levels. This semantic-aware control effectively preserves visual quality without sacrificing attack effectiveness. Extensive experiments on CNNs, ViTs, and VisionMamba validate the superiority of the proposed approach. On average, TRIAD-Attack achieves 2.7% and 3.4% higher attack success rates on CNNs and ViTs, respectively, compared to state-of-the-art methods, while consistently yielding better performance across multiple image quality metrics. Code is available at https://github.com/AdvML-Group/TRIAD-Attack.