Scholarlink

<p>Large language models (LLMs) are widely deployed in intelligent systems but remain vulnerable to adversarial attacks, such as Greedy coordinate gradient (GCG) and Projected gradient descent (PGD). Existing approaches suffer from three critical issues: (1) Gradient methods like PGD use static entropy factors, failing to adapt to the dynamic gap between continuous optimization and discrete evaluation, leading to low success rates; (2) Discrete methods like GCG rely on trial-and-error with 512 candidate tokens, incurring high computational costs; (3) Existing RL attacks optimize only the objective function, lacking component collaboration, which hinders breakthroughs against complex defenses. To address these, we propose Adaptive reinforcement learning-based projected gradient descent (ARL-PGD). It features three logically sequential innovations: (1) A reinforcement learning-based system-level framework that provides discrete evaluations as the system foundation; (2) A distributed discrete loss feedback mechanism to align continuous optimization with discrete objectives, mitigating GCG’s costs and PGD’s feedback gaps; (3)A dynamic entropy factor strategy: adapting entropy via relaxation gaps, it deterministically modulates distribution sharpness (distinct from noise uncertainty) to preserve optimization gains. These components form a closed-loop “evaluate-feedback-adjust” system, enabling nonlinear synergistic optimization and significantly improving attack success rates (ASR). Experiments on mainstream LLM models (Vicuna, Llama, and Gemma series) show ARL-PGD achieves higher ASR than baselines, with more natural and stealthy adversarial prompts. Ablation studies confirm each component’s effectiveness.</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

Adaptive reinforcement learning based projected gradient descent attack

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Zihan Zhu,
  • Yuexin Zhang,
  • Ayong Ye,
  • Xiaoding Wang,
  • Chengling Wang,
  • Tianqing Zhu

摘要

Large language models (LLMs) are widely deployed in intelligent systems but remain vulnerable to adversarial attacks, such as Greedy coordinate gradient (GCG) and Projected gradient descent (PGD). Existing approaches suffer from three critical issues: (1) Gradient methods like PGD use static entropy factors, failing to adapt to the dynamic gap between continuous optimization and discrete evaluation, leading to low success rates; (2) Discrete methods like GCG rely on trial-and-error with 512 candidate tokens, incurring high computational costs; (3) Existing RL attacks optimize only the objective function, lacking component collaboration, which hinders breakthroughs against complex defenses. To address these, we propose Adaptive reinforcement learning-based projected gradient descent (ARL-PGD). It features three logically sequential innovations: (1) A reinforcement learning-based system-level framework that provides discrete evaluations as the system foundation; (2) A distributed discrete loss feedback mechanism to align continuous optimization with discrete objectives, mitigating GCG’s costs and PGD’s feedback gaps; (3)A dynamic entropy factor strategy: adapting entropy via relaxation gaps, it deterministically modulates distribution sharpness (distinct from noise uncertainty) to preserve optimization gains. These components form a closed-loop “evaluate-feedback-adjust” system, enabling nonlinear synergistic optimization and significantly improving attack success rates (ASR). Experiments on mainstream LLM models (Vicuna, Llama, and Gemma series) show ARL-PGD achieves higher ASR than baselines, with more natural and stealthy adversarial prompts. Ablation studies confirm each component’s effectiveness.