arXiv:2505.24369cs.LGcs.AI2025-05ACL被引 12

用对抗训练提升大模型对恶意指令的抵抗力,不依赖人工标注。

Adversarial Preference Learning for Robust LLM Alignment

  • 通过模型自身偏好概率判断有害性,省去人工评估。
  • 生成针对性攻击样本,自动发现并修复模型漏洞。
  • 在保持性能的同时,将有害输出降为0.43%,攻击成功率降低65%。

当前语言模型多依赖人类反馈强化学习(RLHF)以引导安全行为,但仍面临三大挑战:人工标注效率低且成本高、攻击方式多样、反馈偏见与奖励滥用风险。为此,本文提出对抗偏好学习(APL),一种迭代式对抗训练方法,包含三项创新:一是基于模型内在偏好概率的直接有害性度量,无需外部评估;二是条件生成攻击者,可生成针对特定输入的对抗性变体;三是自动化闭环反馈框架,实现漏洞发现与修复的持续迭代。在Mistral-7B-Instruct-v0.3上的实验表明,APL显著提升鲁棒性:相比基线模型,有害性胜率达83.33%(由GPT-4o评估),有害输出从5.88%降至0.43%(由LLaMA-Guard测量),攻击成功率最高下降65%(根据HarmBench)。值得注意的是,APL保持良好实用性,MT-Bench得分为6.59(基线为6.78),LC-WinRate为46.52%(优于基线)。

原文摘要 · Abstract (English)

Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation, (2) the vast diversity of potential adversarial attacks, and (3) the risk of feedback bias and reward hacking. To address these challenges, we introduce Adversarial Preference Learning (APL), an iterative adversarial training method incorporating three key innovations. First, a direct harmfulness metric based on the model's intrinsic preference probabilities, eliminating reliance on external assessment. Second, a conditional generative attacker that synthesizes input-specific adversarial variations. Third, an iterative framework with automated closed-loop feedback, enabling continuous adaptation through vulnerability discovery and mitigation. Experiments on Mistral-7B-Instruct-v0.3 demonstrate that APL significantly enhances robustness, achieving 83.33% harmlessness win rate over the base model (evaluated by GPT-4o), reducing harmful outputs from 5.88% to 0.43% (measured by LLaMA-Guard), and lowering attack success rate by up to 65% according to HarmBench. Notably, APL maintains competitive utility, with an MT-Bench score of 6.59 (comparable to the baseline 6.78) and an LC-WinRate of 46.52% against the base model.

大模型对齐对抗训练安全性

Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品;中文卡片由大模型生成,请以原文为准。