| 引用本文： |
-
丁汝妍,刘玉玲,王颖,廖鑫.基于精准扰动的鲁棒性跨域AIGT检测方法[J].信息安全学报,已采用 [点击复制]
- Ding Ruyan,Liu Yuling,Wang Ying,Liao Xin.Robust Cross-Domain AIGT Detection Method Based on Precise Perturbation[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 随着大型语言模型生成文本的逼真度持续提升，人工智能生成文本（AI-Generated Text, AIGT）与人类文本的辨识难度显著增加，知识产权侵权、虚假信息扩散等问题激增，亟需研发可靠的AIGT检测工具。针对现有检测方法面临跨域泛化性与鲁棒性难以平衡的缺陷，提出一种融合语义感知与可解释性引导的动态扰动网络，从两个维度优化AIGT检测。在内部结构层面，通过引入自注意力机制生成语义感知扰动掩码，自动识别并保护文本关键判别特征，避免噪声对语义空间的无差别破坏；在外部知识引导层面，以思维链（Chain of Thought, CoT）推理挖掘的可解释性特征为扰动靶点，将关键特征转化为强化学习动作空间约束，构建CoT感知奖励函数。实验以M4多领域数据集为基准，构建跨生成器与语料库的不同域检测场景。结果表明，在6个未见目标域上，注意力增强型动态扰动方法（RL_Att）与思维链引导型动态扰动方法（RL_CoT）平均检测F1分别可达87.50 %与88.02 %，较DP-Net方法提高4.54 %与5.06 %；为验证方法对不同辅助大语言模型的适应性，额外引入Deepseek-chat模型提取CoT特征，其对应的RL_CoT变体平均F1进一步提升至88.84 %；当面对同义词替换的对抗攻击时，RL_Att与RL_CoT的鲁棒性最优，F1分别为85.30 %与84.25 %，较DP-Net方法分别提高5.39 %与4.34 %，而RL_CoT_Deepseek模型的平均F1更达到85.95 %；面对改写攻击时，RL_Att与RL_CoT的鲁棒性最优，F1分别为69.81 %与62.11 %，较DP-Net方法分别提高18.26 %与10.56 %，其中RL_CoT_Deepseek模型的平均F1提升至73.31 %，较DP-Net提高21.76个百分点。 |
| 关键词: 人工智能生成文本 跨域人工智能生成文本检测 思维链推理 强化学习 注意力机制 |
| DOI： |
| 投稿时间：2026-02-08修订日期：2026-07-13 |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目）,国家重点基础研究发展计划（973计划） |
|
| Robust Cross-Domain AIGT Detection Method Based on Precise Perturbation |
|
Ding Ruyan, Liu Yuling, Wang Ying, Liao Xin
|
| (Hunan University) |
| Abstract: |
| As the fidelity of Large Language Model outputs continues to improve, the increasing difficulty in distinguishing AI-Generated Text (AIGT) from human-authored content has intensified risks like intellectual property infringement and misinformation, necessitating reliable detection tools. Addressing the trade-off between cross-domain generalization and robustness in current methods, this paper proposes a dynamic perturbation network integrating semantic awareness and interpretability guidance. At the structural level, a self-attention mechanism generates semantic-aware masks to protect key discriminative features from indiscriminate noise. At the knowledge-guidance level, interpretable features extracted via Chain-of-Thought (CoT) reasoning serve as perturbation targets, transforming key attributes into Reinforcement Learning (RL) action space constraints to build a CoT-aware reward function. Experiments on the M4 multi-domain dataset across cross-generator and cross-corpus scenarios show that on six unseen domains, the attention-enhanced (RL_Att) and CoT-guided (RL_CoT) methods achieved average F1-scores of 87.50% and 88.02%, outperforming DP-Net by 4.54% and 5.06%, respectively. To verify adaptability, incorporating Deepseek-chat for CoT extraction further boosted the RL_CoT F1 to 88.84%. Under synonym-substitution attacks, RL_Att and RL_CoT demonstrated superior robustness with F1-scores of 85.30% and 84.25%, while the RL_CoT_Deepseek variant reached 85.95%. Against paraphrasing attacks, RL_Att and RL_CoT achieved F1-scores of 69.81% and 62.11%, surpassing DP-Net by 18.26% and 10.56%, with the RL_CoT_Deepseek model reaching 73.31%, a 21.76 percentage point improvement over the baseline. |
| Key words: artificial intelligence-generated text, cross-domain artificial intelligence-generated text detection, chain of thought reasoning, reinforcement learning, attention mechanism |
