Research Papers 论文研究 3h ago Updated 1h ago 更新于 1小时前 49

DWT-Fusion: A Signal-Based Framework for Training-Free LLM-Generated Text Detection DWT-Fusion:一种基于信号的无需训练的LLM生成文本检测框架

DWT-Fusion introduces a training-free framework for detecting LLM-generated text using discrete wavelet analysis of token-level log-probability sequences. The framework leverages wavelet-based multiresolution signal representations to capture localized probability dynamics, improving detection performance. Four voting variants (equal-weight hard/soft, calibration-weighted hard/soft) are evaluated to combine wavelet configurations without supervised training. Best single wavelet configurations ac 提出DWT-Fusion框架,利用离散小波变换分析token级log-probability序列,实现无需训练的LLM生成文本检测。 通过多分辨率信号表示捕捉局部和跨尺度的概率动态,弥补现有方法仅依赖全局统计量的不足。 采用四种无监督投票策略(等权硬/软投票、校准加权硬/软投票)融合多个小波配置,无需训练元分类器。 在HC3、M4和MAGE数据集上,结合校准加权投票的集成模型分别达到0.9919、0.8477和0.7471的AUROC。 使用GPT-Neo-2.7B、GPT-J-6B、Falcon-7B和LLaMA-3-8B作为代理模型验证泛化能力,显示跨模型与数据集的有效性。

75
Hot 热度
70
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • DWT-Fusion introduces a training-free framework for detecting LLM-generated text using discrete wavelet analysis of token-level log-probability sequences.
  • The framework leverages wavelet-based multiresolution signal representations to capture localized probability dynamics, improving detection performance.
  • Four voting variants (equal-weight hard/soft, calibration-weighted hard/soft) are evaluated to combine wavelet configurations without supervised training.
  • Best single wavelet configurations achieve AUROC of 0.9872 (HC3), 0.8185 (M4), and 0.7138 (MAGE), with calibration-weighted voting further improving AUROC to 0.9919, 0.8477, and 0.7471, respectively.
  • The method demonstrates effective and interpretable signals for training-free LLM-generated text detection across diverse datasets and proxy models.

Why It Matters

This work addresses a critical challenge in AI: detecting LLM-generated text without relying on training data, which is essential for generalization across unseen generators and domains. By exploiting local and multiscale variations in token-level predictability via wavelet analysis, DWT-Fusion offers a novel, interpretable alternative to global statistical measures, advancing the state-of-the-art in zero-shot detection. The framework’s strong performance on multiple datasets and its training-free nature make it highly relevant for practitioners seeking robust, deployable solutions in dynamic AI environments.

Technical Details

  • Core Method: DWT-Fusion uses discrete wavelet transform (DWT) to analyze token-level log-probability sequences generated by a proxy causal language model (e.g., GPT-Neo-2.7B, GPT-J-6B, Falcon-7B, LLaMA-3-8B).
  • Multiresolution Analysis: Wavelet-based multiresolution representations capture localized probability dynamics, enabling detection of subtle patterns in LLM-generated text that global measures may miss.
  • Voting Variants: Four training-free voting strategies are evaluated: equal-weight hard voting, equal-weight soft voting, calibration-weighted hard voting, and calibration-weighted soft voting, which combine multiple wavelet configurations without supervised meta-classifier training.
  • Evaluation Datasets: HC3, M4, and MAGE are used to assess generalization across diverse text domains and generators.
  • Performance Metrics: AUROC is the primary metric, with best single wavelet configurations achieving 0.9872 (HC3), 0.8185 (M4), and 0.7138 (MAGE), and calibration-weighted voting improving these to 0.9919, 0.8477, and 0.7471, respectively.

Industry Insight

  • Generalization Potential: DWT-Fusion’s training-free approach reduces dependency on labeled data, making it adaptable to emerging LLMs and unseen text domains, a critical advantage for real-world deployment.
  • Interpretability: The use of wavelet analysis provides interpretable signals, enabling practitioners to understand detection rationale and refine models based on localized probability dynamics.
  • Scalability: The framework’s lightweight design (no supervised training) allows efficient integration into existing systems, supporting scalable detection solutions for content moderation, plagiarism detection, and AI safety applications.

TL;DR

  • 提出DWT-Fusion框架,利用离散小波变换分析token级log-probability序列,实现无需训练的LLM生成文本检测。
  • 通过多分辨率信号表示捕捉局部和跨尺度的概率动态,弥补现有方法仅依赖全局统计量的不足。
  • 采用四种无监督投票策略(等权硬/软投票、校准加权硬/软投票)融合多个小波配置,无需训练元分类器。
  • 在HC3、M4和MAGE数据集上,结合校准加权投票的集成模型分别达到0.9919、0.8477和0.7471的AUROC。
  • 使用GPT-Neo-2.7B、GPT-J-6B、Falcon-7B和LLaMA-3-8B作为代理模型验证泛化能力,显示跨模型与数据集的有效性。

为什么值得看

该工作为训练-free LLM检测提供了一种新颖的信号处理视角,通过小波分析挖掘token级局部模式,突破了传统基于全局统计量的局限,对构建可解释、零样本部署的检测系统具有重要参考价值。其投票融合策略为多模型协作检测提供了轻量级方案,适用于资源受限或隐私敏感场景。

技术解析

  • 核心方法:将文本视为时间序列,用代理语言模型计算每个token的log-probability,再输入离散小波变换(DWT)提取多尺度特征系数作为检测信号。
  • 信号处理流程:DWT分解产生近似系数(低频趋势)和细节系数(高频波动),后者被用作异常检测指标,因LLM生成文本在局部概率分布上可能呈现规律性平滑。
  • 投票机制:独立训练多个小波配置(不同小波基、分解层级),输出二元预测后通过四种策略融合——硬投票取多数票,软投票平均概率,校准版本则根据各配置在验证集上的AUROC动态加权。
  • 实验设置:代理模型固定为预训练因果LM(如LLaMA-3-8B),不参与微调;测试集覆盖不同领域(HC3=问答,M4=多任务,MAGE=创意写作),评估零样本泛化性。
  • 性能表现:单模型最高AUROC在HC3达0.9872,集成后提升至0.9919;M4和MAGE因数据噪声更大,绝对值较低但相对提升显著(+0.0292和+0.0333),证明融合策略有效。

行业启示

  • 检测系统应向“无训练+可解释”演进:DWT-Fusion证明无需额外标注数据即可构建高性能检测器,适合快速迭代和合规审计场景,建议企业优先采用此类信号驱动方案而非黑盒分类器。
  • 多模型融合是提升鲁棒性的关键路径:单一代理模型易受架构偏差影响,通过小波配置多样性+校准加权投票可显著降低过拟合风险,行业应建立标准化投票框架以适配不同生成器。
  • 需警惕对抗性攻击风险:当前方法依赖token概率平滑性,未来可能遭遇针对小波特征的扰动攻击,建议结合语义一致性检查或动态代理模型切换机制增强防御纵深。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Security 安全 Evaluation 评测 Research 科学研究