基于语言特征集成学习的大语言模型生成文本检测

doi:10.3969/j.issn.1671-1122.2024.07.011

摘要/Abstract

摘要：

大语言模型的快速发展为日常生活和工作提供了极大的便利，但也为个人和社会带来了挑战。因此，迫切需要能够检测大语言模型生成文本的检测器。为了兼具良好的检测性能和泛化能力，文章提出了一种基于语言特征集成学习的大语言模型生成文本检测方法EBF Detection。EBF Detection融合了微调预训练语言模型和高阶自然语言统计特征，利用判决机制，实现了大语言模型生成文本检测。实验结果显示，EBF Detection不仅在域内数据上平均的检测准确率达到了98.72%，而且在域外数据上的平均检测准确率达到了96.79%。

关键词: 大语言模型, 大语言模型生成文本检测, 集成学习, 语言特征

Abstract:

The rapid development of large language model (LLM) has provided great convenience for daily life and work, but has also brought challenges for individuals and society. Therefore, there is an urgent need for detectors that can detect text generated by large language models. For good detection performance and generalization ability, this paper proposed a large language model-generated text detection method based on linguistic feature learning—EBF detection. EBF detection combined the fine-tuned pre-trained language model and higher-order natural language statistical features, and used the decision mechanism to realize the LLM-generated text detection. Experimental results show that EBF Detection not only achieves an average detection accuracy of 98.72% on in-domain data, but also achieves an average detection accuracy of 96.79% on out-of-domain data.

Key words: large language model, LLM-generated text detection, ensemble learning, linguistic feature

中图分类号:

TP309

项慧, 薛鋆豪, 郝玲昕. 基于语言特征集成学习的大语言模型生成文本检测[J]. 信息网络安全, 2024, 24(7): 1098-1109.

XIANG Hui, XUE Yunhao, HAO Lingxin. Large Language Model-Generated Text Detection Based on Linguistic Feature Ensemble Learning[J]. Netinfo Security, 2024, 24(7): 1098-1109.

图/表 11

图1

图2

图3

图4

表1

表2

表3

表4

表5

表6

表7

参考文献 25

[1]	JI Ziwei, LEE N, FRIESKE R, et al. Survey of Hallucination in Natural Language Generation[J]. ACM Computing Surveys, 2023, 55(12): 1-38.
[2]	AZAMFIREI R, KUDCHADKAR S R, FACKLER J. Large Language Models and the Perils of Their Hallucinations[J]. Critical Care, 2023, 27(1): 120-132. doi: 10.1186/s13054-023-04393-x pmid: 36945051
[3]	FLANAGIN A, BIBBINS-DOMINGO K, BERKWITS M, et al. Nonhuman “Authors” and Implications for the Integrity of Scientific Publication and Medical Knowledge[J]. Jama, 2023, 329(8): 637-639.
[4]	KING M R, ChatGPT. A Conversation on Artificial Intelligence, Chatbots, and Plagiarism in Higher Education[J]. Cellular and Molecular Bioengineering, 2023, 16(1): 1-2. doi: 10.1007/s12195-022-00754-8 pmid: 36660590
[5]	The Editor of a Brief History of the Metaverse, Eco. The First AI False Information Case in Gansu, a Man Used ChatGPT to Make Up Fake News for Profit[EB/OL]. (2023-05-09)[2023-12-31]. https://www.thepaper.cn/newsDetail_forward_22999940.
	元宇宙简史编辑 Eco. 甘肃首例AI虚假信息案,男子用ChatGPT编假新闻牟利[EB/OL]. (2023-05-09)[2023-12-31]. https://www.thepaper.cn/newsDetail_forward_22999940.
[6]	CHAKRABORTY S, BEDI A S, ZHU Sicheng, et al. On the Possibilities of AI-Generated Text Detection[EB/OL]. (2023-10-02)[2023-12-31]. https://arxiv.org/abs/2304.04736.
[7]	HOLTZMAN A, BUYS J, DU L, et al. The Curious Case of Neural Text Degeneration[C]// ICLR. In 8th International Conference on Learning Representations. New York: ICLR, 2020: 26-30.
[8]	SOLAIMAN I, BRUNDAGE M, CLARK J, et al. Release Strategies and the Social Impacts of Language Models[EB/OL]. (2019-12-13)[2023-12-31]. https://doi.org/10.48550/arXiv.1908.09203.
[9]	GRITSAY G, GRABOVOY A, CHEKHOVICH Y V. Automatic Detection of Machine Generated Texts: Need More Tokens[C]// IEEE. 2022 Ivannikov Memorial Workshop (IVMEM). New York: IEEE, 2022: 20-26.
[10]	GUO Biyang, ZHANG Xin, WANG Ziyuan, et al. How Close Is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection[EB/OL]. (2023-01-18)[2023-12-31]. https://doi.org/10.48550/arXiv.2301.07597.
[11]	HE Xinlei, SHEN Xinyue, CHEN Zeyuan, et al. MGTBench: Benchmarking Machine-Generated Text Detection[EB/OL]. (2023-03-26)[2024-01-16]. https://doi.org/10.48550/arXiv.2303.14822.
[12]	TIAN Yuchuan, CHEN Hanting, WANG Xutao, et al. Multiscale Positive-Unlabeled Detection of AI-Generated Texts[EB/OL]. (2023-05-29)[2023-09-29]. https://doi.org/10.48550/arXiv.2305.18149.
[13]	ANTOUN W, MOUILLERON V, SAGOT B, et al. Towards a Robust Detection of Language Model-Generated Text: Is ChatGPT That Easy to Detect?[C]// TALN. In Actes de CORIA-TALN 2023. Paris: ATALA, 2023: 14-27.
[14]	GEHRMANN S, STROBELT H, RUSH A. GLTR: Statistical Detection and Visualization of Generated Text[C]// ACL. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics:System Demonstrations. Stroudsburg: ACL, 2019: 111-116.
[15]	ZHONG Wanjun, TANG Duyu, XU Zenan, et al. Neural Deepfake Detection with Factual Structure of Text[C]// ACL. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Stroudsburg: ACL, 2020: 2461-2470.
[16]	FROHLING L, ZUBIAGA A. Feature-Based Detection of Automated Language Models: Tackling GPT-2, GPT-3 and Grover[J]. PeerJ Computer Science, 2021, 7: 443-454.
[17]	MITCHELL E, LEE Y, KHAZATSKY A, et al. DetectGPT: Zero-Shot Machine-Generated Text Detection Using Probability Curvature[C]// ACM. In Proceedings of the 40th International Conference on Machine Learning (ICML’23). New York: ACM, 2023: 24950-24962.
[18]	SU Jinyan, ZHUO T, WANG Di, et al. DetectLLM: Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text[C]// ACL. Findings of the Association for Computational Linguistics:EMNLP 2023. Stroudsburg: ACL, 2023: 12395-12412.
[19]	DESAIRE H, CHUA A E, KIM M G, et al. Accurately Detecting AI Text When ChatGPT Is Told to Write Like a Chemist[J]. Cell Reports Physical Science, 2023, 4(11): 101672-101685.
[20]	WU Kangxi, PANG Liang, SHEN Huawei, et al. LLMDet: A Third Party Large Language Models Generated Text Detection Tool[C]// ACL.Findings of the Association for Computational Linguistics:EMNLP 2023. Stroudsburg: ACL, 2023: 2113-2133.
[21]	ABDELNABI S, FRITZ M. Adversarial Watermarking Transformer: Towards Tracing Text Provenance with Data Hiding[C]// IEEE. 2021 IEEE Symposium on Security and Privacy (SP). New York: IEEE, 2021: 121-140.
[22]	KIRCHENBAUER J, GEIPING J, WEN Y, et al. A Watermark for Large Language Models[C]// ICML. Proceedings of the 40th International Conference on Machine Learning. New York: ICML, 2023: 17061-17084.
[23]	KRISHNA K, SONG Y, KARPINSKA M, et al. Paraphrasing Evades Detectors of AI-Generated Text, But Retrieval Is an Effective Defense[C]// NIPS. Proceedings of the 37th International Conference on Neural Information Processing Systems (NIPS’23). New York: NIPS., 2024: 27469-27500.
[24]	UCHENDU A, LE T, SHU Kai, et al. Authorship Attribution for Neural Text Generation[C]// ACL. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Stroudsburg: ACL, 2020: 8384-8395.
[25]	DONG Xibin, YU Zhiwen, CAO Wenming, et al. A Survey on Ensemble Learning[J]. Frontiers of Computer Science, 2019, 14(2): 241-258.

实验环境	参数
操作系统	Windows 11
CPU	Intel(R) Core(TM) i9-9900K
GPU	2 × GeForce RTX 2080 Ti
内存	64 GB
Python IDE	PyCharm 2023.2.1
Python	3.10.13
PyTorch	2.0.0

数据集		<问题，人类回答，ChatGPT回答>	来源
Wiki-HC	地理	479	维基百科
	生物	355
	物理	585
	国际象棋	157
HC3-ALL	经济学	392	Human ChatGPT Comparison Corpus（HC3）^[10]
	医学	392
	开放问答	392
	计算机科学	392

数据集	检测器	准确率	精确度	召回率	F1值
Wiki-HC	Log-Probability Detector	90.19%	88.34%	92.60%	90.42%
	Log-Rank Detector	90.19%	88.11%	92.93%	90.45%
	Entropy Detector	69.94%	68.24%	74.60%	71.27%
	GLTR	89.07%	86.27%	92.93%	89.47%
	DetectGPT	79.26%	78.09%	81.35%	79.69%
	BERT Detector	98.23%	96.58%	100.00%	98.26%
	EBF Detection	98.55%	98.40%	98.71%	98.56%
HC3-ALL	Log-Probability Detector	96.33%	95.31%	97.44%	96.37%
	Log-Rank Detector	96.65%	96.20%	97.12%	96.66%
	Entropy Detector	87.70%	85.98%	90.10%	87.99%
	GLTR	96.33%	95.31%	97.44%	96.37%
	DetectGPT	86.89%	87.27%	86.38%	86.82%
	BERT Detector	98.24%	97.48%	99.04%	98.26%
	EBF Detection	98.88%	99.35%	98.40%	98.88%

训练集	测试集	检测器	准确率	精确度	召回率	F1值
Wiki-HC	HC3-ALL	Log-Probability Detector	95.15%	93.54%	97.00%	95.24%
		Log-Rank Detector	95.50%	94.40%	96.75%	95.56%
		Entropy Detector	85.84%	83.73%	88.97%	86.27%
		GLTR	95.12%	92.96%	97.64%	95.24%
		DetectGPT	84.60%	91.23%	76.55%	83.25%
		BERT Detector	94.26%	97.66%	90.69%	94.05%
		EBF Detection	97.51%	99.53%	95.47%	97.46%
HC3-ALL	Wiki-HC	Log-Probability Detector	89.77%	90.58%	88.77%	89.67%
		Log-Rank Detector	90.44%	90.08%	90.89%	90.49%
		Entropy Detector	69.85%	69.60%	70.49%	70.04%
		GLTR	88.10%	88.17%	88.01%	88.09%
		DetectGPT	77.71%	72.81%	88.45%	79.87%
		BERT Detector	85.66%	77.80%	99.81%	87.44%
		EBF Detection	96.06%	97.11%	94.93%	96.01%

数据集	检测器	准确率	精确度	召回率	F1值
Wiki-HC	Watermark	96.95%	100.00%	93.89%	96.85%
Wiki-HC	EBF Detection	99.12%	100.00%	98.25%	99.12%
HC3-ALL	Watermark	95.21%	100.00%	90.42%	94.97%
HC3-ALL	EBF Detection	96.00%	96.00%	96.00%	96.00%