DistillSeq: A Framework for Safety Alignment Testing in Large Language Models using Knowledge Distillation

doi:10.1145/3650212.3680304

	DistillSeq: A Framework for Safety Alignment Testing in Large Language Models using Knowledge Distillation
	Yang, Mingke1 ; Chen, Yuqi1 ; Liu, Yi 2; Shi, Ling 2
	2024-09-11
会议录名称	ISSTA 2024 - PROCEEDINGS OF THE 33RD ACM SIGSOFT INTERNATIONAL SYMPOSIUM ON SOFTWARE TESTING AND ANALYSIS
页码	578-589
DOI	10.1145/3650212.3680304
摘要	Large Language Models (LLMs) have showcased their remarkable capabilities in diverse domains, encompassing natural language understanding, translation, and even code generation. The potential for LLMs to generate harmful content is a significant concern. This risk necessitates rigorous testing and comprehensive evaluation of LLMs to ensure safe and responsible use. However, extensive testing of LLMs requires substantial computational resources, making it an expensive endeavor. Therefore, exploring cost-saving strategies during the testing phase is crucial to balance the need for thorough evaluation with the constraints of resource availability. To address this, our approach begins by transferring the moderation knowledge from an LLM to a small model. Subsequently, we deploy two distinct strategies for generating malicious queries: one based on a syntax tree approach, and the other leveraging an LLM-based method. Finally, our approach incorporates a sequential filter-test process designed to identify test cases that are prone to eliciting toxic responses. By doing so, we significantly curtail unnecessary or unproductive interactions with LLMs, thereby streamlining the testing process. Our research evaluated the efficacy of DistillSeq across four LLMs: GPT-3.5, GPT-4.0, Vicuna-13B, and Llama-13B. In the absence of DistillSeq, the observed attack success rates on these LLMs stood at 31.5% for GPT-3.5, 21.4% for GPT-4.0, 28.3% for Vicuna-13B, and 30.9% for Llama-13B. However, upon the application of DistillSeq, these success rates notably increased to 58.5%, 50.7%, 52.5%, and 54.4%, respectively. This translated to an average escalation in attack success rate by a factor of 93.0% when compared to scenarios without the use of DistillSeq. Such findings highlight the significant enhancement DistillSeq offers in terms of reducing the time and resource investment required for effectively testing LLMs. © 2024 Owner/Author.
会议录编者/会议主办者	ACM SIGSOFT ; AITO
关键词	Digital elevation model Model checking Risk assessment Risk perception Automated testing Codegeneration Comprehensive evaluation Diverse domains Extensive testing Knowledge distillation Language model Large language model Natural language understanding Testing/Evaluation
会议名称	33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, ISSTA 2024
会议地点	Vienna, Austria
会议日期	September 16, 2024 - September 20, 2024
URL	查看原文
收录类别	EI
语种	英语
出版者	Association for Computing Machinery, Inc
EI入藏号	20244117161099
EI分类号	1102.1 ; 1106.2 ; 1106.3.1 ; 1108 ; 914.1 Accidents and Accident Prevention
原始文献类型	Conference article (CA)
文献类型	会议论文
条目标识符	https://kms.shanghaitech.edu.cn/handle/2MSLDSTB/436541
专题	信息科学与技术学院_PI研究组_陈宇奇信息科学与技术学院_硕士生
通讯作者	Chen, Yuqi
作者单位	1.ShanghaiTech University, Shanghai, China; 2.Nanyang Technological University, Singapore, Singapore
第一作者单位	上海科技大学
通讯作者单位	上海科技大学
第一作者的第一单位	上海科技大学
推荐引用方式 GB/T 7714	Yang, Mingke,Chen, Yuqi,Liu, Yi,et al. DistillSeq: A Framework for Safety Alignment Testing in Large Language Models using Knowledge Distillation[C]//ACM SIGSOFT, AITO:Association for Computing Machinery, Inc,2024:578-589.