基于大模型敏感数据泄露风险的量化评估方法改进

打开文本图片集

中图分类号：TP309 文献标志码：A 文章编号：2095-2945（2025）34-0123-05

Abstract：WiththewidespreaduseofLarge Language Models（LLMs），theriskofsensitivedataleakagehasbecomean urgentproblem.Aimingattheshortcomingsof existingquantitativeaessmentmethodsforsensitivedataleakagerisks，thispaper proposesanimprovedasessment method.First，weanalyzetheprivacydisclosurerisksthatLLMmaycauseduring thetraining andreasoningphases，includingthedisclosureofsensitiveinformationintrainingdataandcueatacksduringthereasoning phase.Secondly，wesummarizethelimitationsofexistingprotectionmethods，suchastheshortcomingsofdatapreprocessing， diferentialprivacyandforgetingmethods.Onthisbasis，thispaperproposesanimprovedevaluationframework，including technicalmodulessuchassensitivedataidentificationandclassfication，iskscenariosimulation，andquantitativeevaluation indicatorconstructionByintroducingmulti-levelandmulti-dimensionalevaluationindicators，combinedwiththelatestprivacy protectiontechnologies，thispaperaimstoprovideamorecomprehensiveandobjectivequantitativeevaluationmethodforLLMs sensitive data leakage risks，thereby providing support for LLM's security application.

Keywords：largelanguagemodel;sensitivedatadisclosure;privacyprotection;quantitativeriskassessment；diferential privacy

近年来，大型语言模型如ChatGPT、GPT-4等在自然语言处理领域取得了显著进展，广泛应用于文本生成、机器翻译、对话系统等任务。（剩余6780字）