摘要
[研究目的]随着大语言模型(LLMs)的广泛应用,其中蕴含的海量数据普遍存在删除失灵现象,亟需探索有效的数据删除机制保障LLMs的健康发展。[研究方法]通过比较研究、案例分析、文献研究等方法,分析LLMs的数据删除存在数据不当遗留导致泄漏和数据无限聚合引发滥用的安全风险,总结出当前面临删除技术壁垒、成本巨大、技法协同失灵的规制困境。[研究结果/结论]为了实现LLMs数据删除安全与效率的平衡,破解“删除权虚置化”的治理困境,需突破单一治理模式,秉持技法协同的理念进行全流程应对,将数据删除的末端治理转向全流程嵌入:训练阶段优化数据筛选与存储,规范删除请求的发起与确认,建立体系化监管验证机制。
[Research purpose]With the widespread application of Large Language Models(LLMs),the persistent issue of ineffective data deletion has become increasingly critical,necessitating urgent exploration of effective data deletion mechanisms to ensure the secure and sustainable development of LLMs.[Research method]Employing comparative analysis,case studies,and literature review,security risks arising from improper data retention(leading to leakage)and indiscriminate data aggregation(enabling misuse)in LLMs are analyzed.Further,regulatory challenges are summarized,including technical barriers to deletion,prohibitive costs,and the misalignment of legal and technical governance frameworks.[Research result/conclusion]To balance security and efficiency in LLMs data deletion and address the governance dilemma of"ineffective implementation of the right to erasure",it is necessary to break through the single governance model and adhere to the concept of skill coordination to deal with the whole process.Shift the end governance of data deletion to full process embedding:optimizing data screening and storage protocols during the training phase,standardizing procedures for initiating and validating deletion requests,and instituting systematic regulatory verification mechanisms.
作者
张小小
宗绍昊
Zhang Xiaoxiao;Zong Shaohao(Southeast University,Nanjing 211189;China University of Political Science and Law,Beijing 100088)
出处
《情报杂志》
2025年第10期97-104,共8页
Journal of Intelligence
关键词
数据删除
大语言模型
技法协同
安全风险
data deletion
large language models
coordination between technology and law
security risk
作者简介
张小小,女,2001年生,硕士研究生,研究方向:数据法学;通信作者:宗绍昊,男,1998年生,博士研究生,研究方向:数据法学。
