텍스트 기반 스플릿 러닝 환경에서의 데이터 추론 공격

초록

As artificial intelligence technology advances, the demand for extensive data and computing resources for training has significantly increased. Despite these advancements, technology development faces challenges due to limited computing resources and privacy concerns related to data ownership. To address these issues, distributed learning technologies have emerged, with split learning gaining significant attention as a method that allows collaborative model training without sharing raw data. However, recent studies have revealed security vulnerabilities in split learning, where inference attacks using smashed data can secretly extract sensitive information from clients. This paper discusses the risks associated with the security vulnerabilities of split learning and proposes a text-based inference attack framework to demonstrate these vulnerabilities. The proposed attack technique involves the server, acting as an attacker, performing vector matching between the client-side smashed data and a reference smashed dataset to infer the client’s local data. In this paper demonstrates the vulnerabilities in split learning models, including those handling text data, and underscores the need for enhanced data privacy mechanisms to strengthen privacy protection in distributed learning environments.

키워드