Split learning is emerging as a powerful approach to decentralized machine learning, but the urgent task of unlearning to address privacy issues presents significant challenges. Conventional methods of retraining from scratch or gradient ascending require all clients' involvement, incurring high computational and communication overhead, particularly in public networks where clients lack resources and may be reluctant to participate in unlearning processes they have no interest. In this short article, we propose \textsc{SplitWiper}, a new framework that integrates the concept of SISA to reduce retraining costs and ensures no interference between the unlearning client and others in public networks. Recognizing the inherent sharding in split learning, we first establish the SISA-based design of \textsc{SplitWiper}. This forms the premise for conceptualizing two unlearning strategies for label-sharing and non-label-sharing scenarios. This article represents an earlier edition, with extensive experiments being conducted for the forthcoming full version.
View on arXiv