프로토타입 기반 전역 특성 공간 생성을 통한 분할 학습의 연속 학습 기법

초록

Split Learning (SL) enables collaborative model training without sharing private data, but suffers from catastrophic forgetting and feature misalignment in continual learning settings. This study proposes a prototype-based global feature alignment method that stabilizes learning across sequential tasks. The server maintains class-wise prototypes, and clients align their smashed data toward these global centers via prototype regularization and consistency constraints. Experiments on the iCIFAR-20 benchmark demonstrate that the proposed approach improves accuracy by approximately 1.4% over vanilla SL, effectively reducing forgetting and enhancing inter- client consistency.