스플릿 러닝에서의 데이터 프라이버시와 모델 효율성 간 trade-off 분석

Abstract

Split learning, a distributed learning approach that divides the model into client and server components based on a cut-layer, has recently gained attention. Moving the cut-layer closer to the client reduces computation complexity but raises data privacy concerns regarding the original input data. This study analyzes this trade-off through attack simulations on a text-based COVID-19 prediction model. The results show that increasing the distance of the cut-layer from the client side reduces the success rate of inference attacks on the original data, while decreasing model efficiency. Careful consideration of this trade-off is crucial for designing effective split learning models that balance privacy and computation requirements.