Split Learning without Local Weight Sharing to Enhance Client-side Data Privacy

Split learning (SL) aims to protect user data privacy by splitting deep models between client-server and keeping private data locally. SL has been demonstrated to achieve similar accuracy as the centralized learning model. In SL with multiple clients, the local training weights are shared between clients for local model aggregation. This paper investigates the potential of data leakage due to local weight sharing among the clients in SL by performing model inversion attacks. To mitigate the identified leakage issue, we propose and analyze privacy-enhancement SL (P-SL), e.g., SL without local weight sharing, to boost client-side data privacy. We also propose paralleled P-SL to speed up the training process by employing multiple servers without accuracy reduction. Finally, we investigate P-SL with late participating clients and develop a server-based cache-based training to address the forgetting phenomenon in SL. Experimental results demonstrate that P-SL helps reduce up to 50% of client-side data leakage compared to SL. Moreover, P-SL and its cache-based version achieve comparable accuracy to SL under various data distributions with lower computation and communications costs. Also, caching in P-SL reduces the negative effect of forgetting, stabilizes the learning, and enables effective and low-complexity training in a dynamic environment with late-arriving clients.