Enhancing Accuracy-Privacy Trade-off in Differentially Private Split Learning (2310.14434v3)

Abstract: Split learning (SL) aims to protect user data privacy by distributing deep models between client-server and keeping private data locally. Only processed or `smashed' data can be transmitted from the clients to the server during the SL process. However, recently proposed model inversion attacks can recover the original data from the smashed data. In order to enhance privacy protection against such attacks, a strategy is to adopt differential privacy (DP), which involves safeguarding the smashed data at the expense of some accuracy loss. This paper presents the first investigation into the impact on accuracy when training multiple clients in SL with various privacy requirements. Subsequently, we propose an approach that reviews the DP noise distributions of other clients during client training to address the identified accuracy degradation. We also examine the application of DP to the local model of SL to gain insights into the trade-off between accuracy and privacy. Specifically, findings reveal that introducing noise in the later local layers offers the most favorable balance between accuracy and privacy. Drawing from our insights in the shallower layers, we propose an approach to reduce the size of smashed data to minimize data leakage while maintaining higher accuracy, optimizing the accuracy-privacy trade-off. Additionally, a smaller size of smashed data reduces communication overhead on the client side, mitigating one of the notable drawbacks of SL. Experiments with popular datasets demonstrate that our proposed approaches provide an optimal trade-off for incorporating DP into SL, ultimately enhancing training accuracy for multi-client SL with varying privacy requirements.