U-shaped Vertical Split Learning with Local Differential Privacy for Privacy Preserving

Existing distributed learning techniques, including federated learning and split learning, have achieved significant advancements in data privacy protection. However, these methods often neglect the importance of label privacy, with labels typically stored on the server side, particularly in the case of vertically partitioned datasets. This paper presents U-shaped Vertical Split Learning (UVSL), a novel approach designed specifically to address the unmet need for label privacy in vertical data partitioning scenarios. UVSL introduces a U-architecture for preserving label privacy that, combined with local differential privacy, shields client output activations from reconstruction attacks, a vulnerability present in conventional VSL approaches. Our method not only enhances privacy but also strike a balance between privacy preservation and model performance, as evidenced by our extensive experimental evaluations across diverse datasets and partitioning configurations.