Cross-silo Federated Learning with Record-level Personalized Differential Privacy

• DOI: 10.48550/arxiv.2401.16251
• Publication Date: 2024-01-29
• Authors: Junxu Liu; Jian Lou; Li Xiong; Jinfei Liu; Xiaofeng Meng
• Keywords: Differential Privacy; Information silo; Federated Learning; Privacy Protection

Federated learning enhanced by differential privacy has emerged as a popular approach to better safeguard the of client-side data protecting clients' contributions during training process. Existing solutions typically assume uniform budget for all records and provide one-size-fits-all that may not be adequate meet each record's requirement. In this paper, we explore uncharted territory cross-silo FL with record-level personalized privacy. We devise novel framework named rPDP-FL, employing two-stage hybrid sampling scheme both client-level non-uniform accommodate varying requirements. A critical non-trivial problem is select ideal per-record probability q given {\epsilon}. introduce versatile solution Simulation-CurveFitting, allowing us uncover significant insight into nonlinear correlation between {\epsilon} derive an elegant mathematical model tackle problem. Our evaluation demonstrates our can performance gains over baselines do consider preservation.