Physical Layer Security for STAR-RIS-Assisted Federated Learning Systems With Differential Privacy

Zheng Yang

doi:10.1109/TWC.2026.3674798

Physical Layer Security for STAR-RIS-Assisted Federated Learning Systems With Differential Privacy

2026-03-24

0

PRE

AI

DOI：10.1109/TWC.2026.3674798

Original

Original request for help

Save

Abstract

En 中文

In this paper, we propose a federated learning (FL) system enhanced by a simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS), which is designed to protect client-side data privacy and reinforce the security of data exchange over wireless channels. Assuming an honest-but-curious central server that may infer private information from user gradients, we adopt differential privacy (DP) by injecting noise into local updates to safeguard user data. Theoretical results are derived to characterize the systematic privacy guarantees provided by the DP noise power and the gradient information in the proposed STAR-RIS-enabled DP-FL systems. Building on these results, the secrecy sum rate of local information is formulated by jointly optimizing the STAR-RIS coefficient matrices, users’ transmission power, artificial jamming power, and the power of DP noise introduced by the FL users. To tackle the non-convex optimization challenge, we develop a block coordinate descent algorithm that partitions the original problem into four manageable subproblems. The closed-form expressions are obtained for users’ transmit power, artificial jamming power, and the power of DP noise. For the STAR-RIS phase shift design, approximate solutions are derived through semidefinite relaxation combined with a surrogate lower bound method. Finally, simulation results demonstrate that the proposed STAR-RIS-enabled DP-FL systems achieve significantly improved secrecy performance compared to conventional FL systems with randomly configured STAR-RIS amplitude, phase shifts, and transmit power. Furthermore, the proposed FL algorithm achieves model training and testing performance that closely approximates that of FL without DP, highlighting its effectiveness in preserving both data privacy and model utility.

Keywords:

Federated learning

differential privacy

simultaneously transmitting and reflecting reconfigurable intelligent surface

secrecy sum rate

Physical Layer Security for STAR-RIS-Assisted Federated Learning Systems With Differential Privacy

Abstract

Journal

Organization