Green Differentially Private Coded Distributed Learning Over Near-Field MIMO Systems

Abstract

In near-field systems, differentially private coded distributed learning can effectively alleviate straggler issues and ensure privacy. However, it consumes more energy and entails substantial time for transmission. To further enhance the energy and training efficiency of the system, in this paper, we propose GMCDL, a green multiple-input multiple-output coded distributed learning framework based on differential privacy for near-field systems. GMCDL introduces both artificial noise into Lagrange-coded datasets and natural noise throughout the transmission process to achieve privacy protection, while also employing multiple antennas for faster transmission. Consequently, it leads to increased energy efficiency and accelerated transmission speeds. Firstly, we analyze the achievable privacy protection levels across various parameter configurations and explore how adjustments in these parameters influence privacy protection performance. Then, we establish a theoretical upper bound for system convergence, revealing two significant findings: 1) there is a trade-off between convergence performance and privacy protection level, and 2) with other parameters remain fixed, higher constrained power, increased receive antennas, reduced transmit antennas, and a larger power allocation toward effective computing result yield enhanced convergence performance. Finally, we provide experimental results to demonstrate the impact of various parameters on the training process, which aligns with our theoretical analysis.