Privacy Protection for Federated Learning in OFDM Aided Over-the-Air Computation System

Federated learning (FL) an emerging distributed and privacy-protecting machine learning paradigm. In FL, training is completed through iterative local training and gradient aggregation among multiple mobile devices (MDs) and the edge server (ES). Over-the-air computation can help to achieve fast aggregation when multiple MDs need to offload their local gradients to the ES. Orthogonal frequency division modulation (OFDM) can further speed up aggregation through enabling simultaneous symbol transmission over multiple fading blocks. In this paper, we focus on the FL system with OFDM aided over-the-air computation technique implemented. To overcome the challenge of non-linear distortion at the amplifier of OFDM transmitter and protect the privacy of each individual MU, we propose a new aggregation method, which normalizes the local gradient first and then add uniformly distributed artificial noise, before it is offloaded to the ES. With our proposed aggregation method, the performance of differential privacy (DP) is analyzed, which disclose the quantization relationship between the variance of artificial noise and the level of privacy protection. We also analyze the peak power at OFDM transmitter by our proposed method, which is shown to be surely less than traditional method. Numerical results verifies the convergence and effectiveness of our proposed strategy.