Enhancing Federated Learning Robustness in Wireless Networks

To harness the benefits of machine learning (ML), users often face the challenge of sharing their private data with a central entity for model training. However, data sharing can be impractical due to privacy concerns, data size, wireless resource limitations, and other factors. Federated learning (FL) offers an efficient solution. In this approach, edge devices, users, or clients independently train machine learning models locally and iteratively share their model parameters with a central entity or server. The server aggregates these parameters into a global model, which is then distributed to all clients for the next round of training. Since the clients communicate with the server via wireless channels, over-the-air (OTA) computations have emerged as the optimal solution in the context of FL to resolve the challenge of aggregating local models. We investigate the influence of intrinsic noise introduced during OTA computations, focusing on the detrimental effects of impulsive noise on OTA-FL performance. Through a combination of theoretical analysis and experimental validation, we quantify the adverse impact of impulsive noise on convergence. We also introduce an algorithm designed to mitigate these effects. Our empirical results, obtained using CIFAR-10 and MNIST datasets, illustrate both the impact of impulsive noise on OTA-FL and the efficacy of our proposed solution.