Covert Communication over Federated Learning Channel

Abstract

We propose a novel covert communication technique between the federated learning (FL) server and participants without affecting the FL performance. The FL server superimposes the covert message onto the aggregated gradient and broadcasts the superimposed signal to all FL participants. FL participants decode the covert message treating the aggregated gradient as interference, and restore the original global model after removing the covert message from the superimposed signal. Therefore, the FL performance is not affected by sending the covert message. We analyze the covertness of communication against the adversary that monitors the statistical distribution of model updates. We derive the maximum achievable transmission rate of the covert message without being detected by the adversary and without affecting the federated learning performance.