Covertness-Aware Over-the-Air Federated Learning

The over-the-air (OTA) computation which utilizes the waveform-superposition property of wireless signals has been considered as a promising approach to simultaneously accomplish communication and computing tasks in multiple access channels. Equipping federated learning (FL) with the OTA computation allows distributed Artificial Intelligence of Things (AIoT) devices to collaboratively train machine learning models over wireless environments, while preserving data privacy without excessive bandwidth consumption. However, the appearance of wardens in OTA-FL can result in model eavesdropping and the risk of information leakage. In this paper, we investigate how OTAFL can be fulfilled among AIoT devices while ensuring the covertness (i.e., the prevention of eavesdropping) against wardens. To this end, we formulate an optimization problem to maximize achievable data rates subject to covertness and transmit power constraints. Then, we first derive optimal detection thresholds to simplify the covertness constraints, and then propose the covertness-aware transmit power control (CAMEO) algorithm, which iteratively explores transmit powers that yield high target signal-to-interference-plus-noise ratios while satisfying both the covertness and transmit power constraints. Simulation results demonstrate that the CAMEO algorithm outperforms comparison schemes in terms of test accuracy, and it is also shown to converge well under various parameter settings.