Secure AirComp Federated Learning With STAR–IRS

In wireless environments, over-the-air computation (AirComp) can exploit the superposition nature of the channel to facilitate federated learning (FL). When a large number of Internet of Things (IoT) devices participate in the FL process, employing a simultaneously transmitting and reflecting (STAR) intelligent reflecting surface (IRS) can significantly improve FL aggregation performance by enabling wider user participation. However, the STAR–IRS may also introduce substantial security risks as eavesdroppers can also enjoy the benefits of STAR–IRS. In this article, a secure design that enhances the mean-squared error (MSE) performance of the basestation (BS) while disrupting the eavesdropper’s performance is proposed. The BS transmits a jamming noise to the eavesdropper on the same frequency band that devices transmit their local models to the BS. An optimization problem is formulated to minimize the aggregation error at the BS while ensuring the error exceeds a certain threshold at the eavesdropper. Then, a multistep procedure, utilizing alternating optimization (AO) and difference-of-convex (DC) algorithm, is proposed to solve this problem. Numerical results demonstrate the effectiveness of the proposed solution, resulting in low MSE and fast convergence in the learning process for the BS, while also deteriorating the eavesdropper’s learning and detection performance.