GAN-Based Covert Communications Against an Adversary with Uncertain Detection Threshold in Federated Learning Networks

Federated learning (FL) is a promising distributed machine learning paradigm that can address privacy and security issues in the Internet of Things, Beyond 5G and 6G, and so on. Typically, numerous mobile devices in the FL network enable the server to complete model aggregation by uploading trained models obtained from local data. However, due to the broadcast nature of wireless channels, this upload process is vulnerable to adversary's monitoring or potential attacks, and it becomes one of the important risks of model leakage. Physical layer covert communications can protect the existence of covert signals or links, which is one of the effective methods to protect this model uplink. Furthermore, when there is an adversary with an uncertain detection threshold in the network, there is a competitive game between the adversary and mobile devices, which poses a challenge to the transmit power design of covert schemes. Considering the dynamic competition process in covert communications is similar to the zero-sum game process between discriminators and generators in Generative Adversarial Networks (GANs). Therefore, we propose a GAN-based physical layer covert communications against the adversary and protect the uplink in the FL networks. Moreover, the generator imitates the mobile device and the discriminator imitates the adversary. Both are constructed using neural networks, and the learning optimization process is achieved through a zero-sum game between them. Experiment results show that the designed algorithm is convergent and can obtain the optimal covert transmit power and the maximum average communication rate.