Robust Complex-Valued Federated Learning for Secure 6G Mobile Communications
Anders Buvarp, Stefan Werner · 2025
Federated learning (FL) has attracted interest as a decentralized machine learning paradigm, where multiple clients collaboratively train a shared model while retaining their raw data locally. By performing computations on-device and sending only model updates (e.g., gradients or weights) to a central server, FL preserves privacy and supports large-scale distributed training. However, the aggregation step is vulnerable to malicious participants, who can poison the global model. Moreover, most existing FL schemes employ real-valued neural networks with non-robust federated averaging for aggregation, which is illsuited to complex-valued data fundamental to wireless communication systems. For this purpose, we propose robust aggregation schemes for complex-valued FL aimed at secure sixth-generation mobile communications. Our robust aggregation method combines complex-valued projection statistics with a Schweppe-type complex-valued generalized M-estimator featuring regularized scatter matrix estimation. We treat subsets of the network weights and biases as points in high-dimensional complex-valued spaces C p , where adversarial clients produce spatial outliers. Our robust estimators are tuned for the data dimensions and take the C p-points as input. Experiments on dense urban 5G AWGN channels show that the proposed aggregator secures learning in the presence of up to 50% malicious clients, surpassing Krum and outperforming classical federated averaging.