Federated Learning for Enhancing Physical-Layer Security in 6G Networks

This thesis investigates personalized federated learning (PFL) for wireless region classification, where each base station (BS) uses channel state information (CSI) that reflects its local propagation environment to predict which region a user is in. Because the BSs are in different locations, the CSI distributions are heterogeneous. We examine a two-stage PFL approach in which BSs jointly train a shared model and then adapt it using local data. Within this framework, we propose TuneHead, which freezes the backbone and updates only the classification head at each BS during personalization. In comparison with other PFL strategies and baseline methods, we evaluate TuneHead on synthetic channel datasets and on the DICHASUS dataset of real channel measurements. The numerical results show that TuneHead does not always achieve the highest accuracy. We also conduct layer-wise filter ablation to estimate potential reductions in the number of model parameters, and recovery training mitigates the resulting accuracy loss while maintaining the reduced model size.