Federated UE-Side False Base Station Detection via Location-RSRP Consistency

Abstract

False base stations (FBSs) impersonate legitimate cells, making identifier-based detection unreliable. Representative attack data are scarce, while centralized collection of user equipment (UE) locations and reference signal received power (RSRP) measurements exposes sensitive information. This paper introduces location-RSRP consistency, the normal statistical relationship between a UE location and per-cell RSRP, and proposes a federated learning-based UE-side detector trained only on normal data. Regional clients retain raw measurements and collaboratively train a global regression model, which is distributed with region-specific thresholds for local detection. Network Simulator 3 experiments identified forward 1D-CNN-FedDyn as the preferred deployment configuration. Across three repeated runs, forward 1D-CNN-FedDyn achieved an F1-score of 0.987 ± 0.004, whereas forward Transformer-FedDyn achieved the highest F1-score of 0.993 ± 1.3×10−4. Compared with Transformer-FedDyn, 1D-CNN-FedDyn used a 24.4 times smaller model artifact and provided 35.2 times higher inference throughput, offering the most favorable balance between detection performance and deployment efficiency.