Federated Learning for Anomaly Detection in Edge–Cloud Communication Systems.
Nina Großegesse, Christian Maier, Felix Strohmeier, Vinicius Jose De Menezes Pereira (2025): Federated Learning for Anomaly Detection in Edge–Cloud Communication Systems. In: IEEEXplore
Modern communication networks in industrial environments are increasingly heterogeneous and distributed across the Edge–Cloud Continuum (ECC). Ensuring reliable and low latency connectivity in such settings requires effective anomaly detection, yet challenges arise from non-independent and identically distributed (non-IID) data, privacy concerns, and the need to balance accuracy with communication efficiency. This paper presents an anomaly detection approach for edge-cloud communication networks in industrial environments that combines a classical autoencoder with a MLP-based threshold selection mechanism. Decentralized, centralized, and federated learning scenarios were evaluated in Wi-Fi and 5G environments. The proposed MLP-based threshold selection outperforms the standard percentile-based method, achieving up to 8% and 16% improvements in F1-score for anomaly and normal samples, respectively, in the 5G case. Among the federated learning strategies tested, FedAvg achieved the lowest reconstruction error in autoencoder training and the highest F1-scores in MLPbased classification. A comparison between the different scenarios shows that federated learning can achieve performance levels comparable to centralized learning while significantly reducing communication costs, but to the trade-off of increased training time due to slower convergence in non-IID settings. These results demonstrate that the proposed approach enables effective, privacy-preserving anomaly detection in heterogeneous network environments.
