FedEdgeTrans-AD: A Privacy-Preserving Federated Transformer for Edge-Based IoT Anomaly Detection

Modern IoT systems demand high-speed detection of anomalies, but centralized methods have privacy, latency, and communication overhead issues. In this paper, FedEdgeTrans-AD is proposed, which is a federated edgeintelligence framework based on a Transformer-based timeseries model with lightweight deployment on the IoT edge nodes. The use of local training steps and on-the-fly inference enables the system to learn a global model in cooperation with each other without sending the raw data. In order to address the limited availability of labels and non-IID data distributions, FedEdgeTrans-AD combines self-supervised learning (masked reconstruction and contrastive objectives) and federated aggregation (drift-aware proximal regularization). Furthermore, sparse and quantized model updates guarantee communication efficiency for bandwidth-constrained environments. Experimental evaluation in realistic non-IID conditions shows better results, 95.2 % accuracy, 81.3 % F1score, and 0.957 AUROC. In addition, the communication scheme offers a 71% reduction in uplink transmission costs (3.6 MB per round) with an ultra-low edge inference latency of 11.4 ms per window. These results make FedEdgeTrans-AD an efficient, secure, and privacy-preserving real-time anomaly detection solution for heterogeneous smart IoT networks.