FedAdapt-CAD: A Federated Learning Framework for Anomaly Detection with Client-Aware Aggregation and Dynamic Model Adaptation

In this research, we introduced FedAdaptCAD, a novel Federated Learning (FL) framework designed for Anomaly Detection (AD) in critical infrastructure systems. This framework emphasizes robustness, scalability, and privacy preservation in mitigating anomalies. Key innovations include Client-Aware Aggregation (CAA), which adjusts client contributions based on data quality and local model performance, and Dynamic Model Adaptation (DMA), which optimizes global model hyperparameters in real-time to respond to changing attack patterns. Evaluated on three benchmark datasets: WADI, CMAPSS, and BATADAL, FedAdapt-CAD demonstrates superior global accuracy and fairness among heterogeneous clients. The experimental results indicate a significant improvement in detection rates for rare and complex attacks, achieving up to a 15% enhancement in precision and recall compared to traditional FL methods such as FedAvg, FedProx, and FedH2L.