A Federated Continual Learning Framework for Sustainable Network Anomaly Detection in O-RAN
Chafika Benzaïd, Fariha Hossain, Tarik Taleb, Pedro Merino, Michael Dieudonné
Abstract
The distributed and disaggregated nature of 5G and beyond (B5G) networks has spurred interest in federated learning (FL) for empowering privacy-preserving collaborative network anomaly detection at the edge. However, FL is prone to catastrophic forgetting (CF), where prior knowledge is forgotten while sequentially learning new attack patterns from a stream of data. Few studies addressed CF issue in network anomaly detection using Continual Learning (CL), but focusing on centralized models rather than FL and overlooking integration in B5G. To fill this gap, we propose TenaxDoS, a novel framework that combines FL with a replay memory-based CL strategy to foster sustainable and cooperative network anomaly detection in an Open Radio Access Network (O-RAN) environment in B5G networks. The experimental results on a dataset from a real 5G test network show TenaxDoS's superior overall performance, stability and effective mitigation of CF, yielding a remembering of past knowledge of above 98.8%.