Byzantine-Resilient Online Federated Learning with Applications to Network Traffic Classification

Rapid growth in distributed streaming data at the network edge in many applications has prompted the emergence of online federated learning (OFL), a promising distributed machine learning paradigm where multiple agents cooperate to perform online learning via a central server. Despite its distinctive capability in handling various real-world applications, OFL has yet to be widely adopted in the industry due to its vulnerability to the ubiquitous Byzantine attacks initiated by adversarial agents; these agents can influence the OFL process by sending arbitrary updates to the central server. Byzantine attacks in OFL still falls short of being satisfactorily addressed in research as reflected by current literature. In this work, we propose a Byzantine-resilient OFL algorithm called BROFL. Our theoretical analysis shows that BROFL has a sub-linear regret bound under mild conditions and has a polynomial time-complexity in computation. To validate the performance of BROFL, we conduct extensive evaluations based on the tasks of network anomaly detection and application identification and using several real-world datasets. The evaluation results illustrate that BROFL can approach the performance of an offline optimal classification model and achieve F1 score that is at least 37 percent higher than those of two online baselines, which validates the effectiveness of BROFL in practice.