An Efficient Federated Learning System for Network Intrusion Detection
Jianbin Li, Xin Tong, Jinwei Liu, Long Cheng
Abstract
Network intrusion detection is used to detect unauthorized activities on a digital network, with which the cybersecurity teams of organizations can then kick-start prevention protocols to protect the security of their networks and data. In real-life scenarios, due to the lack of high-quality attack instance data, building an in-depth network intrusion detection system (NIDS) is always challenging for a single enterprise, in terms of handling complex network security threats. To remedy the problem, this article proposes an efficient intrusion detection system called dynamic weighted aggregation federated learning (DAFL) based on federated learning. Specifically, DAFL has used the full advantages of federated learning for data privacy preservation. Moreover, compared to a conventional federated-learning based intrusion detection system, our scheme has implemented dynamic filtering and weighting strategies for local models. In this way, DAFL can perform better in detecting network intrusions with less communication overhead. We give the detailed designs of DAFL, and our experimental results demonstrate that DAFL can achieve excellent detection performance with a low network communication overhead, with data privacy preserved.