Unsupervised Lightweight Intrusion Detection Using Federated Learning for Heterogeneous Network Environments

Network intrusion detection systems frequently encounter challenges related to their capacity for generalization across varying contexts. Models that are trained exclusively on data from a particular network may exhibit vulnerability to data drift in other network environments. At the same time, enhancing generalization capability typically necessitates a substantial volume of data, which complicates the balancing of computational efficiency and privacy concerns. To address these challenges, this paper presents a lightweight intrusion detection system that is based on the integration of unsupervised deep autoencoder federated learning. This system enhances the range of models that can be learned by constructing a standard feature dataset in conjunction with federated learning to achieve generalization of heterogeneous networks. Concurrently, a lightweight algorithm is designed to expedite the training of federated learning. By simulating heterogeneous networks using four distinct datasets and performing cross-testing, it is demonstrated that our solution surpasses previous intrusion detection solutions in terms of generalization ability. The unsupervised federated learning-driven intrusion detection system based on lightweight design proposed in this paper can efficiently adapt to diverse network environments and improve computational efficiency.