A Network Intrusion Detection Method for Various Information Systems Based on Federated and Deep Learning

Abstract

Under the premise of ensuring data privacy, traditional network intrusion detection (NID) methods cannot achieve high accuracy for different types of intrusions. A NID method combining transformer and federated learning (FedL) is proposed for this purpose. First, a multi-party collaborative learning framework was built based on FedL, which achieved data exchange and sharing. Then, by introducing the self-attention mechanism (AttM) to improve the traditional transformer, it could quickly converge. Finally, an NID model integrating transformer and FedL was constructed by combining DNN, GRU, and an encoder module composed of improved transformer, achieving accurate detection of network intrusion. The proposed NID method was compared with the other three methods. The results show that the proposed method has the highest NID accuracy and F1 score on the NSL-KDD and UNSW-NB15 dataset, with the highest accuracy reaching 99.65% and 89.25%, while the F1 score has the highest accuracy, reaching 99.45% and 88.13%, outperforming the other three comparative algorithms in terms of performance.