Abstract: Addressing the challenge that privacy protection and recognition accuracy are difficult to balance in distributed environments for network intrusion detection, this paper proposes a hybrid neural network model integrating a federated learning architecture. By integrating residual structures, gated recurrent units, and the squeeze-and-excitation channel attention mechanism, the model addresses the problems of insufficient temporal feature mining and lack of weight perception in network traffic. Meanwhile, the synthetic minority over-sampling technique and the label smoothing cross-entropy loss algorithm are utilized to effectively overcome the problem of model performance degradation caused by heterogeneous data distribution across nodes. Experimental results on the CICIDS2017 dataset demonstrate that the model achieves 99.80% in key metrics such as accuracy, precision, and recall, significantly outperforming existing federated learning and classical detection algorithms. These results validate the feasibility of achieving ultra-high precision collaborative defense without sharing raw sensitive data, which not only effectively alleviates the data island dilemma in the field of network security but also provides a reliable technical basis for constructing privacy-preserving intelligent defense systems.

Key words:  , intrusion detection, federated learning, channel attention mechanism, gated recurrent unit, residual network