Dual Anti-Non-IID in Personalized Federated Learning-based for Flow Attack Detection

Federated learning achieves knowledge sharing and capability sharing through global model aggregation and distribution, and has good performance in cross-organizational and cross-regional network attack detection. In order to solve the Non-IID distribution problem of network attacks on different clients, we proposed a personalized federated learning framework for network attack detection technology to resist Non-IID on the client and server, referred to as pFedRWA. pFedRWA includes three measures. The first measure is to generate a global sample set on the server to build a global initialization model. The second measure is to calculate the cosine similarity between the client local model and the global model as the weight coefficient when aggregating the global model. The third measure is to calculate the KL divergence between the client local model and the global model on the client to construct the regularization constraint term for local model training. Compared with other federated learning algorithms and ablation experiments, pFedRWA has a high accuracy in network attack detection and can effectively solve the Non-IID problem.