The Deep Flow Inspection Framework Based on Horizontal Federated Learning

The deep flow inspection (DFI) can identify abnormal traffic to avoid the data congestion caused by the sudden increase of traffic, which can maintain the stability of 6G network. However, the traditional DFI cannot protect the privacy of clients. This paper proposes a deep flow inspection method based on the horizontal federated learning to achieve the traffic identification locally, which reduces the risk of data leakage. Besides, a lightweight CNN model, Simplified-MobileNet, is proposed to realize the effective traffic identification under the limited hardware environment. Federated aggregation algorithm FedAvg is also applied to promote the communication efficiency during model training. The experimental results demonstrate that the Simplified-MobileNet decreases the training time per round by about 15%, and compared with the standalone mode, the FedAvg algorithm can achieve a higher training accuracy with a limited communication time.