Hier-SFL: Client-edge-cloud collaborative traffic classification framework based on hierarchical federated split learning

Abstract

Network traffic classification technology is widely noticed by internet service providers(ISP), in order to monitor network security status, improving knowledge of users’ traffic demands and then ulterior design policing and prioritization mechanisms. With the continuous development of deep learning technology, network traffic classification technology starts to use supervised learning paradigm to automate traffic identification by extracting some features. However, due to the continuous emergence of new private protocols, network services and access devices, the overall distribution of network traffic data is constantly changing. The training of network traffic classification models usually relies on the labeled data collected by academic organizations, institutions and companies. These data are usually generated in a fixed network environment, and the types of labels collected are limited. Models trained in this way are difficult to adapt to the open and complex network environment. Furthermore, the unevenly distributed computing resources affect the model training efficiency because the collected traffic data is usually concentrated in the end system but the computing power is generally concentrated in the edge computing nodes and cloud centers. Federated learning technology can coordinate data holders to train globally adapted models from discrete data. At the same time, split learning technology can realize more efficient model separation training. In this paper, we designed a new three-tier algorithm Hier-SFL combining and improving these two learning frameworks according to the common client-edge-cloud structure. Hier-SFL algorithm allocated the training task of neural network to each level network device with different computing resources. Also, it took the advantage of federated learning to periodically aggregate network parameters. Thus, we significantly improved the training efficiency of the deep learning model of traffic classification.