Employing Federated Semi-supervised Learning for Network Traffic Classification

Abstract

Network traffic classification is a critical aspect of network management. Software-defined networking (SDN) technology offers a novel approach to network management by separating the control plane from the data plane, enabling controllers to programmatically and efficiently configure the network and enhance its performance. This paper proposes improving network performance through traffic classification in the context of an SDN environment. However, implementing this idea involves several design options for the architecture and classification methods, each presenting unique challenges. For example, the classification module can be deployed on either the controller or the switch. When implemented on the switch, issues related to data labeling arise. In contrast, implementing the module on the controller may restrict traffic feature extraction to single packets. The main contribution of this paper lies in exploring the feasibility of different design options. To this end, this paper proposes a federated semi-supervised traffic classification method. Notably, in this federated semi-supervised learning framework, traffic feature extraction methods and classification models are interchangeable, allowing for substitutions based on specific application scenarios and design requirements. Consequently, the paper compares the performance of network traffic classification in (1) traffic feature extraction methods, (2) traffic classification algorithms, (3) centralized vs. federated learning, and (4) federated supervised vs. federated semi-supervised learning. Finally, while the motivation for this study arises from the context of SDN, the proposed federated semi-supervised traffic classification method is adaptable and applicable to a variety of use cases.