A model of encrypted network traffic classification that trades off accuracy and efficiency

Abstract As the internet industry evolves, the need for effective encrypted traffic classification (ETC) becomes vital for network management and cybersecurity. Meanwhile, existing deep learning (DL)-based methods struggle with balancing model complexity with accuracy. A significant challenge is deploying these models on dominant network devices in a way that ensures rapid and precise traffic classification. In this paper, we propose FasterTrafficNet, a novel DL-based lightweight ETC strategy designed for deployment on low-configuration networking devices. We designed the core component of FasterTrafficNet using a PConv-based concept, which efficiently extracts spatial features from data while reducing unnecessary computations and memory access, further enhancing the model's operational efficiency. Additionally, we integrated Do-Conv in place of conventional non-dot convolution to amplify the model's performance significantly without escalating the computational overhead during inference. We conducted a comparative analysis of FasterTrafficNet against six advanced ETC methods utilizing two publicly accessible benchmark datasets. Experimental results demonstrate that FasterTrafficNet, comprising 1.46 M model parameters, provides superior classification performance relative to the other methods. Consequently, FasterTrafficNet demonstrates a lightweight approach to ETC that can be used on extensive networks of devices.