A Lightweight Deep Learning Framework for Encrypted Traffic Classification via Visual Flow Representation

With the increasing complexity of cybersecurity threats, encrypted network traffic is becoming more and more widely used in all kinds of malicious communications. Traditional feature engineering‐based machine learning methods encounter accuracy and scalability issues when identifying such traffic, while existing deep learning methods achieve high accuracy but have large parameter counts and poor portability. Based on this, this paper for the first time migrates and applies the Lightweight Deep Learning Framework for Encrypted Traffic Classification via Visual Flow Representation to the MobileViT architecture. This paper presents the first application of the lightweight deep learning model MobileViT architecture to the task of classifying encrypted network traffic. We designed structural adaptations tailored for time‐series network data and constructed an enhanced variant, MiTNet, specifically suited for this task, which integrates a convolutional neural network and a visual Transformer architecture to achieve high‐accuracy classification of encrypted network traffic. The model employs 1D depth‐separable convolution to construct lightweight inverted residual blocks, which extract local temporal features, and combines this with the MobileViT module to establish global dependencies, thereby improving the recognition of different traffic types. The experiments are evaluated on four typical network traffic classification datasets, and the model achieves more than 98% accuracy and F1 scores on all four datasets while maintaining a very low number of parameters, balancing accuracy and inference efficiency, demonstrating the feasibility and superiority of the lightweight model for encrypted traffic recognition.