Hybrid compression for LSTM-based encrypted traffic classification model

Abstract

Traditional techniques for network traffic classification are no longer effective in handling the complexities of dynamic network environments. Moreover, deep learning methods, while powerful, demand substantial spatial and computational resources, resulting in increased latency and instability. In this paper, we propose an innovative approach to network traffic classification utilising an LSTM structure. This approach incorporates network pruning, knowledge refinement, and Generative Adversarial Networks (GAN) to reduce model size, accelerate training speed without compromising accuracy, and address challenges associated with unbalanced datasets in classification problems. Our methodology involves the pruning of unimportant filters from the teacher model, followed by retraining and knowledge distillation to generate the student model. Experimental show that the size of the pruned teacher model is only 25.69% of the original, resulting in a noteworthy 28.16% improvement in training speed. Additionally, the classification performance of various unbalanced traffic categories, such as VoIP and streaming, shows significant enhancement.