RWD: Incremental Reinforcement Learning with Wasserstein Distance for Efficient Encrypted Traffic Classification

Network security depends on encrypted traffic classification, which makes it possible to accurately identify and classify applications. Static encrypted traffic classification faces challenges due to the inherent complexity and variability of encrypted traffic, particularly as new applications emerge. While incremental learning offers a potential solution by enabling continuous model updates, its application to traffic classification is often constrained by issues such as traffic ambiguity and privacy concerns, limiting its effectiveness. In this paper, we present an innovative framework that integrates reinforcement learning with Wasserstein Distance knowledge distillation to overcome these challenges. Specially, we utilize the Proximal Policy Optimization agent to dynamically select high-quality samples from historical data while discarding ambiguous instances, enhancing training efficiency and mitigating traffic ambiguity. Concurrently, the Wasserstein Distance knowledge distillation module optimizes knowledge transfer across overlapping feature spaces, reducing transfer errors without extensive memory requirements. Experimental results on multiple benchmark datasets clearly demonstrate state-of-the-art performance improvements of approximately <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2.73 \%, 3.73 \%$</tex>, and 5.35 %.