PP-OpenNet: Privacy-Preserved Open Set Classification for Network Traffic

Network traffic classification is a fundamental and ubiquitous task in a wide range of network applications. It plays a critical role in enhancing the Quality of Service (QoS) for network systems and safeguarding Internet users from potential security threats. However, traditional methods relying on Deep Packet Inspection (DPI) face critical limitations. First, packet header and payload inspection poses severe privacy risks. Second, the rapid proliferation of applications introduces diverse, “unseen” traffic, complicating classification in open-world scenarios. To address these dual challenges, we propose PP-OpenNet, a lightweight deep-learning architecture for open-set traffic classification that relies solely on packet timestamp and length sequences. PP-OpenNet enhances the performance and computational efficiency by introducing disentangled multi-scale feature extraction and recurrent feature fusion mechanisms. We evaluate it on real-world datasets to demonstrate its superiority over state-of-the-art baselines and ablation variants in different open-set metrics. To the best of our knowledge, PP-OpenNet represents a novel solution to the dual challenges of open-set recognition and a privacy-preserving design in network traffic classification.

Recommended citation: Jingze Zhang, Leijie Wu, Xi Peng, Qingqing Yang, Ruilun Liu, and Hong Xu. "PP-OpenNet: Privacy-Preserved Open Set Classification for Network Traffic." IEEE International Conference on Computer Communications (INFOCOM). 2026: 1-6. https://ieeexplore.ieee.org/document/11571190