Anti-Packet-Loss Encrypted Traffic Classification via Masked Autoencoder

Chen Yang, Li Guo, Zheyuan Gu, Gaopeng Gou, Xinjie Lin, Gang Xiong, Yangyang Guan

Abstract: Encrypted traffic classification is an important technique for network management and network security assurance. Current mainstream methods utilize machine learning and deep learning techniques for classification, and train models with complete data pre-collected in experimental environments to achieve excellent classification performance. However, real network environments are widely affected by packet loss, which causes traffic data to exhibit a different distribution compared to the training and negatively impacts existing methods. To address this problem, in this paper, we propose APMA, an Anti-Packet-loss encrypted traffic classification method based on Masked Autoencoder, which utilizes two traffic-specific mask pre-training tasks to extract generalized knowledge from large-scale unlabeled traffic, and simulates packet loss to learn robust traffic representations that are applicable to different network environments. Extensive evaluations show that APMA can be applied to a variety of classification scenarios and network environments, achieving over 90% classification accuracy even at 15% packet loss.

External IDs:dblp:conf/wasa/YangGGGLXG24