Universal Patch Construction Method for Intelligent Obfuscation of Network Traffic Features

Deep learning-based encrypted network traffic identification technologies may lead to the leakage of sensitive network information. Existing adversarial sample defense schemes generally face challenges such as high bandwidth overhead, poor adaptability in black-box scenarios, and poor generality of defense methods. To address these issues, this paper proposes a universal patch construction method (UPCM) for intelligent obfuscation of network traffic features, based on a feature-reversible traffic representation graph, to construct an offline, universal, and undirected adversarial patch generation framework for confusing traffic at both the feature level and packet level. A Gaussian noise-driven adaptive perturbation strategy is designed to support intelligent obfuscation of the temporal features of network traffic. Experiments on real-world network traffic datasets show that UPCM achieves a defense success rate of over 85% in typical network traffic environments, with bandwidth overhead controlled within 10%. Furthermore, the adversarial patches generated by UPCM exhibit strong generality and transferability: a single patch can defend against all types of network traffic, and when migrated to other deep learning models based on temporal features, the defense success rate remains approximately 85%.