Abstract:

While deep learning-based traffic analysis technology improves network management efficiency, it also opens up new intrusion paths for malicious attackers. Users’ sensitive information can be extracted by analyzing the temporal characteristics of encrypted traffic, thereby posing a serious threat to individual privacy and security. The current defense strategies mainly relied on adversarial example to mislead adversaries’ classifiers. However, the application of these strategies encountered significant limitations in real-world scenarios. On the one hand, existing strategies confine to perturbing the feature space and are unable to impact real traffic. On the other hand, defense methods depend on understanding the attacker model, only proving effective in white-box environments. Given the insufficient research on obfuscating real traffic in black-box environments, the paper proposed a traffic obfuscation method for temporal features based on adversarial example named TAP. TAP was capable of generating effective adversarial perturbations targeting temporal features without requiring access to the adversary’s classifier. The core concept of TAP involved inserting a small number of packets into unidirectional communication flows, effectively resisting traffic analysis based on temporal features without disrupting normal communication. The experimental results show that TAP significantly reduce the accuracy of adversary traffic classification methods, with a bandwidth overhead of no more than 7%.

Key words: traffic obfuscation, adversarial example, generative adversarial network, traffic analysis