A Dynamic Traffic Camouflage Method Based on Adversarial Machine Learning for Connected Vehicles

While improving the convenience of Internet of Vehicles applications, its communication process faces serious privacy and security threats. Especially in the process of network layer transmission, the analysis of side-channel characteristics (such as packet size and traffic transmission mode) may lead to user privacy leakage. Most of the existing research focuses on website fingerprint camouflage or overall traffic statistical feature camouflage, and usually only designs a camouflage mechanism for a single application layer protocol, which is challenging to achieve unified protection in the multi-protocol scenario of the Internet of Vehicles. In this paper, we propose a method called LeakGAN Packets Traffic Camouflage (LPTC), which can provide traffic camouflage for multiple application-layer protocols simultaneously. The LPTC learns the sending pattern of target traffic packets and generates camouflaged traffic through an adversarial game between a generator and a discriminator, to effectively resist side-channel attacks without affecting communication performance. Experimental results indicate that the proposed LPTC scheme significantly degrades the classification performance of traffic transmission models based on K-Nearest Neighbors, Support Vector Machine, Bagged Trees, and Random Forest algorithms, reducing classification accuracy by 41.23%. In addition, the precision, recall, F1 score, and AUC value decreased by 41.82%, 61.85%, 49.57%, and 42.00%, respectively. Moreover, the LPTC can effectively withstand adaptive attackers who retrain deep learning classifiers on camouflaged traffic. The LPTC substantially enhances the indistinguishability between target and camouflaged traffic, thereby effectively mitigating user privacy leakage caused by side-channel analysis. Meanwhile, the system's bandwidth and delay performance is sufficient to support the transmission requirements of real-time vehicular networks.