GAN Tunnel: Network Traffic Steganography by Using GANs to Counter Internet Traffic Classifiers

Abstract

In this paper, we introduce a novel traffic masking method, called Generative Adversarial Network (GAN) tunnel, to protect the identity of applications that generate network traffic from classification by adversarial Internet traffic classifiers (ITCs). Such ITCs have been used in the past for website fingerprinting and detection of network protocols. Their use is becoming more ubiquitous than before for inferring user information. ITCs based on machine learning can identify user applications by analyzing the statistical features of encrypted packets. Our proposed GAN tunnel generates traffic that mimics a decoy application and encapsulates actual user traffic in the GAN-generated traffic to prevent classification from adversarial ITCs. We show that the statistical distributions of the generated traffic features closely resemble those of the actual network traffic. Therefore, the actual user applications and information associated with the user remain anonymous. We test the GAN tunnel traffic against high-performing ITCs, such as Random Forest and eXtreme Gradient Boosting (XGBoost), and we show that the GAN tunnel protects the identity of the source applications effectively.

Keywords

• Generative adversarial networks
• GAN
• Internet packet classification
• machine learning
• online privacy
• statistical traffic properties