A Realistic Approach for Network Traffic Obfuscation Using Adversarial Machine Learning

Abstract

Adversaries are becoming more sophisticated and standard countermeasures such as encryption are no longer enough to prevent traffic analysis from revealing important information about a network. Advanced encryption techniques are intended to mitigate network information exposure, but they remain vulnerable to statistical analysis of traffic features. An adversary can classify different applications and protocols from the observable statistical properties, especially from the meta-data (e.g. packet size, timing, flow directions, etc.). Several approaches are already being developed to protect computer network infrastructure from attacks using traffic analysis, but none of them are fully effective. We investigate solutions based on obfuscating the patterns in network traffic to make it more difficult to accurately use classification to extract information such as protocols or applications in use. A key problem of using obfuscation methods is to determine an appropriate algorithm that introduces minimal changes but preserves the functionality of the protocol. We apply Adversarial Machine Learning techniques to find realistic small perturbations that can improve the security and privacy of a network against traffic analysis. We introduce a novel approach for generating adversarial examples that obtains state-of-the-art performance compared to previous approaches, while considering more realistic constraints on perturbations.