Using machine learning techniques for censorship-resistant communication

Abstract

To prevent citizens from accessing unfiltered information over the internet, repressive governments usually deploy national level censorship. To counter these censors, a number of systems have been introduced that make use of various allowed applications as covert channels and tunneling data through them. However, a drawback of these covert channels is their failure to maintain un-observability. The censors these days are capable enough to identify covert traffic amongst the stream of regular traffic. With censors getting more rigorous, some new ways of hiding blocked content in legitimate data streams have been introduced. These systems have very closely fooled the similarity based censors. However, Machine learning systems were still unbeaten. A recent study showed that some state-of-the-art Machine learning systems have been able to identify covert traffic, deeming the censor ship resistant models useless. In this thesis, we have used the strongest known Censorship-resistant system till date, to show that the system can be tweaked to defeat a machine learning based anomaly detection system. We have made use of adversarial machine learning techniques to beat some well-known classification techniques. Our results show that Adversarial examples can not only be used to hide blocked content in an application data stream but can also remain undetected.