Censorship Circumvention Using Generative Adversarial Networks

Abstract

Censorship circumventing technologies have been developed in response to attempts to censor Internet communication, but the technological capabilities of censors have continued to advance. Recent approaches to censorship circumvention have focused on multimedia protocol tunneling as a means to transmit covert information while evading detection by censors. One such approach, Voiceover, is an audio-based protocol tunnel that encodes covert data in audio signals and shapes the audio signals to match the timing properties of human speech in order to mitigate a censor's ability to identify Voiceover traffic. However, any censorship circumvention regime needs to also provide reliable communication. Voiceover as currently proposed does not possess any guarantees of data integrity or the reliability of the protocol amid application-layer transformations and disruptions. This thesis aims to be a continuation and evaluation of the work done in Voiceover. Our first contribution is to implement a rudimentary reliability layer within Voiceover that provides for message integrity and increased message recoverability through notions of data framing, checksums and redundancy. Our second contribution is to improve the usability of Voiceover through automation, maximizing throughput, improving demodulation time, and increasing the robustness of bidirectional communication. Our third contribution is to demonstrate the value of choice of protocol tunnel and the flexibility of the reliability layer by showing that Skype for Web provides a transmission channel unobservable to packet size analysis. Our fourth contribution is to demonstrate the value of the novel audio shaping approach by showing that audio shaping decreases the ability of a classifier to identify Voiceover transmissions based on inter-packet timing statistics. These results demonstrate that the design choices of Voiceover go a long way to achieving unobservable communication.