Learning to Behave: Improving Covert Channel Security with Behavior-Based Designs

Censorship-resistant communication systems generally use real-world cover protocols to establish a covert channel through which uncensored can occur. Unfortunately, many previously proposed inconsistently with the way humans normally those protocols, leading anomalous network traffic patterns that have been shown be discoverable by censors. In this paper, we argue censorship-resistant should follow two behavior-based design properties: (i) behavioral independence: isolate operation of their channels from and (ii) realism: either opportunistically existing genuine protocol instances or run new are modeled after ones. These properties ensure behavior system’s users will not degrade its security. We demonstrate how achieve these evaluation Raven, messaging system uses email identically email. Raven generative adversarial is trained on data control timing sizes messages it sends receives, transferred independently user actions. Our shows that, compared state-of-the-art email-based Mailet system, raises false-positive rate 3% 50% when detecting usage 100% recall.