Geneva: An AI Approach to Evading Nation-State Censorship

Abstract

Nation-state censorship is one of the greatest threats to free and open communication. For years, censors have engaged in an arms race with researchers and practitioners, resulting in myriad techniques for evading censorship, as well as counter-measures to stop them. Unfortunately, censors have long had the advantage, because evading censorship has required researchers to first measure and understand how censors operate before they can intuit methods of evading them. In this talk, we will present a drastic departure from the previously manual evade/detect cycle. We will introduce Geneva, a novel genetic algorithm we developed that automatically discovers how to manipulate packet streams to evade censors. We deployed Geneva around the world live against real-world censors in China, India, and Kazakhstan. Geneva discovered dozens of novel strategies, including many that exploit what appear to be bugs in nation-state censors. We will also present the first known strategies for evading censorship strictly from the server: without requiring any deployment of censorship evasion configuration, software, or hardware within the censoring regime at all, Geneva has learned strategies can subvert nation-state censorship at the server on behalf of clients. Geneva represents an important first step towards automating censorship evasion through the application of AI. We will end the talk discussing how these results may shape the future of the censorship arms race.