Automating Attack Analysis on Blockchain Incentive Mechanisms with Deep Reinforcement Learning
- Mingxun Zhou, Turing Class (Class of 2016)
- Time: 2020-08-06 10:00
- Host: PKU Turing Class Research Committee
- Venue: Online Talk
Incentive mechanisms are central to the functionality of permissionless blockchains: they incentivize participants to run and secure the underlying consensus protocol. Designing incentive-compatible incentive mechanisms is notoriously challenging, however. As a result, most public blockchains today use incentive mechanisms whose security properties are poorly understood and largely untested. We proposed SquirRL as a new framework for using deep reinforcement learning to analyze attacks on blockchain incentive mechanisms. In this talk, I will introduce several novel empirical results we discovered by applying SquirRL. 1) The attacking strategies learned by SquirRL have the best performance in dynamic environment. 2) A counterintuitive ﬂaw in the widely used rushing adversary model when applied to multi-agent Markov games with incomplete information. 3) The optimal selﬁsh mining strategy is actually not a Nash equilibrium in the multi-agent selﬁsh mining setting. In fact, SquirRL suggests that there are no profitable NE when more than two parties are competing. 4) A novel attack on a simpliﬁed version of Ethereum’s ﬁnalization mechanism, Casper the Friendly Finality Gadget (FFG) that allows a strategic agent to amplify her rewards by up to 30%. Altogether, those results demonstrate SquirRL’s ﬂexibility and promise as a framework for studying attack settings that have thus far eluded theoretical and empirical understanding.