Security and reliability of rational players in distributed consensus

Abstract

Distributed ledgers and their applications in solving centralization problems in both financial and non-financial domains has been in the forefront of information security research since the emergence and the subsequent popularity of Blockchain. While the Proof of Work protocol has been successfully utilized for cryptocurrencies, the requirement for higher throughputs in non-financial domain based distributed ledgers favor alternate protocols whose consensus assumptions usually come with thresholds of Byzantine agents (faulty inputs) the consensus can withstand. Proof of Work is designed so that financial gain from conducting a successful attack is less than what honest participation would provide, eliminating any motivation an adversary might have to attack (within the context of direct gain). This assumption fails for non-financial solutions since resourceful malicious participants may exist where their gain may lie in manipulation of the distributed ledger or the order in which the transactions are recorded. A resourceful attacker could selectively convert rational agents to byzantine agents until the tolerance threshold is exceeded. Therefore, we propose that completeness assurance, and the overall reliability of distributed consensus requires rational and foresighted players to be sufficiently incentivized in affording costs of self-protection. We present a dynamic, complete, and imperfect information game to study the relationships between individual costs and utilities, tolerance threshold of the protocol and environment volatility in terms of exogenous attack probabilities, and observe conditions under which a mixed strategy equilibrium that preserves completeness would be stable. Our research extends existing literature by obtaining realistic resilience measures when considering rational player behavior in volatile environments, and provide a better understanding of mandatory security requirements that need to be implemented by a protocol designer for security in distributed consensus. We evaluate our proposed model using efficiency measurement concepts such as Price of Anarchy and Price of Malice, alongside learning methodologies such as regret matching and bounded rationality for extended insight. Our evaluations follow the theoretical predictions of the proposed model. Our results confirm reputation optimization to be capable of completeness assurance when the benefits are carefully assigned with consideration to tolerance threshold of the network. Our experiments also indicate that reputation optimization has attractive stability and convergence properties that are absent in other learning methodologies considered for evaluation