A game-theoretic method based on Q-learning to invalidate criminal smart contracts

Criminal smart contracts, severely threatening the security of cyberspace, allow criminals to maximize their utilities through illegal behaviors. The validity of criminal smart contracts is an indicator of criminals’ success. While the validity of criminal smart contracts heavily hinges on parameters derived from data feed. Therefore, criminals have incentives to increase these contracts’ validity by biasing the parameters therein. In this paper, we formalize data feed parameters by utilizing stochastic distributions, allowing us to analyze criminal smart contracts as state-based games and evaluate their validity through state-arrival probabilities. The main target of this paper is to decrease the validity of criminal smart contracts to prevent criminals’ illegal behaviors. To this end, Q-learning is utilized to train distribution parameters so that criminals have a low probability of reaching their desirable state. This impairs the validity of criminal smart contracts and minimizes criminals’ utilities. To the best of our knowledge, it's the first implementation of machine learning in the analysis of smart contracts. The experiments show that our method is at least an order of magnitude lower than previous works under the same settings with respect to the validity of criminal smart contracts.