Secure Federated Deep Reinforcement Learning with Blockchain

This paper introduces {\sf FedRLChain}, a novel framework for blockchain-based secure federated deep reinforcement learning, which allows users to securely and collaboratively train a Deep Reinforcement Learning (DRL) model by plugging appropriate aggregation and verification algorithms for specific problems. Unlike existing systems, {\sf FedRLChain} adopts (1) a novel verification algorithm to prevent malicious clients, (2) an aggregation weight scheme from preventing the global model from getting biased toward any client, and (3) a variant of traditional FedAverage algorithm to accelerate the convergence process. We perform a rigorous experimental evaluation of {\sf FedRLChain} considering the classic cart-pole problem, and we show a significant improvement in the number of epochs and time required for model convergence w.r.t. the state-of-the-art frameworks - DDQL, BAFFLE, and BASE-PIoT.