RL-Based Federated Learning Framework Over Blockchain (RL-FL-BC)

Abstract

Federated learning (FL) paradigms aim to amalgamate diverse data properties stored locally at each user, while preserving data privacy through sharing users’ learning experiences and iteratively aggregating their local learning models into a global one. However, the majority of FL architectures with centralized cloud do not guarantee the trust in sharing users’ models, and hence, open the door for slowing and/or contaminating the global learning experience. In this paper, we propose a decentralized Blockchain (BC)-based framework and define a comprehensive protocol for exchanging local models, in order to guarantee users’ mutual trust while sharing their local learning experiences. We then propose a technique to optimize the global learning experience using Reinforcement Learning (RL), namely RL-FL-BC, to tackle the trade-off between information age of the learning parameters, data skewness (i.e., non-iid), and BC transaction cost (i.e., Ether price). We implement the proposed framework in a realistic containerized environment to facilitate the comparative study of the RL-FL-BC technique with baselines techniques. Our results show the efficacy of the BC-based protocol to facilitate the exchange of both the models’ and the optimization parameters to guarantee users’ mutual trust, while improving global learning performance compared to baselines techniques.