ForkRL: Deep Reinforcement Learning-Based Forking Prevention in Blockchain-Enabled Federated IoMT

The frequent occurrence of forking events in the blockchain-enabled federated Internet of Medical Things (IoMT) paradigm is one of the significant factors contributing to the latency and energy consumption overhead in the network. The occurrence of a forking event in the network necessitates restarting the entire Federated Learning (FL) process to ensure consistent global model updates in subsequent rounds. This imposes a substantial latency and energy consumption overhead, further straining the network’s efficiency. To address the aforementioned issue, this paper proposes a deep reinforcement learning-based device selection strategy to determine the optimal set of edge devices and miners for participating in an FL round that can minimize the likelihood of forking occurrences. The proposed Deep Reinforcement Learning (DRL)-based approach is evaluated on the PathMNIST dataset. Empirical results show that optimally selecting miners and edge devices in each round significantly reduces the number of forks, yielding a reduction in the network latency and energy consumption while preserving scalability as compared to the existing baseline approaches. The proposed approach achieves near-optimal performance, with a latency overhead of only 15% and energy consumption overhead of 25% compared to the optimal solution of the brute-force approach.