Node-Aware Dynamic Weighting Methods for Federated Learning on the Blockchain to Improve Performance

Abstract

In contrast to the traditional centralized learning approach, federated learning (FL) is a decentralized learning strategy. Every device’s local learning progress when the FL interacts with the central server, progressively improving the learning model. Nonetheless, it may cause network congestion because of limited transmission capacity and the participation of a large number of users. The model’s ability to converge quickly, steadily, and with target learning accuracy is one method for reducing the network load. We suggest a scenario for federated learning using blockchain in this study. Every user who participates in learning may be distinguished as a “node” via blockchain, which effectively encourages user participation. Integrity, stability, and other goals can also be pursued. When choosing between the two sorts of weights, the consumer in charge of refreshing the global representation should be consulted. When determining how much each customer matters, we first factor on how well they've learned locally.. Second, we take each client’s involvement frequency into consideration when determining the weight. To compare the effectiveness of our suggested system with that of other schemes, we select two critical performance indicators: learning rate and standard deviation. The simulation results demonstrate that, in comparison to existing schemes, our suggested system delivers greater stability and quick convergence time for the desired accuracy.