Decentralized Federated Learning
Gunwant Singh · 2025
The rapid advancement of distributed machine learning has created new opportunities for enhancing data privacy and fostering collaborative intelligence. However, significant challenges remain in achieving scalability, efficiency, security, and trust. This thesis aims to enhance Decentralized Federated Learning by addressing core challenges related to communication efficiency, convergence time, scalability, security, and privacy. The study involves a comprehensive review of existing literature to identify key gaps and challenges, followed by the development of innovative methods and frameworks designed through experiments to overcome these limitations. The proposed techniques integrate Blockchain technologies, such as sharded Blockchains and IPFS-based off-chain storage, to improve the transparency and scalability of federated systems. By systematically comparing synchronous, asynchronous, and hybrid federated approaches across Blockchain-based and non-Blockchain scenarios, this research uncovers critical trade-offs between speed, accuracy, convergence, and security. Experiments conducted on real-world IoT datasets reveal that asynchronous Blockchain-based Federated Learning achieves up to 7.93% faster convergence when compared to the baseline setup. Methods such as delta compression with sparsification and sharded Blockchains reduce convergence time by 2.91% and 4.94%, respectively. When the use of sharded Blockchains is combined with delta compression and sparsification, it achieves the highest overall improvement of 7.93% in convergence time and up to 8.89% reduction in epoch time, highlighting its efficiency in large-scale distributed environments. However, the variability introduced by stale updates in asynchronous systems underscores the importance of uniform contribution from nodes. Techniques such as adaptive aggregation strategy, weighted aggregation, partial updates, and compression of updates emerge as crucial strategies for enhancing model performance in heterogeneous environments. The findings of this study provide actionable insights for the development of next-generation, privacy-preserving AI systems capable of operating efficiently at scale.