Fully decentralized privacy-enabled Federated Learning system based on Byzantine-resilient consensus protocol

We present a novel blockchain-based Federated Learning (FL) system that introduces a Byzantine-resilient consensus protocol that performs well in the presence of adversarial participants. Unlike existing state-of-the-art, this system can be deployed in a fully decentralized manner, meaning it does not rely on any single actor to function correctly. Using a Smart Contract-driven workflow coupled with a commitment scheme and a differential privacy- based solution, we ensure training integrity, prevent plagiarism, and protect against leakage of sensitive data while performing effective federated training. We demonstrate the system's effectiveness by performing simulation and implementation of an end-to-end proof of concept. Our practical implementation showcases the system's efficiency on a single computer with multiple trainers, revealing low memory demands and manageable network and block I/O, which suggest scalability to larger, more complex networks. The paper concludes by exploring future enhancements, including advanced cryptographic methods for enhanced privacy and potential applications extending the system's utility to broader domains within FL. Our work lays the groundwork for a new generation of decentralized learning systems, promising increased adoption in real-world scenarios where data privacy and security are of paramount concern.