Blockchain-Powered Federated Learning: A Secure and Decentralized Approach to Distributed AI

Abstract

Integrating both federated learning and blockchain technology proposed a writing solution to the inescapable challenges of data privacy, security, and trust in FL-based decentralized machine learning environments. This paper proposes a BEFL framework with a blockchain network to ensure the trustworthiness and accountability of federated learning. The results presented in the experimental evaluation, conducted on MNIST, CIFAR-10, and a healthcare dataset, demonstrate the superiority of BEFL over FL. The BEFL framework achieved 98.5% on MNIST, CIFAR-10 82.3%, and the healthcare dataset 87.6%while surpassing standard FL by 3% on average. Regarding the convergence speed, the proposed BEFL reached the target accuracy level in 40 communication cycles for MNIST and 100 rounds for CIFAR-10, while the standard FL was 60 and 140 rounds, respectively, which is approximately 30% faster. While BEFL has a higher communication cost (180 MB for MNIST and 300 MB for CIFAR- 10) than the standard FL (150 MB and 250 MB, respectively), the authors consider it a worthy tradeoff with benefits in terms of security and transparency. The above results under adversary scenarios showed that, as with FL, the proposed BEFL framework was more robust to data poisoning, with the accuracy drop to a mere 1.5% as opposed to 5.2% of FL and a model inversion attack reconstruction accuracy of only 20% as compared to 60% of FL. The above outcomes show that through BEFL, it is possible to support distributed learning while maintaining the security of the fields of study, such as health and finance, as illustrated above.