Abstract

As federated learning gains adoption in sensitive domains, ensuring trust and privacy remains a major challenge. This paper presents ChainTrust-FL, a novel framework that integrates Zero Trust Architecture (ZTA), Blockchain, and Differential Privacy to enhance the security, privacy, and performance of machine learning models. We evaluate ChainTrust-FL against three baseline models: Vanilla Federated Learning (FL), FL with Blockchain, and FL with Differential Privacy using the CIFAR-10 dataset. Our experiments show that ChainTrust-FL achieves an accuracy of 93.7% after 20 communication rounds, outperforming the baselines (FL with Blockchain: 89.2%, FL with Differential Privacy: 88.4%, and Vanilla FL: 82.1%). The model also exhibits superior resilience to model poisoning attacks, maintaining an accuracy of 90.2% under adversarial conditions. Moreover, ChainTrust-FL effectively manages the privacy-accuracy tradeoff, maintaining high accuracy even with strict privacy settings, and introduces minimal latency (0.15 seconds) due to its blockchain integration. These results demonstrate that ChainTrust-FL is an effective and efficient solution for privacy-preserving federated learning, providing a secure, robust, and privacy-aware machine learning framework suitable for real-world applications.