Enhancing Privacy in Distributed AI Systems Through Decentralized Federated Learning

In this paper, we provide a decentralized federated learning (FL) framework that targets reducing privacy breaches as well as enhancing the efficiency of communication in distributed AI systems. Our work is along a different approach from conventional federated learning methodologies that rely on various forms of secure aggregation and centralized optimization strategies compared to model. We address major challenges, including data distributions that are non IID, bringing privacy vulnerabilities with increased overheads in terms of communication. Through mathematical modeling along with extensive experimentation on benchmarks like MNIST and CIFAR-10, we validate that the proposed framework gives an improvement of several times in privacy preservation, reduces communications overhead by up to 30%, and achieves models competitive in accuracy. Hence, these results present this decentralized federated learning towards being a scalable and more privacy-aware solution for distributed systems in AI.