P2PFL: A Modular Framework for Decentralized Federated Learning

Abstract

[Abstract] Decentralized Federated Learning (DFL) offers a promising paradigm for collaborative machine learning training without centralized coordination, addressing critical limitations of centralized federated learning such as single points of failure, communication bottlenecks, and scalability constraints. However, the adoption of DFL has been hindered by the lack of robust, flexible, and user-friendly frameworks. This paper introduces P2PFL, an open-source framework that enables broader adoption of DFL through a highly modular and extensible architecture. P2PFL decouples communication protocols, learning modules, aggregation algorithms, and orchestration workflows into independent, interchangeable components. This modularity enables researchers and practitioners to efficiently prototype and deploy federated learning systems tailored to their specific requirements. A key contribution is the introduction of optimized gossip-based protocols with incremental partial aggregation whose equivalence to centralized federated learning is formally proven, with communication savings that scale with network size, achieving up to 88% reduction in model data transfer at 128 nodes. We demonstrate P2PFL’s effectiveness through functional validation on standard benchmarks (MNIST and CIFAR-10) under various network topologies and data heterogeneity settings, a scalability analysis from 8 to 128 nodes quantifying communication overhead and round duration across topologies, and a real-world deployment on the CASA IoT dataset across 10 logical nodes over geographically distributed machines. The results reveal that topology selection directly affects control overhead, while the model data itself remains independent of the selected topology, confirming that the benefits of partial aggregation scale with network size. By providing flexible deployment interfaces and integrated monitoring capabilities, P2PFL bridges the gap between DFL theory and practice, accelerating research and facilitating the deployment of privacy-preserving, decentralized machine learning systems.