Phoenix: A Modular Framework for Security Evaluation in Decentralized Federated Learning

This paper presents Fenics, a comprehensive open-source frame- work for evaluating Decentralized Federated Learning (DFL) net- works in the presence of malicious users. The framework implements core DFL functionalities, including data distribution strategies, dynamic node participation, model aggregation, and attack model simulation, to assess DFL network resilience under various attack scenarios. Phoenix supports multiple communication proto- cols and custom network topologies, providing potential avenues to investigate the impact of malicious nodes based on their placement within the network. To the best of our knowledge, Phoenix is the first fully open-source and modular framework of its kind, allowing diverse topologies and attack scenarios to be easily incorporated. We demonstrate the framework’s capabilities through experimental evaluation using the FashionMNIST dataset for poisoning and delay attacks. Experimental results reveal critical insights into DFL network vulnerabilities. For example, poisoning attacks exhibit topology-dependent impacts, with accuracy dropping by over 55% in certain scenarios, leading to derailed convergence. Conversely, delay attacks significantly degrade computational efficiency by forcing nodes to idle during malicious delays. The framework’s modular architecture, comprehensive visualization capabilities, and scalability make it a lightweight and accessible tool for insightful analysis of DFL network security and further research.