Towards a scalable and robust federated learning system

Abstract

This project enhances a prior peer-to-peer federated learning (P2PFL) prototype by addressing key challenges in scalability, transparency, and robustness. On the development side, the native Android application is extended with an improved asset management system for handling models and datasets, a scalable HTTP-based model transfer protocol, and a centralised web-based registry to facilitate model discovery. To improve traceability and trust, a graph-based model lineage visualisation system is implemented using Neo4j. From a research standpoint, the project examines three types of model poisoning attacks—random label, flip label, and backdoor—to determine the minimum effective poisoning threshold. These attacks are evaluated against three defence strategies: performance-based (RONI), weight-based (Dim-Krum), and representation-based (activation clustering). Experimental findings reveal that no single method offers complete protection, but a hybrid strategy combining performance and representation-based defences provides the most resilient safeguard. Collectively, these contributions advance the scalability and security of decentralised federated learning systems in real-world applications.