Federated learning (FL) has emerged as a key approach for training machine learning models without sharing raw data, making it highly relevant for privacy-sensitive applications. However, many existing FL frameworks rely on a central coordinator, which can introduce bottlenecks and single points of failure—especially in heterogeneous and dynamic network environments. To address these challenges, this work introduces NEBULA, a decentralized federated learning platform designed to operate efficiently across diverse and unstable network conditions. Unlike traditional approaches, NEBULA unifies both centralized and peer-to-peer (decentralized) FL paradigms, allowing flexible deployment depending on the network structure and application needs.
A key feature of NEBULA is its network-aware orchestration. The system continuously monitors network conditions—such as latency and connectivity—and uses this information to dynamically adjust communication patterns, node selection, and aggregation strategies. This enables more efficient and resilient training, particularly in environments with fluctuating connectivity or resource constraints. The platform also integrates robustness mechanisms to handle adversarial behavior. Techniques such as robust aggregation algorithms (e.g., Krum), reputation-based filtering, and gradient norm bounding help detect and mitigate malicious updates during training. The demonstration setup highlights NEBULA’s capabilities in real-world-like scenarios, including heterogeneous hardware environments (e.g., containerized clients and Raspberry Pi devices), adversarial attacks, and mobile clients with changing network conditions. A web-based dashboard provides real-time visualization of node participation, communication links, and model performance.
Overall, NEBULA demonstrates how combining decentralization, network awareness, and adaptive mechanisms can improve the resilience and efficiency of federated learning systems, making it a promising approach for future distributed and privacy-preserving AI applications.
DEMO NEBULA Decentralized Federated Learning for Heterogeneous Networks