Federated Learning for a Dynamic Edge: A Modular and Resilient Approach.

Area of Science:

• Computer Science
• Artificial Intelligence
• Distributed Systems

Background:

• Growing demand for distributed machine learning (ML) in edge environments.
• Challenges in fault tolerance, elasticity, and communication efficiency for Federated Learning (FL).
• Need for adaptable frameworks in dynamic, resource-constrained settings.

Purpose of the Study:

• Propose a novel modular and resilient FL framework.
• Address challenges of fault tolerance and communication efficiency in edge ML.
• Enhance the adaptability of FL systems for diverse environments.

Main Methods:

• Developed a modular FL framework with decoupled core functionalities.
• Integrated various communication protocols (Zenoh, MQTT, Kafka) and FL paradigms.
• Simulated probabilistic worker failures to test resilience and performance.

Main Results:

• Framework demonstrated flexibility in integrating diverse communication protocols and FL approaches.
• Protocol selection significantly impacts performance, especially in high-volume communication.
• Zenoh showed the lowest overhead and highest efficiency among tested protocols.
• Model training achieved convergence with a Matthews Correlation Coefficient (MCC) of 0.9453 despite simulated failures.

Conclusions:

• The proposed modular framework enhances resilience and efficiency in Federated Learning.
• Optimized communication protocol selection is crucial for performance in edge environments.
• The framework effectively maintains FL operations under disruptive conditions, ensuring model convergence.