FedEach: Federated Learning with Evaluator-Based Incentive Mechanism for Human Activity Recognition

Area of Science:

• Artificial Intelligence
• Machine Learning
• Decentralized Systems

Background:

• Federated learning (FL) enables collaborative model training without data sharing.
• Byzantine clients pose a significant threat to FL integrity and performance.
• Existing FL defenses against malicious clients often require server-labeled data or lack incentives for reliable clients.

Purpose of the Study:

• To introduce a robust federated learning framework, FedEach, that effectively handles Byzantine clients.
• To develop an evaluator-based incentive mechanism that identifies and utilizes reliable clients without relying on server-labeled data.
• To ensure computational efficiency for practical FL applications.

Main Methods:

• A novel federated learning framework (FedEach) with an evaluator-based incentive mechanism.
• Server-based selection of evaluators and participants using performance criteria (test score, reputation).
• Evaluator-driven assessment of client reliability (malicious vs. reliable).
• Aggregation of models exclusively from identified reliable participants and evaluators.
• Client contribution calculation for fair recognition and penalization.

Main Results:

• FedEach demonstrates significant robustness against malicious clients, particularly in high-maliciousness environments.
• The framework effectively aggregates updates from reliable clients, enhancing global model performance.
• FedEach maintains computational efficiency, suitable for real-time FL applications.
• Empirical validation on human activity recognition (HAR) datasets confirms the framework's effectiveness.

Conclusions:

• FedEach provides a robust and incentive-aligned solution for federated learning in the presence of Byzantine clients.
• The evaluator-based mechanism enhances security and reliability without increasing time complexity or requiring public validation datasets.
• FedEach is a practical and efficient framework for real-world FL applications like sensor-based HAR.