FedKDC: Consensus-Driven Knowledge Distillation for Personalized Federated Learning in EEG-Based Emotion Recognition

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neuroscience

Background:

• Federated learning (FL) enables secure, decentralized training for electroencephalogram (EEG)-based emotion recognition.
• Traditional FL struggles with model and data heterogeneity in healthcare, impacting convergence and performance.
• Heterogeneity arises from varied computational resources and diverse EEG data across institutions.

Purpose of the Study:

• To propose FedKDC, a novel FL framework addressing heterogeneity challenges in EEG emotion recognition.
• To enhance model convergence speed and reduce performance degradation caused by data and model variations.
• To improve the security and efficiency of collaborative EEG data analysis in smart healthcare.

Main Methods:

• Developed FedKDC, a framework integrating clustered knowledge distillation (CKD) with consensus-based distributed learning.
• Implemented intraclass distillation for faster convergence and interclass distillation to mitigate heterogeneity.
• Introduced DriftGuard mechanism to combat client drift and an entropy reducer for aggregated knowledge.

Main Results:

• FedKDC demonstrated effectiveness on SEED, SEED-IV, SEED-FRA, and SEED-GER datasets under heterogeneous conditions.
• Achieved a maximum average accuracy of 85.2% in emotion recognition, outperforming existing FL frameworks.
• Showcased superior convergence efficiency, characterized by faster and more stable convergence rates.

Conclusions:

• FedKDC effectively addresses model and data heterogeneity in FL for EEG emotion recognition.
• The proposed framework offers enhanced accuracy, faster convergence, and improved stability.
• FedKDC represents a significant advancement for secure and efficient collaborative emotion recognition in smart healthcare.