Comparative Analysis between Individual, Centralized, and Federated Learning for Smartwatch Based Stress Detection.

Area of Science:

• Wearable technology
• Machine learning applications
• Biomedical signal processing

Background:

• Machine learning models for stress detection require substantial data, often necessitating centralized data collection.
• Centralized data aggregation raises significant user privacy concerns due to the sensitive nature of personal health information.
• Federated learning offers a decentralized approach to train models without compromising user data privacy.

Purpose of the Study:

• To implement and evaluate a federated learning approach for stress detection using smartwatch sensor data.
• To compare the performance of federated learning against individual and centralized learning strategies for stress detection.
• To analyze the trade-offs between privacy preservation and model accuracy in stress detection.

Main Methods:

• The study utilized the WESAD dataset, which includes multi-modal sensor data from smartwatches.
• Logistic Regression was employed as the classification algorithm for stress detection.
• A comparative analysis was performed between individual learning, centralized learning, and federated learning paradigms.

Main Results:

• Individual learning demonstrated superior performance, achieving an average accuracy of 0.9998 and an F1-measure of 0.9996.
• Federated learning, while enabling decentralized training, did not attain the accuracy levels of individual or centralized learning.
• The results indicate a performance gap between federated learning and traditional centralized/individual methods for this specific stress detection task.

Conclusions:

• Individual learning provides the highest accuracy for stress detection with the WESAD dataset.
• Federated learning presents a privacy-preserving alternative but currently underperforms compared to centralized and individual methods in terms of accuracy.
• Further research is needed to optimize federated learning algorithms for improved performance in wearable-based health monitoring.