Privacy-preserving Speech-based Depression Diagnosis via Federated Learning

Area of Science:

• Artificial Intelligence
• Computational Psychiatry
• Machine Learning

Background:

• Depression diagnosis is crucial, and speech analysis shows promise.
• Centralized deep learning models for speech analysis risk data breaches and privacy concerns.
• Federated learning (FL) offers a decentralized alternative for training models.

Purpose of the Study:

• To demonstrate the feasibility of privacy-preserving depression diagnosis using federated learning and speech analysis.
• To evaluate the performance and robustness of federated learning models for depression detection.
• To address privacy concerns associated with centralized voice data collection.

Main Methods:

• Implemented a federated learning framework for training speech-based depression diagnosis models.
• Integrated privacy-enhancing techniques including norm bounding, differential privacy, and secure aggregation.
• Trained and evaluated models on the DAIC-WOZ dataset under various federated learning settings.

Main Results:

• The federated learning model achieved high performance comparable to centralized approaches.
• Privacy-preserving methods did not significantly compromise the diagnostic utility of the models.
• The system demonstrated robustness against potential attacks in federated settings.

Conclusions:

• Federated learning is a viable and effective method for privacy-preserving depression diagnosis via speech analysis.
• This approach mitigates privacy risks associated with centralized data collection.
• The developed system offers a secure and accurate tool for mental health monitoring.