An Explainable Framework for Mental Health Monitoring Using Lightweight and Privacy-Preserving Federated Facial...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Psychology

Background:

• Facial emotion recognition (FER) systems are crucial for mental health monitoring but face privacy and trust issues due to centralized data.
• Lack of transparency in current FER systems hinders their adoption in sensitive applications like mental health evaluation.

Purpose of the Study:

• To propose a federated, explainability-driven FER framework that ensures privacy and trustworthiness.
• To develop a lightweight Convolutional Neural Network (CNN) for real-time, accurate emotion recognition.
• To evaluate the framework's performance and the reliability of its explanations.

Main Methods:

• Development of a lightweight Convolutional Neural Network (CNN) for FER.
• Implementation of a federated learning approach to preserve data privacy.
• Utilizing Grad-CAM++ for model explainability and evaluating explanations with perturbation-based metrics (IAUC, DAUC, AD, IC, ADA).

Main Results:

• The proposed model achieved average accuracies of 75.5% (centralized) and 74.3% (federated) across multiple datasets (RAF-DB, ExpW, FER2013).
• Demonstrated improved cross-dataset generalization compared to existing methods.
• Quantitative analysis confirmed that model explainability enhances transparency and correlates with improved performance.

Conclusions:

• The federated, explainability-driven FER framework offers a trustworthy and privacy-preserving solution for mental health monitoring.
• The lightweight CNN model provides real-time inference with high accuracy and robust generalization.
• Model explainability is key to building trust and improving the reliability of FER systems in psychological well-being applications.