Project Overview
This research project focused on developing a privacy-preserving machine learning system to recognize mental stress levels in students using data collected from their smartphones. The challenge was to analyze behavioral patterns without compromising user privacy.
Technical Approach
Data Collection & Processing
- Collected multimodal sensing data from smartphone sensors in real-world conditions
- Processed time-series data using advanced feature extraction techniques
- Handled high-frequency data streams with proper preprocessing pipelines
Machine Learning Pipeline
Models Evaluated:
- K-Nearest Neighbors (KNN)
- Random Forest
- XGBoost
- Neural Networks with LSTM and Attention mechanisms
Privacy-Preserving Innovation
The key innovation was implementing Federated Learning to train models without centralizing sensitive user data:
- Models trained locally on each device
- Only model updates shared with central server
- Achieved personalization while preserving privacy
- Reduced data transmission requirements
Technologies Used
- Languages: Python
- ML Frameworks: PyTorch, Scikit-learn
- Data Processing: Pandas, NumPy
- Techniques: Time series analysis, LSTM, Attention mechanisms, Federated Learning
Key Outcomes
- Successfully classified mental pressure levels from smartphone sensor data
- Demonstrated feasibility of federated learning for health monitoring
- Maintained user privacy throughout the ML pipeline
- Thesis research contributing to the field of affective computing