DP-FL on Human Activity Recognition
Differentially Private Federated Learning for HAR
This project applies Differentially Private Federated Learning (DP-FL) to Human Activity Recognition (HAR) using wearable sensor data, exploring the privacy-utility tradeoff in a cross-device FL setting.
Research Questions
- How does DP noise (Gaussian mechanism, ε-δ privacy) impact HAR model accuracy across heterogeneous clients?
- Can adaptive clipping strategies reduce accuracy loss while preserving privacy guarantees?
Setting
- Dataset: UCI HAR (accelerometer + gyroscope, 6 activities, 30 subjects)
- Federated setup: heterogeneous data partitions per client, FedAvg with DP-SGD