The problem
Activity data from wearables is personal health data: it should stay on the device. Federated learning trains a shared model by exchanging model updates only, but these updates are large for small devices and users behave differently (non-IID data).
Who needs it
Health and fitness device makers, elderly-care and fall-detection services, insurance wellness programmes.
Idea of the approach
- Use public HAR datasets (UCI-HAR, WISDM, PAMAP2) split per user.
- Implement FedAvg with Flower or a simple PyTorch loop.
- Reduce communication with low-rank updates (as in ProLoRA / FedQR) and quantization; add personalization layers.
- Measure accuracy per user, bytes sent and device time on a Raspberry Pi.
What you will build and deliver
- The federated training system and an on-device demo.
- Results: accuracy vs communication.
- Paper.
Skills needed
Python, PyTorch; signal-processing basics helpful.
Research contribution & where to publish
Contribution: combining low-rank compression with personalization for heterogeneous wearable users. Venues: IEEE MSN, ACM UbiComp/IMWUT workshops, IEEE IoT Journal, Pervasive and Mobile Computing.
Key references — start reading here
- B. McMahan, E. Moore, D. Ramage, S. Hampson, B. Agüera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. AISTATS, 2017.
- X. Ouyang, Z. Xie, J. Zhou, J. Huang, G. Xing. ClusterFL: A Similarity-Aware Federated Learning System for Human Activity Recognition. ACM MobiSys, 2021.
- H. Djigal et al.. ProLoRA: Resource-Efficient Personalized Federated Learning for Sensor Based Human Activity Recognition. , 2024.
- H. Djigal et al.. FedQR: Communication-Efficient Federated Learning via QR Factorization. Communications in Computer and Information Science, 2025.
Interested?
Send the topic title and the courses you have taken.
Contact Dr. Hamza Djigal