MASTER THESIS ASSIGNMENT: WorkshopFL - Privacy-Preserving HAR on Earables via Personalized Federated Learning
INTRODUCTION
Activity recognition (HAR) in workplaces often requires sensitive data. Personalized Federated Learning (PFL) can train models across users/sites without centralizing raw data.
OBJECTIVES
- Implement a PFL pipeline (e.g., Flower) for earable-based HAR with different users.
- Compare FedAvg vs. personalized schemes (FedPer, FedProx, FedYogi..).
- Quantify privacy–utility–communication trade-offs.
PROJECT DESCRIPTION
1. Literature Review: Federated/PFL for time-series; Earable HAR.
2. Dataset Setup: Use existing earable datasets and/or small-scale new recordings for non-daily actions.
3. PFL Implementation: Train baseline centralized model; then federated variants with personalization and regularization.
4. Evaluation: Macro-F1, personalization gains, communication rounds, robustness to client dropouts.
5. Extensions: Differential privacy noise; secure aggregation; on-device inference profiling.
PRE-REQUISITES
Python, PyTorch/TF, ML; familiarity with federated learning frameworks is a plus.
WORK
30% Theory, 50% Simulations, 20% Writing
CONTACT
Egemen İşgüder (egemen.isguder@utwente.nl)
Özlem Durmaz İncel (ozlem.durmaz@utwente.nl)
Rob H. Bemthuis (r.h.bemthuis@utwente.nl)