2024·ML researcher
Fault-Tolerant Federated Learning
Federated learning approach for sensor systems that handles missing data and sensor dropouts using imputation and LLM-based filling, evaluated on the MHEALTH dataset.
Stack
Python,Federated Learning,PyTorch,TensorFlow,DistilGPT-2
Timeline
2024
Impact
- Handles missing data & sensor dropouts gracefully
- Imputation + LLM-based value filling
- Evaluated on MHEALTH activity data
Context & Goals
Real-world sensor networks are messy — devices drop offline and streams go missing. This research explores a federated learning approach that stays robust when participating sensors deliver incomplete data, without centralizing raw signals.
Approach
- Fault tolerance: the training procedure tolerates sensor dropouts and partial observations across federated clients.
- Missing-data handling: combines classical imputation with an LLM-based filling strategy (DistilGPT-2) to reconstruct plausible values for absent readings.
- Evaluation: benchmarked on the MHEALTH dataset using both PyTorch and TensorFlow pipelines for activity recognition.
Highlights
- Compared imputation strategies head-to-head to quantify the accuracy cost of dropout.
- Kept data decentralized — clients contribute model updates, not raw sensor traces.
What I'd do next
- Add differential-privacy guarantees on client updates.
- Test on additional multi-sensor wearable datasets.