The problem
Each site has few examples of failures, so a model trained at one site is weak; sites will not share raw data (confidentiality, size). Federated learning lets them train one detector together.
Who needs it
Facility management, manufacturing (predictive maintenance), energy companies.
Idea of the approach
- Public time-series datasets (e.g. NASA bearing / SKAB / building energy data) split into several "sites".
- Train an autoencoder or LSTM anomaly detector locally, then federated (FedAvg / FedProx).
- Compare local-only, centralized and federated detection quality; measure communication.
- Dashboard showing alerts per site.
What you will build and deliver
- Federated detector, dashboard and evaluation.
- Paper.
Skills needed
Python, PyTorch or Keras, time-series basics.
Research contribution & where to publish
Contribution: evaluation of federated anomaly detection under unequal site sizes and rare anomalies. Venues: IEEE IoT Journal, IEEE Transactions on Industrial Informatics, IEEE ICC.
Key references — start reading here
- T. D. Nguyen, S. Marchal, M. Miettinen, H. Fereidooni, N. Asokan, A.-R. Sadeghi. DÏoT: A Federated Self-learning Anomaly Detection System for IoT. IEEE ICDCS, 2019.
- Y. Liu, S. Garg, J. Nie, Y. Zhang, Z. Xiong, J. Kang, M. S. Hossain. Deep Anomaly Detection for Time-Series Data in Industrial IoT: A Communication-Efficient On-Device Federated Learning Approach. IEEE Internet of Things Journal, 2021.
- P. Kairouz, H. B. McMahan, et al.. Advances and Open Problems in Federated Learning. Foundations and Trends in Machine Learning, 2021.
- T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, V. Smith. Federated Optimization in Heterogeneous Networks. MLSys, 2020.
Interested?
Send the topic title and the courses you have taken.
Contact Dr. Hamza Djigal