LIA-FL

LIA-FL is a containerised platform for deploying and orchestrating federated learning workflows on distributed IoT devices. Each node trains locally and only model parameters are exchanged with the aggregation server (raw data never leaves the device).

The problem

Federated learning usually trains a single global model shared by every device. In AIoT environments, where nodes can differ a lot in behaviour, location, or hardware, that one-size-fits-all model tends to end up mediocre for everyone.

What it does

• Group-based federation. Instead of a single global model, the platform trains a specialised model per cluster of contextually similar nodes.
• Federated nodes and aggregation server. Nodes train locally with TensorFlow/Keras and exchange only model weights with a central aggregator over MQTT (Mosquitto), never raw data.
• Weighted FedAvg. Aggregation combines each cluster’s member weights, weighted by local dataset size.
• Reliable weight transfer. Model weights travel in chunks over MQTT with SHA-256 integrity checks.
• Local persistence and tracking. Each node keeps its state in local SQLite storage, with structured JSONL logging and optional MLflow experiment tracking.
• Container-first deployment. Every component (nodes, aggregator, broker) ships as a Docker container.

Validation

Validated on temperature forecasting with 83 weather stations from the AVAMET network in Castellón, comparing centralised, global federated, and cluster-based federated strategies using GRU networks.

Publications

This work is published in SoftwareX and was presented at FLICS 2026:

• Ruiz-Fas, À., Goterris, J., Díaz-Rivero, A., Sáenz-Martínez, C., & Trilles, S. (2026). LIA-FL: A containerised federated learning platform for AIoT time-series forecasting. SoftwareX. doi.org/10.1016/j.softx.2026.103000
• Goterris, J., Díaz-Rivero, A., Ruiz-Fas, À., Sáenz-Martínez, C., & Trilles, S. (2026). A Preliminary CoI-Based Federated Learning Platform for Artificial Intelligence of Things. Proceedings of FLICS 2026, pp. 464–469. doi.org/10.1109/FLICS70075.2026.11621955

Status

Research project developed during a research collaboration scholarship with the GEOTEC research group at Universitat Jaume I. Source code released under the MIT license.