Local Intelligence for AIoT through Federated Community Learning (LIA) 

Funding institution: Convocatoria 2024 – «Incentivación de la consolidación investigadora» – Ministerio de Ciencia, Innovación y Universidades
Duration: 1 April 2025 – 31 March 2027

LIA investigates the design, development and validation of a machine-learning platform for IoT (AIoT) that increases adaptability, privacy and scalability by combining local learning with Federated Learning (including Federated Transfer Learning) across Communities of Interest (CoI) defined by purpose, time and geospatial context. The goal is to make advanced ML usable in heterogeneous IoT ecosystems without requiring deep data-science expertise, while keeping data local and enabling low-latency intelligence at the edge/fog.

The design is based on three basic pillars:

– Edge/Fog AIoT for low latency and operational autonomy. LIA follows a two-layer architecture (edge + fog, with cloud only for auxiliary tasks) where fog nodes act as gateways and coordinators for geographically distributed devices, enabling scalable ML execution close to where data is generated.

– Federated Community Learning for adaptable models. LIA leverages FL/FTL to generate context-adapted local models (TinyML-friendly micro-models) within each CoI, addressing dynamic environments (“CACE”) and device heterogeneity. Devices can consume and further refine models (fine-tuning) suited to their current location and time context.

– Decentralised model repository + trustworthy inference support. LIA proposes a distributed repository deployed across fog nodes to publish, discover and reuse ML models using rich metadata (purpose, data features, temporality, location) and a recommender. The platform supports inference via (a) decentralised on-device inference (TinyML + OTA) and (b) semi-distributed inference using FaaS at the fog for constrained devices, incorporating explainability (XAI) to improve transparency and reuse.

Website: https://lia.uji.es/

This project code is CNS2024-154145

IP and technical contact

Sergi Trilles (strilles@uji.es)

Funding