Platform for Artificial Intelligence usage on the Internet of Things through the sharing of local micro models between device communities

Funding institution: European Commission – Ministerio de Ciencia, Innovación y Universidades – Agencia Estatal de Investigación
Duration: 1 September 2023 – 31 August 2027

CoIoTIA investigates the design, development and validation of an ML analytics platform for IoT devices capable of improving adaptability in dynamic and heterogeneous environments by leveraging local/context-dependent learning and the Community of Interest (CoI) concept (time and location). The goal is to facilitate the publication, discovery, reuse and reproducibility of ML models for IoT, while enabling low-latency inference at the edge/fog and increasing trust and traceability through blockchain.

The design is based on three basic pillars:

– Internet of Things and Edge/Fog computing for low-latency AIoT. CoIoTIA follows a distributed architecture where edge devices (smart and basic) operate close to data sources, while fog nodes act as gateways that manage IoT devices under their geographical influence. This enables distributing computation depending on available resources, reducing latency and avoiding unnecessary cloud dependency (cloud is reserved for auxiliary tasks).

– Machine Learning with local micro-models and CoI-based sharing. CoIoTIA proposes generating micro ML models using local learning / decomposition strategies, making models more robust to changing contexts (“Change Anything Changes Everything”). These micro-models are shared and consumed within device communities (CoI) defined by purpose plus temporal and geospatial context. A decentralised repository in fog nodes supports model publication and discovery, using metadata and a recommender driven by purpose, data characteristics, time and location.

– Blockchain and smart contracts for trust, traceability and Proof of Inference. Fog elements form a (private/hybrid) blockchain network to link models with the datasets/code/parameters used for training and versioning, improving transparency and reproducibility. Based on smart contracts and geospatial areas (e.g., geohashes), CoIoTIA introduces a Proof of Inference mechanism to validate that an inference was executed using a specific model under certain location/time conditions; otherwise, the interaction can be invalidated.

CoIoTIA supports two inference strategies:

• Decentralised inference (TinyML + OTA): models are deployed to smart edge devices (e.g., microcontrollers / mobile devices) to run inference locally.
• Semi-distributed inference (FaaS at fog): basic devices call functions hosted at fog nodes to perform inference when edge resources are limited.

Demo

Funding

IP and technical contact

Joaquín Huerta (huerta@uji.es)

Sergi Trilles (strilles@uji.es)