FedCoRE: Effective Federated Learning for constrained RESTful environments in the Artificial Intelligence of Things
Résumé
Federated Learning (FL) empowers Internet-of-Things (IoT) devices to train intelligent models without sharing sensitive data, facilitating the transition to an Artificial Intelligence of Things (AIoT) ecosystem. However, FL demands significant storage, computation, and communication resources, which often exceed the capabilities of resource-constrained IoT devices. In this work, we introduce FedCoRE, an effective and practical FL architecture tailored for IoT environments. FedCoRE leverages standards for constrained RESTful environments, such as the Constrained Application Protocol (CoAP), to optimize communication and applies model quantization to address computation and storage limitations. FedCoRE has been implemented on resource-constrained IoT devices with 256 KB of RAM and evaluated on a human activity recognition task using a deep neural network. Extensive evaluations conducted in a real-world IoT environment, comprising 10 Thunderboard Sense 2 nodes, demonstrate the feasibility and effectiveness of our proposal. Notably, compared to FL, FedCoRE achieves up to a 60% reduction in communication cost, while maintaining model accuracy and requiring only approximately 75 KB of RAM and 438 KB of ROM.