Preliminary Analysis of Federated Learning in Agrifood: From MNIST to Real-World IoT Applications in a Heterogeneous Environment

Federated Learning (FL) is a promising approach for decentralized Machine Learning (ML), enabling collaborative model training while keeping data localized. However, the integration of heterogeneous devices, particularly resource-constrained Internet of Things (IoT) clients like microcontrollers, poses significant challenges. A simulation of an FL framework involving IoT clients is presented in this paper to train a Neural Network (NN) model optimized for microcontrollers. In the first round, an STM32F722ZE microcontroller is used as one of the clients, alongside standard clients, to simulate device heterogeneity. Initially, FL is demonstrated on the MNIST dataset as a proof of concept, before the framework is extended to real-world applications in the agrifood sector, using IoT devices in a heterogeneous environment. The simulation includes real-world constraints such as training delays, energy consumption, and system heterogeneity. The system is evaluated based on parameters like energy consumption, inference time, and model performance. The results provide insights into the trade-offs between resource usage and model accuracy, demonstrating the feasibility of implementing FL in agrifood applications with heterogeneous IoT devices.