Elevating IoT Efficiency: Fusing Multimodal Data with Federated Learning Algorithms

The increasing use of smart elements in some industries generates high amounts of complex data, which challenges data privacy and scalability processing. This work presents the Multimodal Federated Learning (MFL) framework designed to leverage multimodal input data in a decentralized manner while preserving users' privacy. We assessed the suggested architecture using three authentic IoT datasets. The three main categories that it covered include Healthcare, Smart City, and Autonomous Driving. The MFL framework displayed exceptional performance, achieving an overall accuracy of 90% over all the datasets, with specific accuracies of 92 % in Healthcare, 88% in Smart City, and 90% in Autonomous Driving, a performance that is 12 % above the normal Federated Learning models. The results have been quite impressive, and the precision and recall scores are high: precision 91 %, 87%, and 89%, and recall 90%, 86%, and 88% for the respective datasets. The F1-Score achieved by the framework is as follows: Healthcare modality scored 90.5%, Smart City scored 86.5%, while the Autonomous Driving modality score was 88.5%, which shows the framework's effectiveness in many modalities. The MFL framework successfully reduced the COM overhead by 30% for Healthcare, which needed only 500 MB; Smart City needed 400 MB, while Autonomous Driving needed 450 MB in comparison to the 700 MB, 600 MB, and 650 MB taken by other centralized learning approaches. The study demonstrated that the MFL can integrate different IoT data into one set without violating data privacy and at an affordable communication cost, making it possible to provide security measures for implementing scalable and complex IoT environments using the machine learning algorithm.