Federated Learning-based Techniques for Healthcare IoV: Challenges and Future Directions -A Knowledge Review

Federated Learning (FL) is a cutting-edge method for creating Machine Learning (ML) models using networked data sets. It supports advanced Artificial Intelligence (AI), keeps information neighborhoods, protects personal information, and allows for ownership. FL's features make it a good candidate for IoT networks due to its inherent distributed infrastructure. On the other hand, FL provides several challenges, the most significant of which is training on IoT devices with mostly heterogeneous data sources. The diversity of devices and algorithms in complex Internet of Vehicles (IoV) networks significantly impacts the FL training process and renders traditional FL unsuitable for direct deployment. While numerous recent publications claim to lessen the negative impacts of variety in FL networks, the efficacy of these proposed ways has never been tested and assessed. In this study, we examine the impact of heterogeneity in FL in depth and provide an overview of the actual challenges imposed by the system and statistical heterogeneity. We researched on cutting-edge algorithms, emphasizing their practical application in IoT networks. We also compared the popular FL algorithms on a dynamic IoV network. Our study shows that existing methods fail to alleviate the issue appropriately, emphasizing the significance of considering system and statistical diversity while developing FL systems. To properly investigate the consequences of the network heterogeneity, we tested 10%, 30%, 50%, 70%, and 90% stragglers. These conditions serve as a training ground, allowing us to see how various degrees of network heterogeneity impact a technique's efficiency. The test-bed results show that the Scafold technique outperforms existing techniques in terms of training time reduced to 18% and accuracy increased to 21%