An Adaptive Federated Learning Approach for Efficiency and Privacy Preservation of Dynamic Network of IoT

For training machine learning models in networked systems like the Internet of Things (IoT), the approach of federated learning (FL) has become intriguing. FL makes it possible to learn models collaboratively across many devices by providing efficiency in training and testing. The standard FL technique, however, has significant challenges due to the dynamic nature of IoT networks. In this research paper, we propose an adaptive federated learning (AFL) approach that is capable of accommodating the dynamism of IoT networks along with privacy-preserving. AFL utilizes dynamic model training by using data streaming at edge devices of IoT networks. The proposed algorithm efficiently trains the AFL model at the cloud data store by receiving the encrypted data with homomorphic cryptography from an edge device. It also preserves the privacy of data in transit. The AFL model gets trained by the encrypted data only by maintaining the bucket of data streaming. This paper demonstrates the effectiveness of AFL on a real-world IoT data-set, where it achieves higher accuracy with data privacy.