Achieving High Efficiency and Privacy in IoT With Federated Learning: A Verifiable Horizontal Approach

Abstract

The exponential growth of the Internet of Things (IoT) domain has raised the question of the immediate need for complex and privacy-sensitive data processing methodologies. Federated Learning (FL) makes learning decentralized and supports individual privacy; hence, it proves to be an optimistic approach. However, according to recent research, federated learning systems are vulnerable to attacks that may breach client privacy. Traditional federated learning models face issues such as verification of aggregated results and high computational and communication demands, making them less practical in large-scale IoT implementations. We propose a new federated learning framework: Verifiable Horizontal Federated Learning (VHFL). The main goals of VHFL include data privacy, reducing computational and communication-side overheads, and ensuring the accuracy of aggregated results. To improve the data's privacy, VHFL uses single-mask encryption techniques together with group-key techniques. Another step will be to further integrate Latin Squares Design to reduce client-side computational and communication overheads. The system introduces a new verification scheme that generates Hamiltonian graphs from LSD to ensure that the VHFL aggregation result is correct. We tested the proposed system on several datasets, including MIMIC-III and HAR and compared to traditional federated learning models. More importantly, our proof-of-concept provided evidence that VHFL is robust in ensuring high efficiency and strong privacy within IoT environments. These experiments confirm that VHFL is efficient in handling the most demanding issues posed to federated learning and can be viable for both secure and efficient IoT applications.