A Homomorphic MAC-based verifiable secure aggregation for federated learning in cloud–edge AIoT

Abstract

The cloud–edge collaborative Artificial Intelligence of Things (AIoT) architecture addresses challenges in managing vast data storage, intelligent information processing, device interconnectivity within the Internet of Things. For its security risks and data privacy, federated learning emerges as a promising solution for ensuring data privacy in AIoT. However, susceptibility to malicious attacks during data transmission poses a significant challenge and a semi-trusted server may deviate from the specified protocol leading to inaccurate aggregation parameters returned to clients. Our proposed solution introduces a federated learning integrity verification scheme based on homomorphic Message Authentication Code (MAC) within a cloud–edge collaborative AIoT architecture. Homomorphic MAC ensures secure aggregation and integrity verification, even when distinct clients possess different keys, emphasizing integrity verification by edge node, contributes to reduced client computing costs. Further verifying of the aggregated parameters by users prevents untrusted transmission from edge node. Leveraging data integrity verification proves effective in mitigating challenges associated with parameter security, especially in scenarios involving inaccurate aggregation of local model parameters within federated learning. Our solution is free bilinear pairing, resulting in a significant reduction in computational overhead. We evaluate accuracy on the MNIST dataset through comparison with the FedAVG plaintext scheme, showing that our approach ensures parameter integrity while maintaining model performance, numerical simulations also confirm its efficiency.