Efficient Dropout-resilient and Verifiable Federated learning scheme for AIoT healthcare system

Abstract

Federated learning (FL) with healthcare AIoT data from multiple healthcare institutions can enhance device intelligence and further protect people’s health. The preservation of confidentiality in healthcare data has emerged as a central concern in healthcare FL systems. Under the premise of emphasizing privacy protection, it is particularly inefficient to maintain system robustness after a participant drops out. Furthermore, it proves more complex to confirm whether the aggregation server faithfully executes the aggregation task, especially given that the aggregation server may collude with corrupt users to intentionally return carefully crafted aggregation results. At present, few works focus on the two issues concurrently. To address this challenge, an efficient dropout-resilient and verifiable FL scheme (EDV-FL for short) is proposed in this paper. Our scheme addresses the issue of dropped users rejoining in the future, while reducing both communication and computation overhead. Moreover, we ensure that even if the server colludes with corrupt users to forge the aggregation result, users can still detect the correctness of the aggregation result. We theoretically demonstrate the effectiveness of EDV-FL and reproduce the scheme using Convolutional Neural Network (CNN) models on the MNIST, CIFAR-10, and Fashion-MNIST datasets. Theoretical proofs and experimental analyses show that our EDV-FL is an efficient, dropout-resistant, and collusion-resistant, verifiable FL scheme.