An efficient and secure adaptive federated learning method based on CKKS for data processing in the Internet of Things

Abstract

Federated learning (FL) provides a new paradigm for solving the security of private data in the internet of things (IoT). However, the huge consumption of computing resources and communication cost makes the FL process inefficient. To solve above problems, this paper proposes an efficient and secure adaptive federated learning method based on CKKS homomorphic encryption (HE) for data processing in IoT. Inspired by dropout, we propose an adaptive inactivation of weights strategy. Through adaptive change of inactivation parameter, part of the weights after reorganization are encrypted and uploaded in each communication. The dual protection of reorganization operation and HE can better protect the weights information. Then, to alleviate the impact of the above methods on the performance of FL, the local data distribution and the change of model accuracy are considered, we propose the federated aggregation method with reward and punishment factor, and the historical information of the local model is employed to design a weights correction strategy. Finally, we use MNIST dataset, fashion-MNIST dataset, GTSRB dataset and CSE-CIC-IDS2018 dataset to design non-independent and identically distributed data scenarios, and a large number of experiments are carried out to verify the effectiveness of the proposed method. Our method not only protects the privacy of weights information, but also reduces the communication cost and the local resource consumption caused by the encryption, which provides a good reference for the follow-up development of FL in IoT.