HEND-FL: Accurate Federated Learning Using Homomorphic Encryption and a New Distributed Protocol
Dehua Zhou, Zexiao Chen, Di Wu, Yingwei Yu, Qingqing Gan, Botong Xu · 2022 IEEE Intl Conf on Dependable, Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech) · 2022
Federated learning (FL) is the latest development of distributed machine learning(ML), in which data is acquired and processed locally on the client, and then updated ML parameters are transmitted to a central server for aggregation. However, FL also brings some challenges, since it may reveal private information by analyzing upload parameters (such as weights trained in deep neural networks) from the client. To tackle the problem, this paper introduces a decentralized system, HEND-FL, which relies on a group of computing nodes to achieve aggregate calculations of parameters. In order to ensure the data confidentiality and the privacy of the data provider, this proposed system combines an interactive protocol and homomorphic encryption, and uses the Chinese Remainder Theorem to optimize the speed of decryption. Based on the discrete logarithm problem of elliptic curves, the scheme can be proved to be secure under the defined threat model. We evaluate the effectiveness of our plan and compare it with existing related work on the MNIST dataset. The results show that our scheme has almost the same accuracy in the plaintext and encrypted state. Compared with other encryption schemes, due to our smaller ciphertext size and the assistance of multiple computing nodes, the burden on the central server is significantly reduced.