Federated Learning based on Homomorphic Encryption and Digital Signatures

Abstract

Federated learning aims to train a centralized federated model using decentralized data sources and to ensure the security and privacy of user data during system training process. Federal learning still faces four challenges: high communication costs, system heterogeneity, data heterogeneity, and data security, The primary attacks facing federal learning include confidentiality, integrity, and usability attacks. Attackers mainly target confidential attacks, while malicious attackers target integrity attacks and usability attacks. Homomorphism encryption refers to the encryption algorithm that satisfies the nature of the homomorphism operation, that is, to the ciphertext of the data after the homomorphic encryption. This article first introduces the federated learning and its possible attacks in the process of modeling behavior, and then introduces how to use homomorphic encryption and digital signature algorithm to prevent the attack method, finally through the experimental reality of federated learning encryption and digital signature and analyzed the above operation on the performance of the federated learning system.