Design and Implementation of a Cloud Computing Privacy-Preserving Machine Learning Model for Multi-Key Fully Homomorphic Encryption

The rapid development of AI, the Internet, and cloud computing has made the research of distributed machine learning a key driver, greatly facilitating human life. Meanwhile, certain emergent security issues are very striking. In Ring Allreduce model architecture, impairment of integrity on the gradient parameters results in an incorrect training model. In federated learning, confidentiality and integrity threats are found on the transmission process of gradient parameters, which may disclose privacy or result in incorrect model training. It, therefore, examines integrity protection algorithms for the training data and gradient parameters of distributed machine learning systems that are beset with challenges such as privacy leakage, key escrow, and dynamic worker node management. In particular, it introduces the RAA-DIP scheme to protect the integrity of the gradient parameters under the Ring Allreduce architecture. It securely negotiates session keys through an identity-based group key agreement algorithm and achieves dynamic management of worker nodes and key escrow problem solution. Further, the FL-DCIP scheme is proposed to provide confidentiality and integrity for gradient parameters in federated learning systems. It uses ElGamal encryption algorithm for encrypting gradient information and integrates sampling audits with BLS signatures by a trusted third party for integrity. It also provides both dynamic management of worker nodes and key agreement.