A privacy protection method for large-scale graph data supported by homomorphic encryption
TL;DR
Combined with privacy leakage probability verification, the protection strategy is optimized to form a full-process privacy protection system and shows that the research method keeps the privacy leakage rate within 0.375%.
Abstract
Given the characteristics of large-scale graph data, such as high-dimensional sparsity, topological coupling perturbations, and inherent heterogeneity in sensitive contexts, traditional coarse-grained privacy identification rules are unable to characterize these non-orthogonal semantic features. This leads to technical bottlenecks in privacy protection, including "coarse granularity, diffuse boundaries, and inaccurate identification," significantly limiting security. To address these issues, a privacy protection method for large-scale graph data supported by homomorphic encryption is proposed. A privacy information identification index system is constructed and a random forest algorithm is adopted to classify the privacy level of graph data. Privacy information and non-privacy information are separated and a bidirectional mapping relationship is established. The separated privacy information is preprocessed, such as denoising, to avoid destroying the graph structure during the separation process. The Paillier homomorphic encryption algorithm is used to generate a key to encrypt the preprocessed privacy information. The security and correctness of the encrypted data are ensured through key offset and ciphertext verification mechanisms, generating directly computable ciphertext graph data. Ciphertext domain graph computation is realized based on the additive homomorphic property of the Paillier algorithm. A role-based access control mechanism is designed to prevent unauthorized access. A dynamic update mechanism for ciphertext and key is established. Combined with privacy leakage probability verification, the protection strategy is optimized to form a full-process privacy protection system. Experiments show that the research method keeps the privacy leakage rate within 0.375%. Compared with the comparison method, the privacy protection accuracy rate is up to 12.6 percentage points higher at 5000 nodes. The encryption time increases more gradually and is lower with each training round, effectively ensuring security.