机器学习中核函数的隐私保护计算方法及应用OA

The kernel function maps data into a higher-dimensional space to address linear insepara-bility by quantifying similarity between cross-domain samples.However,its conventional computation on plaintext involving multi-party data interaction poses significant privacy leakage risks.This study proposes a privacy-preserving kernel function computation framework under the semi-honest model.First,three interactive sub-protocols are designed based on Paillier homomorphic encryption and ran-dom perturbation factors:secure inner product computation,secure power function computation,and secure Euclidean distance computation.By partitioning the plaintext space into congruence classes of positive/negative numbers and introducing floating-point scaling factors,compatibility issues of traditional encryption algorithms with real-world datasets are resolved.A nonlinear computation framework is constructed using interactive protocols,enabling secure computation of complex kernel functions through two-party computation and random perturbation techniques combined with Taylor polynomial approximation,while relying solely on additive homomorphic encryption.The framework supports secure computation of linear kernels,polynomial kernels,and Gaussian kernels within a single cryptographic system.The correctness,security,and computational complexity of the scheme are ana-lyzed,and its application scenarios are discussed.Experimental results show that the proposed scheme achieves the privacy protection goal effectively while maintaining the accuracy of kernel function model,and has the advantages of low computational complexity and low time overhead.