A unified privacy-preserving data mining framework with multi-noise injection and hybrid deep learning for robust privacy-utility trade-offs.

This paper introduces a single Privacy-Preserving Data Mining (PPDM) framework that attempts to balance the privacy-utility dilemma between privacy and utility of the data. This methodology will combine a stringent preprocessing pipeline, which uses k-NN imputation and Tomek Link SMOTE to balance the classes and a multi-regularized approach to feature selection (L1, L2, and Elastic Net) to boost the significance of the attributes. As an extensive multi-noise injection layer is required to offer strong privacy assurances, Laplace, Cauchy, t-distribution, intuitionistic fuzzy, exponential, speckle, and Gaussian perturbations are adopted. Studying is carried out through a hybrid 1D CNN-LSTM system, which was designed to identify patterns in perturbed high-dimensional data. Comparison between the Breast Cancer, Adult, and Customer Churn datasets demonstrate that the framework has high predictive accuracy and Kappa scores with a minimum variance in the data utility. Experimental findings show that such multi-layered construction has a better defence-in-depth mechanism than traditional PPDM pipelines, which are capable of ensuring strong performance in sensitive analytical conditions.