Multi Attribute Stratified Sampling: An Automated Framework for Privacy-Preserving Healthcare Data Publishing with Multiple Sensitive Attributes

The accumulation and analysis of large-scale patient data have led to breakthrough discoveries in potential flags for diseases based on pattern recognition, highlight medication efficacy, and local population health trends that would be impossible with traditional paper-based records. However, these benefits come with unique challenges posed by the application of data sharing for research and analysis, and mandatory requirements that require careful balance between privacy protection and usefulness of data especially when the data contains several sensitive information. We propose a framework, Multi Attribute Stratified Sampling (MASS), to achieve automatic parameter optimization by separating the sanitization process from manual privacy parameter configuration. Most traditional privacy-preserving techniques require experts to specify privacy parameters such as k, l, and t values for k-anonymity, l-diversity, and t-closeness respectively based on intuition or trail and error resulting in sub-optimal privacy-utility-tradeoffs. In contrast our framework employs a self-tuning paradigm which uses GetAnonymized, CandidateBuilder, and Optimizer modules. CandidateBuilder produces multiple anonymized versions of the original preprocessed data by iteratively calling GetAnonymized on a range of anonymization levels creating a solution space. The Optimizer then scans through the solution space using an objective function to determine the optimal anonymized version. The objective function utilizes privacy and information losses along with the classification recall to discover the privacy parameters that yield the best balance between privacy protection and data utility, eliminating the burden of manually fine tuning the privacy parameters and ensuring reproducible outcomes across various healthcare datasets and analytical contexts. Experimental validation on four datasets (1k, 10k, 100k and 10M) demonstrates that MASS achieves strong privacy protection across datasets with a privacy loss < 0.25 while maintaining >95% recall retention on datasets exceeding 10k records. Given that the computational complexity to generate anonymized dataset is NP-hard, MASS presents a polynomial-time heuristic solution which validates practical implementation and scalability for real-world deployment.