Privacy concern has gained increased attention in data analysis, prompting the application of privacy-preserving methodologies. This includes private dataset generation techniques designed to conceal sensitive information, such as anonymization, Differential Privacy (DP), Generative Adversarial Networks (GANs), and Differentially Private GANs (DPGANs). Nonetheless, the utilization of these techniques can influence the importance of features within the privatized dataset, potentially impacting the accuracy and dependability of subsequent data analysis and machine learning models. This study presents a comprehensive and detailed comparative examination to explore the preservation of features’ significance between the privatized dataset and its original counterpart, thus addressing the challenge of information hiding in privacy-preserving techniques. Through a series of experiments, we aim to offer valuable insights into the application of private data generating techniques to uphold the relevance of features, thereby advancing privacy-conscious data analysis across diverse applications.
In distributed data analysis, clustering plays a vital role in uncovering patterns and relationships across datasets. However, ensuring individual privacy in decentralized settings remains a challenge. Existing approaches often fall short in protecting individual privacy, requiring continuous user engagement, or being limited to specific use cases. To address these shortcomings, we propose a novel framework called nD-Laplace, leveraging Geo-Indistinguishability for privacy preservation. Our framework enables non-interactive privacy-preserving clustering while addressing challenges associated with perturbing data. We introduce grid-remapping to handle out-ranged perturbed points, ensuring clustering utility and privacy. We provide theoretical proof of adherence to generalized differential privacy principles and validate the efficacy of our methodology through real-world dataset evaluations and simulated attacks.