Privacy Preserving Clusters with Generative Adversarial Network Based Data Transformation

Preserving data from privacy leakage at same time without affecting data centric utilities like clustering is an important requirement in enterprise data analytics.Privacy preserving data analytics solves this problem by transforming the data and execute data centric operations over transformed data.Recently Local Differential privacy (LDP) has gained lot of attention due to its strong privacy guarantee.Data is perturbed locally and sent to servers for data utilities.Because of high statistical variance and minimal consideration for distance property between data records at global level, accuracy of data utilities like clustering and classification is lower in data perturbation techniques of LDP.Also these schemes don't refine the perturbation means for streaming data.In this work, a novel Deep learning DL-3A (Approximate, Adapt and Anonymize) framework is proposed with goal of providing privacy without affecting data utilities.The framework uses generative adversarial network (GAN) based data transformation scheme to solve issues in LDP data perturbation techniques.The representative elements in the datasets are selected applying Steiner minimal tree and GAN transformer is built over representative elements to generate syntactic data.The generated syntactic data preserves privacy without affecting the data utilities.Through experimental analysis with real and synthetic datasets in both static and streaming mode from multi sites, the proposed solution is found to have at least 9% higher perturbation strength with clustering and classification accuracy of at least 3% higher over perturbed data compared to most recent works.