Differential privacy scheme using Laplace mechanism and statistical method computation in deep neural network for privacy preservation

Abstract

Mountainous amounts of information are now available in hospitals, finance, counter-terrorism, education and many other sectors. Those information can offer a rich source for analysis and decision making. Such information contains user's sensitive and personal data as well. This emanates direct conflict with the user's privacy. Individual's privacy is their right. The existing privacy preserving algorithms works mainly on the numerical data and doesn't care about the categorical data. In addition, there is a heavy trade-off between privacy preservation and data utility. To overcome these issues, a deep neural network - statistical differential privacy (DNN−SDP) algorithm is proposed as the solution to disguise the individual's private and sensitive data. Both the numerical and categorical based human-specific data are considered and fed to the input layer of the neural network. The statistical methods weight of evidence and information value is applied in the hidden layer along with the random weight (wi) to get the initial perturbed data. This initially perturbed data is taken by Laplace computation based differential privacy mechanism as the input and provides the final perturbed data. Census income, bank marketing and heart disease datasets are used for experimentation. While comparing with the state-of-the-art methods, DNN−SDP algorithm provides 97.4% of accuracy with 98.2% of precision, 99% of recall rate and 98.7% of F-measure value. In addition, the fall-out rate, miss rate and false omission rate of the proposed algorithm are less than 4.1%. The DNN−SDP algorithm guarantees the privacy preservation along with data utility.