Generative Adversarial Network-based Privacy and Utility Preserving Data Transformation for Time Series Data

Time series data are collected and distributed for various utilities in many application domains like healthcare, smart grids etc. Distribution of raw time series data can reveal sensitive information about the users and compromise user's privacy.The existing works for privacy preservation are based on noise addition, compression-based perturbation and geometric transformation techniques.The existing approaches apply transformation uniformly across the dataset without considering the application utilities.This work proposes a privacy & utility preservation technique (PUPT) which is adaptive to application utilities.Sensitive time segments in the series are detected and transformed by deep learning Generative Adversarial Networks (GAN) with multiple objectives of privacy protection, utility maximization adaptive to application semantics and resistance against inference attacks.Through experimental analysis with real and synthetic datasets, the proposed solution is found to have 8.5% higher utility and 10.5% lower perturbation cost compared to recent privacy preserving approaches for time series data.