Differentially Private Noise Addition On Smart Meter Data For Effective Privacy Research

Official URL: https://risc01.sabanciuniv.edu/record=b3400653

Smart meters measure utility consumption, like electricity, gas, or water. Utility providers publish smart meter data to contribute to research and innovation by performing analysis on the data. Data owners utilize limited privacy techniques when publishing smart meter data, such as anonymization, which is susceptible to linkage attacks that allow for the re-identification of individuals. As a result, making smart meter data publicly available raises privacy concerns. Smart meter data could be misused to reveal personal information about daily routines, activities, and private characteristics of households. Differential privacy is a framework that balances the conflicting goals of data utilization and individual privacy. In this thesis, we aim to show to what extent differential privacy can effectively balance household privacy while providing efficient data utilization and information extraction. For this purpose, we use household electricity consumption data. The data set was unbalanced, so Synthetic Minority Oversampling Technique (SMOTE) was used to balance it. Moreover, since the data set was small, Generative Adversarial Network (GAN) technique was used to generate synthetic data based on the real data. Using IBM’s diffprivlib library, conducted various experiments for adding noise to the data and performed machine-learning-based classification over noisy data. We evaluated various noise levels to determine the optimal one that gives a similar classification performance as the original data. It has been determined that the Gaussian Naive Bayes model with differential privacy provides a better differential privacy level (smaller ε) than the Logistic Regression model with differential privacy. Furthermore, it has been shown that the Gaussian noise addition mechanism is the best among the other mechanisms for achieving differential privacy