Differentially private noise addition on smart meter data for effective privacy research

Smart meters measure utility consumption, like electricity, gas, or water. Utilityproviders publish smart meter data to contribute to research and innovation by performinganalysis on the data. Data owners utilize limited privacy techniques whenpublishing smart meter data, such as anonymization, which is susceptible to linkageattacks that allow for the re-identification of individuals. As a result, makingsmart meter data publicly available raises privacy concerns. Smart meter data couldbe misused to reveal personal information about daily routines, activities, and privatecharacteristics of households. Differential privacy is a framework that balancesthe conflicting goals of data utilization and individual privacy. In this thesis, weaim to show to what extent differential privacy can effectively balance householdprivacy while providing efficient data utilization and information extraction. Forthis purpose, we use household electricity consumption data. The data set wasunbalanced, so Synthetic Minority Oversampling Technique (SMOTE) was used tobalance it. Moreover, since the data set was small, Generative Adversarial Network(GAN) technique was used to generate synthetic data based on the real data. UsingIBM’s diffprivlib library, conducted various experiments for adding noise to the dataand performed machine-learning-based classification over noisy data. We evaluatedvarious noise levels to determine the optimal one that gives a similar classificationperformance as the original data. It has been determined that the Gaussian NaiveBayes 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 bestamong the other mechanisms for achieving differential privacy.