An evaluation of the effectiveness of smart meter data perturbation mechanisms using a unified stochastic framework
The growing energy needs are forcing utility companies towards the use of Advanced Metering Infrastructure (AMI). Smart meters are a part of the AMI, that capture fine grained electric consumption data from households and share it with the utility company, operations center, and other third party entities in the smart grid network, in order to make the grid more efficient and reliable. Yet, sharing such fine grained data gives rise to privacy concerns from the consumer's perspective. As a response to such privacy concerns, several smart meter data perturbation techniques have been proposed to fortify against extraction of sensitive personal information from the consumers' electric consumption data. However, due to the lack of a unified framework for assessing and comparing different perturbation techniques, consumers do not have access to a readily available tool to measure and evaluate the privacy provided by these perturbation techniques. We introduce a unified and practical data centric framework for evaluating and measuring the privacy offered by different smart meter data perturbation techniques. The framework trains multiple smart meter data models based on past smart meter data and other auxiliary information (e.g., time, temperature, location, etc.) that may be easily available to the adversary. The framework then evaluates the privacy offered by different perturbation techniques by carrying out reconstruction attacks (usually the adversary's first step before performing more advanced inference attacks) using trained models suitable for the characteristics of the perturbed data. Accuracy of reconstruction and the loss of privacy is measured in terms of well-known metrics, such as R-squared correlation and relative entropy. The framework is said to be unified because it considers all the elements required for evaluation such as the prior information available to the adversary, perturbation techniques, and the reconstruction strategy. To validate the effectiveness of our framework, for evaluating and measuring privacy offered by some of the popular perturbation techniques, we test the framework using real smart meter power consumption data collected from twenty-two households over a period of two-years.
Thesis (M.S.)--Wichita State University, College of Engineering, Dept. of Electrical Engineering and Computer Science