Statistical appliance inference in the smart grid by machine learning

Smart Grid has been attracting more interest than ever thanks to emergence of enabling technologies such as 5G and IoT. Yet, there are some long-standing privacy concerns about revealing habits and lifestyles of people from fine-grained power consumption data collected through smart meters. In this context, the contribution of this work is twofold: First, we empirically demonstrate how appliance-level fine-grained power consumption data can reveal households' routines simply using probability density estimations derived from consumption data without requiring any complex analysis. Second, we point out that appliance types can be identified in a targeted house using circuit-level consumption data of other houses. We show how machine learning can be used maliciously to realize this threat in an automatic manner and achieve high success rate even with limited amount of training data on the public REDD dataset. In addition, we provide discussions on possible countermeasures against the threats examined in this study.