Avoiding Occupancy Detection From Smart Meter Using Adversarial Machine Learning

More and more conventional electromechanical meters are being replaced with smart meters because of their substantial benefits such as providing faster bi-directional communication between utility services and end users, enabling direct load control for demand response, energy saving and so on. However, the fine-grained usage data provided by smart meter brings additional vulnerabilities from users to companies. Occupancy detection is one such example which causes privacy violation of smart meter users. Detecting the occupancy of a home is straightforward with time of use information as there is a strong correlation between occupancy and electricity usage. In this work, our major contributions are twofold. First, we validate the viability of an occupancy detection attack based on a machine learning technique called Long Short Term Memory (LSTM) method and demonstrate improved results. In addition, we introduce an Adversarial Machine Learning Occupancy Detection Avoidance (AMLODA) framework as a counter attack in order to prevent abuse of energy consumption. Essentially, the proposed privacy-preserving framework is designed to mask real-time or near real-time electricity usage information using calculated optimum noise without compromising users' billing systems functionality. The results show that without the use of the proposed AMLODA approach, our occupancy detection attack models using LSTM achieve a high detection rate with Matthews Correlation Coefficient (MCC) value of 0.89 on average for the five different households energy consumption data under investigation captured during the winter and summer seasons. With the proposed AMLODA approach working to protect consumers' privacy, occupancy detection attacks are demonstrated to be mitigated with the MCC values of the attack models converging to zero with no significant change over the actual consumption data and thus protecting needed functionalities of the utility companies.