User-perceptional privacy protection in NILM: A differential privacy approach

This paper presents an approach to achieving user-perceptional privacy protection in Non-Intrusive Load Monitoring (NILM) datasets within the Differential Privacy (DP) framework. Our method is tailored to align privacy protection with genuine user expectations while simultaneously ensuring data authenticity and utility for downstream tasks dependent on pattern recognition. Specifically, based on DP, our privacy-preserving mechanism protects sensitive and critical user behaviors of appliance usage—namely, appliance start times, usage durations, and power levels. Based on user preferences, our mechanism allows for the flexible allocation of privacy budgets to these behaviors, offering customizable privacy protection. As protecting privacy harms data authenticity and utility for pattern recognition tasks, to solve this problem, we first utilize probability mass functions (pmfs) of user behaviors to effectively characterize appliance usage patterns. Then, our post-processing optimization utilizes a network flow framework to align the pmf of privacy-preserving behaviors closely with that of real behaviors. This alignment guarantees the authenticity and utility of appliance usage patterns in privacy-preserving data, ensuring that our privacy-preserving measures do not compromise the data’s accuracy and usefulness for analysis. To manage the significant computational burden of the network flow, we utilize the inherent sparsity of pmfs to reduce computational complexity, thereby enhancing the efficiency of our approach. Our approach allows users to prioritize the privacy of specific behaviors, demonstrating the method’s flexibility, efficiency, and applicability in real-world scenarios. Numerical validations reveal our method’s superiority in privacy preservation, data utility, and computational efficiency. Ultimately, this work underscores the importance of a user-centric approach to privacy that aligns with authentic user concerns and needs.