Towards Privacy-preserving Framework for Non-Intrusive Load Monitoring
Henrique Pötter, Stephen L. Lee, Daniel Mossé · 2021
The proliferation of smart meters has led to a vast collection of fine-grained energy data. While energy consumption data has many useful applications that can improve energy-efficiency, it also leaks private information, such as user behavior and occupancy. Such privacy concerns may prevent users from sharing raw energy data. In recent years, federated learning has emerged as a solution that mitigates the problem of sharing raw data by training models in a decentralized manner. In this framework, all data is local to the client, and a centralized server builds a model by exchanging model parameters. Unfortunately, studies have shown that federated models are also prone to privacy attacks. Thus, in this paper, we explore the use of differentially private federated learning (DPFL) and study its effectiveness in mitigating privacy risk for energy-efficiency applications. In particular, we study the effectiveness of the DPFL on Non-Intrusive Load Monitoring (NILM) models' accuracy and whether it can decrease privacy risks. In doing so, we develop an open-source framework that streamlines NILM models' training in a privacy-preserving federated manner. Further, we evaluate our approach on different datasets and compare them against the non-private federated approach as a baseline.