User privacy leakages from federated learning in NILM applications

Abstract

Non-intrusive load monitoring (NILM) is a technology that estimates the energy consumed by each appliance in the building from the main electricity meter reading only. Federal Learning (FL) is increasingly employed to construct a distributed learning environment to address the lack of data issues in NILM applications. Although FL inherently provides client privacy by sharing training parameters instead of raw data with the federated server, this does not ensure the user privacy is in absolute security. This work aims to investigate the potential user privacy leakage issues of NILM applications using the federated learning frameworks. We experimentally study what data can be revealed and how vulnerable they can be. We are also towards building a new federated learning framework to provide better security for NILM applications.