Privacy-Preserving Household Characteristic Identification With Federated Learning Method

Abstract

Understanding residential household characteristics is crucial for retailers to provide customers personalized services. Current methods infer household characteristics from smart me-ter data in a centralized manner that requires the data of all re-tailers to be gathered together for model training. This may raise privacy concerns since the privacy-sensitive data are owned by dif-ferent retailers, and they may be unwilling to share the raw data. This paper proposes a federated learning (FL) based deep learning model to identify household characteristics. A hybrid model com-bining the convolutional neural network and long short-term neu-ral network is designed to learn spatial-temporal features from load profiles. It is implemented in a decentralized manner based on the FL framework. To improve the training speed and accuracy, an asynchronous stochastic gradient descent with delay compen-sation method is proposed to update the global model parameters. Comprehensive experiments are conducted to verify the effective-ness of the proposed method.