Artificial Intelligence Based Smart Communities with Privacy Enhancement

Abstract

This thesis introduces Artificial intelligence-based smart communities, including the optimization of smart home and home microgrid applications, thus allowing communities to be more comfortable and affordable for their residents while protecting user privacy. For these purposes, a three-layer smart community framework is proposed.

The smart home layer focuses on a household device action recommendation system. The household appliance data is organized into a knowledge graph. A framework, `DARK' (Device Action Recommendation with Knowledge graph), is proposed, including three parts. Firstly, a household device action recommendation algorithm is proposed to make accurate household appliance recommendations. Next, graph interpretable characteristics are developed in DARK using trained graph embeddings. Lastly, with the recommendation expectations, the consumers' degree of satisfaction and the appliances' average power load are modelled as a multi-objective optimization problem in DARK to participate in demand response.

The home microgrid layer concentrates on the home microgrid scheduling algorithm. The proposed microgrid model includes energy storage systems, PV panels, loads, and the connection to the main grid. Firstly, a multi-objective deep reinforcement learning architecture is proposed for accumulated carbon emissions and electricity costs optimization. Secondly, data privacy is protected by federated learning, in which the original data is not uploaded to the server but remains locally stored. Finally, the optimization results are selected and stored to form Pareto fronts, allowing users to bias towards their preferred optimization goals.

The privacy protection layer focuses on the measurement of potential privacy leakage in federated learning. With the usage of an explainable algorithm, a framework named SAFE-Home (Smart Applications in Federated learning with Emphasis on Home privacy) is proposed to locate the key factors that affect privacy leakage and vulnerable data in federated smart communities. The theoretical analysis and simulation results are consistent, demonstrating the similarity in the privacy leakage trends in different applications and situations.