Towards Private and Accurate IoT Applications

Over the past years, the fast-growing trend of Internet of Things (IoT) is bringing millions of new smart devices and sensors into homes, office buildings and industries. These smart devices and sensors enable smart IoT applications (e.g., energy prediction, activity recognition, etc.) to increase the quality and efficiency of our lives. To achieve promising performance for smart IoT applications, it requires massive data from different users and sensors to guarantee the performance due to machine learning and deep learning purposes. However, the edge devices of IoT applications often collect and store only limited data, which is insufficient for training modern learning models. Collaboratively training sets steps to achieve better application performance among different devices, while introducing the concern of data privacy. On the other hand, directly applying privacy-preserving techniques such as differential privacy can dramatically degrade the performance of IoT applications.

In this dissertation, we aim to achieve privacy-first smart IoT applications while ensuring their accurate performance for multi-user and multi-sensor scenarios. First, we propose Personalized Federated Deep Reinforcement Learning (PFDRL), a system that helps local users to achieve private and accurate energy management. PFDRL replaces the central server with decentralized federated learning (DFL) framework and enables a personalized federated reinforcement learning to tackle the standby energy reduction in residential building. Next, we present Atlas, a private and accurate personalized federated local differential privacy (LDP) framework for IoT applications. We first design a layer-sharing mechanism called layer importance mask to separate the local model into global and personalized layers. Second, we design a weighted LDP mechanism and add noise to the global layers before transmitting them to the federated learning framework for aggregation. Third, we combine local personalized layers and aggregated global layers to perform IoT tasks. Finally, we introduce Privatehub, a system that utilizes contrastive learning with diffusion models for synthetic data generation in multi-sensor scenarios. Privatehub helps to prevent the private applications' identification from multi-sensor environments while ensuring the accurate performance of the non-private applications.