A framework reforming personalized Internet of Things by federated meta-learning
Linlin You, Zihan Guo, Chau Yuen, Calvin Yu‐Chian Chen, Yan Zhang, H. Vincent Poor
Abstract
Advances in Artificial Intelligence envision a promising future, where the personalized Internet of Things can be revolutionized with the ability to continuously improve system efficiency and service quality. However, with the introduction of laws and regulations about data security and privacy protection, centralized solutions, which require data to be collected and processed directly on a central server, become impractical for personalized Internet of Things to train Artificial Intelligence models for a variety of domain-specific scenarios. Motivated by this, this paper introduces Cedar, a secure, cost-efficient and domain-adaptive framework to train personalized models in a crowdsourcing-based and privacy-preserving manner. In essentials, Cedar integrates federated learning and meta-learning to enable a safeguarded knowledge transfer within personalized Internet of Things for models with high generalizability that can be rapidly adapted by individuals. Through evaluation using standard datasets from various domains, Cedar is seen to achieve significant improvements in saving, elevating, accelerating and enhancing the learning cost, efficiency, speed, and security, respectively. These results reveal the feasibility and robust-ness of federated meta-learning in orchestrating heterogeneous resources in the cloud-edge-device continuum and defending malicious attacks commonly existed in the Internet, thereby unlockingthe potential of Artificial Intelligence in reforming personalized Internet of Things.