Federated Meta-Learning for Privacy-Preserving AI in Smart Home Ecosystems

Abstract

With a continuing evolution of smart home ecosystems, the enhancement of the artificial intelligence (AI) concept has recently become a key in terms of providing intelligent automation, adaptive control, and personal experiences of users. Such AI-based services largely depend on data about the users, detect patterns, estimate requirements, and make the system performance. But, the traditional approach of putting this information on cloud servers causes severe problems of privacy of the users and ownerships of the data, latency, and scale of the system. To overcome such limitations, the proposed paper presents a new privacy-aware AI model to utilize both Federated Learning (FL) and Meta-Learning (ML) to provide personal, efficient, and secure AI services in smart homes. Federated Learning That is used to train a model over multiple devices in collaboration, without revealing the original data, thus securing better privacy and adherence to data protection standards. Nevertheless, a classic FL is aimed at overcoming non-IID distribution of data and slow convergence to heterogeneous environments. So as to resolve these problems, in our framework, we are including a model-agnostic meta-learning framework that provides each device in the smart home with the capability of adapting rapidly to its local environment with minimal data samples. In this federated meta-learning approach, smart devices are given the power to customize models, but they also enjoy global, shared knowledge base. The architecture proposed incorporates light on device computation and secure aggregation protocols and differential privacy to guarantee sophistication against inference attacks. Our framework is effective, which is confirmed by test simulations and real experiments on large datasets, including CASAS and synthetic smart home activity log. Evaluation outcomes indicate that our method of evaluation is disproportionately high compared to the traditional FL and centralized model in terms of model accuracy, adaptation speed, communication efficiency and privacy protection. The study forms a solid basis of scalable, personalised and trustable AI in intelligent homes, which can provide insightful information on how federated meta-learning systems can be deployed in various environments, especially those that are privacy-sensitive.