Cedar: A Federated Meta-Learning Framework for Secure and Scalable Personalized IoT

The Personalized Internet of Things (PIoT) demands intelligent learning models that can adapt to highly heterogeneous user data while preserving privacy, scalability, and security. Centralized learning approaches are impractical in PIoT settings due to privacy regulations, non-independent and identically distributed (non-IID) data distributions, and vulnerability to adversarial attacks. To address these challenges, this paper proposes Consent-Driven Ethical Data Access and Regulation (CEDAR), a federated meta-learning framework that enables secure and adaptive personalization across the cloud–edge–device continuum. CEDAR integrates meta-learning with federated learning to extract transferable representations from distributed data, allowing rapid adaptation to individual user contexts with minimal local updates. A layer-wise adaptive uploading mechanism selectively communicates model updates based on parameter importance, substantially reducing communication overhead and accelerating convergence. In addition, asymmetric uploading and anomaly-aware aggregation enhance robustness against gradient inversion and model poisoning attacks. Extensive evaluations on six benchmark datasets covering regression, text classification, and image recognition tasks demonstrate that CEDAR achieves up to 60.39% higher accuracy compared to FedAvg-based federated learning, while reducing communication cost by 23.36% and improving adversarial robustness relative to other state-of-the-art baselines. Ablation studies further confirm the complementary contributions of CEDAR’s core components. By jointly optimizing personalization, privacy, efficiency, and security, CEDAR provides a scalable and ethically aligned learning framework for next-generation PIoT applications in domains such as smart mobility, healthcare, and the digital economy.