FedUED: Enhancing Generalization of Federated Learning via Uncertainty-Guided Evidential Distillation in Consumer Electronics
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The proliferation of intelligent consumer electronics has created vast volumes of edge data, while strict privacy constraints prohibit centralized collection. Federated learning (FL) offers a natural solution, but data heterogeneity, which arises from diverse users, regions, and form factors, often degrades global generalization and stability. To address this, this paper introduces FedUED, a framework designed to enhance the generalization of FL in consumer electronics. The core of FedUED is an Uncertainty-guided curriculum, which quantifies sample-level uncertainty by combining feature density with predictive confidence, enabling clients to prioritize challenging data. Guided by uncertainty, local optimization adopts a stable-plus-focused objective where cross-entropy ensures convergence on the full dataset, an evidential distillation(ED) loss. FedUED synergizes evidential learning and knowledge distillation, refines high-uncertainty samples to suppress noise and align them with the global model. Unlike existing approaches that treat all samples uniformly, our method adaptively balances efficiency and robustness. Experiments on CIFAR-10/100, SVHN, and TinyImageNet, which serve as representative proxies for visual tasks in consumer electronics, demonstrate that FedUED achieves consistent improvements over baselines, with better convergence.
Fields of Study & Subjects
Privacy-Preserving Technologies in Data
Domain Adaptation and Few-Shot Learning
Machine Learning and Data Classification
Publication details
- Journal / source
- International Conference on Consumer Electronics
- Pages
- 1–6
- Publication date
- 3 February 2026
- Language
- EN