A Review of Lightweight Federated Learning Research for Home Edge IoT

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

With the development of 5G and the widespread adoption of home IoT devices and wearable smart devices, home edge IoT devices generate a large amount of data. This data contains sensitive personal privacy and cannot be uploaded to public servers for training. Distributed machine learning technology, specifically federated training, is designed to protect data privacy, ensuring that data does not leave the local environment while multiple parties collaboratively train and share the model. However, home IoT devices differ significantly from industrial-grade IoT devices, exhibiting substantial heterogeneity, primarily in computing power, memory, power consumption, network connectivity, and in user data distribution, which makes it impossible to deploy federated training directly. This paper systematically reviews lightweight federated learning techniques for home edge IoT, analyzing two technical approaches: "device-aware adaptive lightweight training" and "data distribution-aware weighted aggregation." This paper proposes a co-optimization research framework, incorporating pruning rate, quantization bit width, and aggregation weights into a unified optimization space. The goal is to strike a balance among global model accuracy, training energy consumption, and convergence speed under device resource constraints, thereby providing a theoretical reference for future research.