Intelligent Edge-Based Deep Learning Framework for Scalable Internet of Things: An Adaptive Model Compression and Federated Optimization Approach

Abstract

In the era of large-scale connected devices, the Internet of Things (IoT) has evolved into a critical infrastructure supporting smart cities, industrial automation, and intelligent healthcare systems. However, deploying deep learning models on IoT devices remains a major challenge due to limited computation, energy, and bandwidth constraints. This paper proposes an intelligent edge-based deep learning framework for scalable IoT, integrating adaptive model compression with federated optimization. The proposed framework leverages an edge-cloud collaborative architecture, where deep models are trained across distributed devices without sharing raw data, ensuring privacy preservation and communication efficiency. A self-adaptive pruning and quantization mechanism is introduced to dynamically adjust model complexity according to device capability. Furthermore, an improved federated averaging algorithm is employed to enhance convergence stability under heterogeneous data distributions. Experimental results demonstrate that the proposed system achieves a 38.5% reduction in communication overhead, a 25.2% increase in inference efficiency, and maintains 97.3% of baseline accuracy on the Edge-IoT benchmark dataset. The integration of adaptive compression and federated learning significantly improves scalability, energy efficiency, and privacy protection in edge intelligence, paving the way for sustainable and secure IoT systems.