Lightweight Deep Neural Networks for Real-Time IoT Sensing and Analytics

Abstract

Recent advancements in deep learning have significantly enhanced intelligent data analysis within Internet of Things (IoT) ecosystems. However, conventional deep neural networks (DNNs) are computationally expensive, energy-consuming, and unsuitable for resource-constrained IoT edge devices. This paper proposes a lightweight deep learning framework designed for real-time IoT sensing and analytics. The proposed architecture integrates model compression, knowledge distillation, and quantization-aware training to achieve low-latency inference with minimal memory footprint. Moreover, an adaptive edge-cloud collaboration mechanism dynamically offloads complex computation based on device workload and network conditions. Experimental results on benchmark IoT datasets demonstrate that the proposed method achieves 28.7% faster inference speed and 32.1% lower power consumption compared to baseline DNNs while maintaining comparable accuracy. These findings indicate that lightweight neural architectures can significantly enhance IoT performance and enable intelligent, sustainable, and scalable sensing systems.