Enhancing Efficiency and Robustness of Compressed Deep Learning Models on Edge Devices for IoT Sensor Data

Abstract

The proliferation of Internet of Things devices has amplified the need for efficient and reliable edge-based intelligence. Compressed deep learning models, particularly those using post-training quantization and quantization-aware training, offer promising solutions for resource-constrained edge deployments. However, real-world IoT sensor data are often noisy, incomplete, or subject to drift, which can significantly degrade model performance. This study systematically investigates the trade-offs between computational efficiency and robustness of compressed models under realistic sensor noise conditions. Through controlled experiments, we demonstrate that PTQ, while highly efficient in terms of latency, is prone to substantial accuracy loss in noisy environments. In contrast, QAT preserves model accuracy while maintaining low inference latency, providing a more reliable solution for practical edge deployments. The results highlight the necessity of robustness-aware compression and underscore the limitations of evaluating models solely on clean datasets. This work offers a deployment-centric framework for assessing model trustworthiness and provides guidance for developing lightweight, robust deep learning models for safety-critical IoT applications, including industrial monitoring and healthcare.