Adversarial Error Mitigation Technique for Sensor Fusion Framework on Dynamic Edge Computing Platform

Abstract

Edge computing environments encounter significant challenges in processing unreliable sensor data, while dealing with dynamic systems like health monitoring or IoT networks. This paper suggests a novel architecture that addresses the challenges of inconsistent sensor data in self-configurable, dynamic, and resource-limited edge processing environments. To validate the feasibility of the proposed sensor fusion framework in a real-world, resource-constrained setting, we hypothesized that the model could operate efficiently on low-power edge devices while maintaining high classification performance. One of the key aspects of this model is its data imputation module, which employs K-Nearest Neighbors to handle missing and incomplete sensor data. This enhances the reliability of the fusion process. Performance evaluations conducted on the health dataset demonstrate that the proposed model outperforms conventional deep learning models such as Convolutional Neural Network, Feedforward Neural Network, Long Short-Term Memory, and Gated Recurrent Unit across various performance metrics. The proposed model achieved 99.42% accuracy, 99.43% precision, 99.42% recall, and 99.42% F1-score across multiple health parameters including temperature, systolic and diastolic blood pressure, pulse rate, and SpO2. Additionally, the model showed 99.44% sensitivity, 99.46% specificity, and 99.39% resistivity to noisy or adversarial inputs. The system offers a 20% reduction in inference time and 15% lower energy consumption. These results highlight the potential of proposed deep learning model for real-time, energy-efficient sensor fusion in critical applications like healthcare, autonomous systems, and industrial IoT, where reliability, performance, and adaptability are paramount.