Lightweight and Scalable Feature Reduction Approach for Cyber Defense System

Abstract

The growing use of Internet of Things (IoT) and Wireless Sensor Networks (WSNs) has increased cyber threats while imposing strict hardware limitations on devices. Traditional Machine Learning (ML) models are too resource-intensive for such Resource-Constrained Devices (RCDs). This paper presents a Lightweight and Scalable Feature Reduction Approach (LSFRA) that enhances cyber defense efficiency through optimized Feature Selection (FS) and Feature Extraction (FE). Using the Extra Tree Classifier (ETC), feature dimensionality for malware detection was reduced by 82% with minimal accuracy loss and a 73% reduction in execution time. In WSN intrusion detection, Sequential Filtering achieved a 91.5% feature reduction while maintaining 97% accuracy. Additionally, a Deep Asymmetric Convolutional Autoencoder (DACA) with Deep Reinforcement Learning (DRL) ensured high adaptability and low-latency detection. Results confirm that effective feature reduction is vital for achieving lightweight, scalable, and energy-efficient cyber defense in IoT and WSN environments.