An Efficient Hybrid MobileNetV3–LightGBM Framework for Image-Based Malware Detection

Abstract

Traditional signature-based malware detection techniques have become inadequate in addressing the complexity of modern cybersecurity threats. To overcome these limitations, this paper presents an intelligent malware classification framework that leverages computer vision and deep learning. The Malimg dataset, consisting of grayscale images representing diverse malware families, was utilized to facilitate structural and behavioural feature extraction. The hybrid MobileNetV3—LightGBM model proposed in this paper combines the lightweight MobileNetV3 architecture for efficient deep feature representation with the Light Gradient Boosting Machine (LightGBM) for robust and accurate classification. Experimental results demonstrate that the proposed model outperforms conventional deep learning approaches such as CNN and CNN—SVM, achieving an accuracy of 97.6%, with precision, recall, and Fl-score averaging 98%. These findings confirm that integrating lightweight convolutional networks with gradient-boosted decision techniques significantly enhances malware detection performance and generalization. The proposed framework provides a scalable and effective solution for real-time malware analysis and establishes a foundation for future research on adaptive and explainable Al-driven cybersecurity systems.