Neural Network Innovations in Image-Based Malware Classification: A Comparative Study

As the digital landscape continuously evolves, the field of cybersecurity faces the growing challenge of effectively detecting and classifying malware. Malware visualization is an emerging technique that involves converting malware’s binary structure into images. This technique uses deep learning as an appealing approach and has been providing promising results. However, the intricate effects of recent deep neural network models on classification performance need to be explored. This paper presents a comprehensive study of malware classification using recently developed models such as ConvNeXt V1 and V2. We initiate our investigation by examining the effectiveness of various neural network architectures in malware detection. We compare traditional models such as VGG16 and ResNet18 with recent ones like ConvNeXt V1 and V2. Our results demonstrate that ConvNeXt V2 significantly outperforms all other examined neural networks in both effectiveness and efficiency. For instance, the ConvNeXt V2 model achieves remarkable F1 scores, excelling on the Malimg and MaleVis datasets with scores of 99.47% and 98.91%, respectively. Our study also includes a comparative analysis between the ConvNeXt V2 model and the recent research works in this area. This analysis reveals that ConvNeXt model substantially outperforms all malware classification methods, achieving an F1 score improvement of 0.1% to 6%. In brief, this research contributes valuable insights to the ongoing efforts aimed to advance malware classification techniques.