An Analytical Study of Deep Convolutional Neural Networks Architectures for Automated Malware Identification

Abstract

Cybersecurity continues to face significant challenges from the rapid proliferation of malware, driven by automated generation techniques and mutation mechanisms that allow adversaries to produce large volumes of malicious binaries in a short time. With the advancement of high-performance computing resources, the sophistication and diversity of these malicious programs have also increased, making effective detection more complex. Malware classification therefore plays a crucial role in identifying and categorizing these threats by analyzing distinctive patterns and features among malicious samples.

In this study, we explore the use of transfer learning with the EfficientNetV2 architecture for automated malware family classification using the Malimg dataset. The dataset is divided into training, validation, and testing subsets with a 70/15/15 stratified split to ensure balanced representation across classes. The proposed model achieves impressive performance, obtaining a 98.4% accuracy on the test set, with a macro-averaged F1-score of 0.983, precision of 0.985, and recall of 0.982.

By transforming malware bytecode into visual representations, the model effectively leverages object recognition capabilities of deep convolutional neural networks to distinguish between malware families. This visualization-based approach not only simplifies the feature extraction process but also enhances classification accuracy. Furthermore, the proposed method can be easily extended to other cybersecurity datasets and adapted for real-time threat detection, offering a promising pathway for future advancements in automated malware analysis and intelligent cyber defense systems.