The advancement of artificial intelligence has achieved significant success across various domains, including computer vision, natural language processing, and generative AI. This advancement is also extending to the field of cybersecurity. In particu...
The advancement of artificial intelligence has achieved significant success across various domains, including computer vision, natural language processing, and generative AI. This advancement is also extending to the field of cybersecurity. In particular, AI-based malware attacks enable more sophisticated and complex attack scenarios, such as automated malware generation and propagation. These attacks are particularly threatening because they allow attackers to execute attacks using AI tools even without expertise in cyberattacks. Furthermore, attackers can rapidly generate malware variants or apply novel obfuscation techniques to easily bypass existing malware defense systems. Therefore, it is necessary to develop malware analysis methods that are robust against obfuscation techniques.
Among deep learning-based approaches, image-based malware analysis methods are effective at detecting new malware variants by extracting and analyzing malware features based on visual similarities and texture patterns in malware images. Convolutional Neural Networks (CNNs), a representative vision model widely employed in these methods, have shown high malware classification performance. However, CNNs have limitations in effectively utilizing global features of images, making them vulnerable to obfuscation techniques such as code insertion or transposition. To address these limitations, Vision Transformer (ViT)-based methods have been researched for malware analysis. However, these methods fail to capture local features of images or have generalization performance limitations of ViTs, making effective malware analysis difficult. Therefore, to employ ViTs in the malware domain, it is necessary to develop ViT-based methods that can provide effective and robust malware analysis.
Therefore, in this paper, we propose Vision Transformer (ViT)-based methods for robust malware analysis. The contributions of this research are as follows: (1) We introduce ViT-based methods for malware analysis. Additionally, we develop a novel model called MalFormer, specifically designed for image-based malware analysis, and propose a new analysis method based on it. (2) To overcome the limitations of ViT-based models and enhance their generalization performance, we introduce Sharpness-Aware Minimization (SAM) into the training process. (3) We validate that the proposed methods achieve high performance and robustness against obfuscation techniques across various malware datasets, and analyze the model’s predictions based on Explainable AI (XAI).
From the experimental results, we show that the proposed methods achieve high performance. In particular, MalFormer, specifically designed for malware analysis, achieves the highest performance and provides strong robustness against obfuscated samples by enhancing sensitivity to key features while maintaining a well-balanced integration of local and global features. We show that applying SAM optimizer overcomes the limitations of ViT and improves generalization performance, enabling the effective use of ViT-based methods for malware analysis. We validate that the proposed methods provide strong robustness through visualization-based analysis including XAI. Therefore, we provide Vision Transformer-based methods that can be effectively applied to malware analysis, and we expect this work to contribute to future research in vision model-based malware analysis.