models

A research team from Turkey, led by Erdal Başaran and Ömer Okucu of Ağrı İbrahim Çeçen University, along with Yusuf Alaca from Hitit University, has developed an innovative detection framework for Android malware that utilizes a visual approach. This framework transforms key features of Android applications into images, allowing a Vision Transformer to differentiate between benign software and malware. The researchers created two types of images from static and dynamic analysis features: a conventional 2D grayscale image and a QR code image. Separate Vision Transformer models were trained on each image type, and spatial pooling features were extracted from deeper layers to create a comprehensive representation of application behavior. To refine the fused feature vector, Recursive Feature Elimination (RFE) was employed, enhancing classification clarity and reducing computational load. The final classification utilized a majority-voting ensemble strategy, achieving a detection accuracy of 98.72 percent, surpassing standalone models. The study demonstrates that the multimodal system outperforms individual grayscale or QR code pathways. The dataset used is publicly available, promoting reproducibility. The framework's reliance on static and dynamic analysis features allows it to adapt to new threats, and the novel QR code representation offers distinctive visual signatures of malicious behavior. However, challenges remain regarding feature extraction quality and computational demands.