A Hybrid Feature Extraction Framework for Android Malware Detection
Zicheng Ma, Kui Zhao, Jin Yang, Siyu Li, Chen Chen · 2023
In the task of Android malware detection, Convolutional Neural Networks (CNNs) serve as one of the methods for feature extraction. By transforming DEX text into images, the software classification problem can be reframed as an image classification task. However, CNN models require a fixed input image resolution, which necessitates downs amp ling of DEX text, leading to information loss. To address this issue, we augment the CNN model with a Graph Attention Network (GAT) to capture structural features inherent in DEX. In this paper, we propose a hybrid feature extraction framework that leverages both ConvNeXt and GAT for extracting content and structural features of the software, respectively. To evaluate the performance of this framework, we conducted experiments on the CICMalDroid 2020 dataset, achieving an accuracy rate of 98.20%. To further investigate the effectiveness of the structural feature extraction method proposed in this paper, we conducted ablation experiments. Additionally, we compared the ConvNeXt architecture with other CNNs, i.e. VGGNet, ResNet, ResNeXt, and EfficientNet. The results demonstrated that ConvNeXt out-performs other CNNs in extracting content features from DEX text.