Android malware detection with image-based features using transfer learning methods

The increase in mobile device usage today has also brought security threats to Android devices. Ensuring the security of Android devices is crucial for protecting user data. Numerous studies have been conducted that encompass both traditional and AI-based methods to secure Android devices. This study aims to enrich the literature by evaluating and comparing the performance of three transfer learning models (ResNet50, DenseNet201, and VGG16) on a hybrid dataset adapted with specific parameters. The dataset obtained from AndroZoo and Drebin classifies the data as benign and malicious for effective malware detection. This research used grayscale images from the dataset to train the aforementioned transfer learning models. The results show that transfer learning models can be successfully used in malware detection. It has been observed that the VGG16 model achieved the best performance in malware detection with an accuracy of 97.24%, a precision of 97.26%, and an AuRoc value of 99.33%. These findings may also provide valuable contributions to developing mobile security applications.