Ses ve görüntü dönüşümü kullanilarak android kötücül yazilim tespiti

Mobile devices have started a new era with their hardware and various software developed for them. A security vulnerability in these devices could lead to the theft of personal information, breaches of privacy, and even financial loss. Therefore, ensuring that the apps downloaded to the devices are reliable and safe is very important. For this purpose, within the scope of this thesis, an investigation into the image and audio-based approaches for Android malware detection and family classification is conducted. In the image-based approach, an end-to-end method is proposed that treats Android application files as binary sequences. In the method, grayscale image representations were created for each sample, and training and testing processes were carried out with CNN. Image representations of malware can be made by treating the files as binary sequences or the extracted static features in matrix form. For this reason, the impact of static feature set combinations on classification performance is also investigated. Initially, all possible combinations of four different feature sets obtained from Android application files are considered, and their effects on classification performance are investigated. Effective feature set combinations are determined by evaluating all combinations with ten different classification algorithms. Subsequently, RGB images are created with feature set combinations, and training and testing processes are carried out using CNN. In the results obtained, it was seen that with different feature set combinations, a performance above 99% could be obtained. In addition to image representations of Android applications, audio representations can also be created. Although audio-based approaches are less common than image-based ones in the literature, they can achieve similarly high classification accuracies. In this context, an audio-based method is proposed, treating the Android malware family detection problem as a music category classification problem. Android application files were converted to audio files, and their audio-based attributes were extracted. Then, features with high discrimination were determined with four different feature selection algorithms, and the classification processes were carried out. Family detection was performed with 96.6% accuracy in experiments on an eight-class data set. At the end of the thesis, discussions were made about the methods used in the study.