| Popis: |
The file attached to this record is the author's final peer reviewed version. The Publisher's final version can be found by following the DOI link. The ever-increasing use of mobile phones running the Android OS has created security threats of data breach and botnet-based remote control. To address these challenges, numerous countermeasures have been proposed in the domain of image-based Android Malware Detection (AMD) applying Deep Learning (DL) approaches. This paper proposes, implements and evaluates a solution based on pre-trained CNN models using Transfer Learning feature to identify botnets from the ISCX Android Botnet 2015 dataset. More specifically, we study the performance of 6 prominent pre-trained CNN models namely, MobileNetV2, RestNet101, VGG16, VGG19, InceptionRestNetV2 and DenseNet121, in terms of training accuracies, computation time complexity and testing accuracies. The maximum classification accuracy obtained was 91% for Manifest dataset using the MobileNetV2 model. Also, in terms of computational complexity the MobileNetV2 yielded the lowest training time of 16 ms per sample and testing time of 0.9 ms per sample. In order to improve the testing accuracies we plan to further augment these pre-trained models with larger datasets or fine-tune the model parameters for enhanced performance. |
