Abstract
To enable more effective mitigation of Android botnets, image-based detection approaches offer great promise. Such image-based or visualization methods provide detection solutions that are less reliant on hand-engineered features which require domain knowledge. In this paper we propose Bot-IMG, a framework for visualization and image-based detection of Android botnets using machine learning. Furthermore, we evaluated the efficacy of Bot-IMG framework using the ISCX botnet dataset. In particular, we implement an image-based detection method using Histogram of Oriented Gradients (HOG) as feature descriptors within the framework, and utilized Autoencoders in conjunction with traditional machine learning classifiers. From the experiments performed, we obtained up to 95.3% classification accuracy using train-test split of 80:20 and 93.1% classification accuracy with 10-fold cross validation.