Abstract
Android is arguably the most widely used mobile operating system in the world. Due to its widespead use and huge user base, it has attracted a lot of attention from the unsavory crowd of malware writers. Traditionally, techniques to counter such malicious software involved manually analyzing code and figuring out whether it wasmalicious or benign. However, due to the immense pace at which newer malware families are surfacing, such an approach is no longer feasible. Machine learning offers a way to tackle this issue of speed by automating the classification task. While several efforts have been made to use traditional machine learning techniques to Android malware detection, no reasonable effort has been made to utilize the newer, deep learning models in this domain. In this paper, we apply several deep learning models including fully connected, convolutional and recurrent neural networks as well as autoencoders and deep belief networks to detect Android malware from a large scale dataset of more than 55 GBs of Android malware. Further, we apply Bayesian machine learning to this problem domain to see how it fares with the deep learning based models while also providing insights into the dataset. We show that we are able to achieve better results using these models as compared to the state-of-the-art approaches. Our best model gets an F1 score of 0.986 with an AUC of 0.983 as compared to the existing best F1 score of 0.875 and AUC of 0.953.