Evaluating Deep Learning Classification Reliability in Android Malware Family Detection

Artificial intelligence techniques are nowadays widespread to perform a great number of classification tasks. One of the biggest controversies regarding the adoption of these techniques is related to their use as a 'black box' i.e., the security analyst must trust the prediction without the possibility to understand the reason why the classifier made a certain choice. In this paper we propose a malicious family detector based on deep learning, providing a mechanism aimed to assess the prediction reliability. The proposed method obtains an accuracy of 0.98 in Android family identification. Moreover, we show how the proposed method can assist the security analyst to interpret the output classification and verify the prediction reliability by exploiting activation maps.