Leveraging attention-based deep neural networks for security vetting of Android applications

Many traditional machine learning and deep learning algorithms work as a black box and lack interpretability. Attention-based mechanisms can be used to address the interpretability of such models by providing insights into the features that a model uses to make its decisions. Recent success of attention-based mechanisms in natural language processing motivates us to apply the idea for security vetting of Android apps. An Android app’s code contains API-calls that can provide clues regarding the malicious or benign nature of an app. By observing the pattern of the API-calls being invoked, we can interpret the predictions of a model trained to separate benign apps from malicious apps. In this paper, using the attention mechanism, we aim to find the API-calls that are predictive with respect to the maliciousness of Android apps. More specifically, we target to identify a set of API-calls that malicious apps exploit, which might help the community discover new signatures of malware. In our experiment, we work with two attention-based models: Bi-LSTM Attention and Self-Attention. Our classification models achieve high accuracy in malware detection. Using the attention weights, we also extract the top 200 API-calls (that reflect the malicious behavior of the apps) from each of these two models, and we observe that there is significant overlap between the top 200 API-calls identified by the two models. This result increases our confidence that the top 200 API-calls can be used to improve the interpretability of the models.