Towards Improved DNN-Based Android Malware Detection via Uncertainty Estimation

The long-lasting Android malware threat has attracted significant research efforts in malware detection. In particular, by modeling malware detection as a classification problem, machine learning-based approaches, especially deep neural network (DNN)-based approaches, are increasingly used for Android malware detection and have achieved significant improvements over traditional methods such as signature-based detection. However, due to the rapid evolution of Android malware and the presence of adversarial samples, DNN models trained on historical samples often yield poor predictions when used to detect newly emerging samples. Fundamentally, this phenomenon stems from the uncertainty in the data (noise or randomness) and the limitations of the training process (insufficient training data). Overlooking these uncertainties poses risks to the model predictions. In this paper, we take the first step to estimate the prediction uncertainty of DNN models in malware detection and leverage these estimates to enhance Android malware detection techniques. Specifically, besides training a DNN model to predict malware, we employ several uncertainty estimation methods to train a Correction Model that determines whether a sample is correctly or incorrectly predicted by the DNN model. We then leverage the uncertainty estimates from the Correction Model to correct the prediction results, improving the accuracy of the DNN model. Experimental results show that our proposed MalCertain effectively improves the accuracy of the underlying DNN models for Android malware detection by around 21% and enhances the detection of adversarial Android malware samples by up to 94.38%. Our research sheds light on a promising direction of leveraging prediction uncertainty to improve prediction-based software engineering tasks.