OI-AADX: an optimised image-based Android Adware detection using eXplainable AI

Existing Machine learning (ML)/Deep Learning (DL) based android malware detection approaches suffer from a series of issues viz. Adversarial Robustness, Higher Computational Complexity, Data Scarcity and Imbalance, Explainability and Interpretation, Higher False Positives and Meta/Polymorphic Nature of Malware. Our focus in this research is on two aspects viz. (a) addressing the Explainability and Interpretation issue that typically the DL-based techniques suffer from and (b) optimisation. In this paper, we propose a CNN-based adware detection method that uses an Android permission-based dataset (MadDroid). We convert the dataset permissions into black and white images and use them to train the proposed CNN model. We use eXplainable Artificial Intelligence (XAI) techniques to get SHAP (SHapley Additive exPlanations) values and use them to find the feature’s contribution in the prediction of our model. By following this, the number of features reduces from 440 to 144 key features. We apply statistical methods to analyse SHAP values and highlight the most significant features by training and testing our model on each reduced dataset. We systematically find out the important features using CNN and XAI, and we have maintained the sequence of permission in the image to maintain the same sequence as the original feature set.