Adaptive Malware Detection in Android: An Active Learning and XGBoost Approach

The growing dependence on mobile devices like smartphones has caught the consideration of cybercriminals. It creates numerous chances for intruders to infiltrate and compromise them. This current continuing challenge demands the need for more efficient and effective methods to shield Android users by recognizing various malwares. Though researchers and companies strive for a steady and consistent approach to increase the classification algorithms accuracy to categorize malicious Android applications, the search for the ultimate answer continues. Hence, we introduce an active learning-based classifier to advance the accuracy of classification in Android mal-ware. In our proposed system, we find the prediction probability of each test instance. If this probability falls below a certain threshold, we generate a similar test instance, incorporate it into the original dataset and retrain the model to get better accuracy of the classification. We evaluated the proposed system on CIC-Maldroid2020 dataset which comprises of 17,341 instances with five various android application categories consisting of Adware, Banking, SMS, Riskware, and Benign. This experiment results imply that the XGBoost (eXtreme Gradient Boosting) algorithm outperforms the various other five machine learning algorithms. Further, the active learning component of our proposed system improves accuracy from 95.34% to 98.5%.