Threshold-Optimized Ensemble Learning for Accurate and Secure Android Malware Detection

This paper proposed a machine learning approach for the classification of Android applications as benign or malicious based on permission counts as features. The method proposed in this work utilizes a stacking classifier ensemble, which combines Logistic Regression, Random Forest, and SVM to leverage their strengths and enhance predictive accuracy. The model was fine-tuned with a decision threshold of 0.44, which is important for achieving real-world malware detection that balances detection accuracy and false positives. The stacking classifier reached an accuracy of 84.2%, precision of 82.48%, recall of 96.93, and ROC AUC of 0.92, pointing out that it did well in terms of high recall, which is critical to keeping the number of false negatives to a bare minimum for malware detection. The experiments carried out demonstrate the potential and efficiency of employing ensemble learning with threshold optimization for enhanced Android malware detection. The work extends ensemble methods towards better security in Android applications.