A Hybrid AI System for Early Malware Detection and Automated Cybersecurity Risk Assessment

Abstract

The development of improved algorithms for malware identification and classification is necessary the considerable challenge posed by the increasing proliferation of malware in terms of digital security. When dealing with uneven datasets or complicated virus patterns, traditional detection approaches could fail. This research seeks to create efficient ML models for Android malware detection by utilizing two suggested approaches: a Multilayer Perceptron (MLP) model and a Hybrid XGBoost-Random Forest stacking model. Common classification metrics and five-fold cross-validation are utilized to train and assess the models. The Hybrid model showed impressive detection capabilities with an accuracy of 97.87%, while the MLP attained an accuracy of 97.38%. Results from classification using the MLP were comparable, and its performance was consistent throughout validation folds. The Hybrid model's interpretability was further enhanced by using SHAP analysis to uncover key Android properties. Overall, the research contributes two complementary machine-learning approaches for Android malware detection, offering accurate, reliable, and interpretable solutions for strengthening mobile application security.