Machine Learning-Powered Dynamic Analysis Framework for Effective Android Malware Detection

Abstract

Since Android smartphones are so widely used, hackers have turned them into their main target, which has resulted in a sharp increase in sophisticated mobile malware. Only static analysis techniques are not enough to detect evolving threats due to obfuscation and code transformation techniques. We explore a dynamic analysis-based approach for Android malware detection using machine learning algorithms in this research, leveraging the CICMalDroid 2020 dataset. Various classifiers, including Random Forest, Decision Tree, Naive Bayes, Logistic Regression, AdaBoost, Extra Trees, and Gradient Boosting Machine, were trained using dynamic features that captured runtime behavior, such as system calls, API calls, and other runtime behaviors. An extensive assessment of model performance is made possible by the dataset's inclusion of a wide range of malware families and benign applications. Common metrics like Accuracy, Precision, Recall, F1-score, Specificity, Matthew Correlation Coefficient, Cohen Kappa, and ROC Score were used to evaluate the models. The results of experiments show that machine learning models trained on dynamic features are capable of accurately differentiating between malicious and benign applications, with the highest detection accuracy being attained by the Random Forest classifier.