Abstract
Android losing its resistance to complex malware has caused grave doubts about the privacy of users, the security of their data and financial security. As malicious applications and obfuscation techniques continue to develop at a high pace, traditional malware detection methods are unable to remain accurate in real-time settings. This paper provides an in-depth guide to identifying malware written in Android, utilizing machine learning and large language model (LLM) methods. The methodology will include the data collection, vocabulary construction, preprocessing permission and API call, computation of risk score, and extraction of features using TFIDF to guarantee adequate representation of the input. Some of the popular measures of performance are accuracy, precision, recall, and F1-score to train and evaluate the proposed models, LLM-Guard, Autoencoder (AE), Decision Tree (DT), and Multi-Layer Perceptron (MLP). The dataset was split into subsets, training and testing of which were to be applied in order to support supervised learning and objective evaluation. The experimental results present that the proposed LLM-Guard achieved the highest accuracy of 98.41 against AE (90.71%), DT (93.47%), and MLP (89.2%). This evidence demonstrates the efficiency and scalability of the LLM-Guard framework and makes it the best solution for detecting Android malware in realtime in the dynamic mobile security setting.