Android Malware Detection Using an Optimized and Explainable Attentional Convolutional LSTM Model

Abstract

Due to the increasing number of Android applications, its usage has increased which are exposing ourselves into complex mobile malware which is a serious threat on user privacy and system integrity. Machine Learning (ML) based detection methods are considered traditional and generally perform poorly in modeling the complex sequential activity of API-call patterns, typically having low interpretability of their decisions. To address these challenges, this work proposes an optimized and interpretable attentional CNN–LSTM framework for Android malware detection. It utilizes convolutional layer to extract spatial features, Long Short-Term Memory (LSTM) networks to capture patterns in sequential behavior and attention mechanism to highlight distinctive sequences of API calls. Moreover, Manta Ray Foraging Optimization (MWO) is applied for adaptive parameters fine-tuning and convergence improvement to enhances detection performance. We validate our approach with a wide range of experiments on two absolutely popular datasets for API call–based malware detection and result show high accuracy, precision, recall as well as F1-score owing to the 90:10 train–test split. Randomization-based statistical validation confirms the robustness and significance of results (p < 0.05, non-parametric tests). Besides classification performance, the proposed framework has an innate explainability as examining attention weights through API sequences allows for transparent interpretation of detection decisions. Malicious and benign applications exhibit different concentration patterns in the usage of APIs as revealed by attention heatmaps and top-weighted API analysis Hence, the proposed approach provides end to end, optimized and interpretable solution for reliable android malware detection.