Android Malware Detection Method Combining Multi-Frequency Features and Convolutional Neural Networks

With the rapid evolution of Android malware variants, traditional detection methods exhibit significant limitations in addressing complex patterns and adversarial attacks. This paper proposes a dual-modal feature extraction and fusion approach that integrates Convolutional Neural Networks (CNN) with Fourier frequency domain analysis for Android malware detection. Firstly, a novel feature extraction framework based on frequency domain analysis is constructed, which dynamically adjusts the boundary between high-frequency and low-frequency components through adaptive frequency selection. This framework effectively captures both the global patterns and local details of malware, significantly enhancing the robustness of frequency domain features. Secondly, a bidirectional recursive optimization-based feature fusion mechanism is designed, enabling deep integration of high-frequency, low-frequency, and spatial domain features through multiple rounds of iterative interaction. Experimental results demonstrate that the proposed method outperforms existing approaches in terms of accuracy, recall, and adversarial sample detection capability. Notably, it exhibits outstanding stability and performance in handling malware variants, offering an efficient and reliable technical framework for Android malware detection.