Convolutional neural network based android malware detection andlocalization

Android malware poses significant threats to mobile users by compromising privacy and security. Identifying the precise location of malicious code within an application is crucial for malware analysis and the development of effective countermeasures. This is particularly crucial for artificial intelligence (AI)-based malware detection techniques, as it improves the interpretability and transparency of such solutions. Among these techniques, Convolutional Neural Networks (CNNs) have emerged as a promising approach for malware detection due to their ability to capture spatial patterns in image-like data. By transforming code into visual representations, CNN-based approaches enable automatic feature extraction, reducing reliance on manual analysis. Additionally, these methods exhibit resilience against code obfuscation, as they focus on structural patterns rather than specific code syntax. However, no existing study has focused on localizing malicious code within malware using this approach. In this study, we propose a CNN method leveraging multiple saliency methods (Grad-CAM++, Score-CAM, SmoothGrad, Vanilla Saliency, Integrated Gradients) attention mechanism to localize hot pixels in Android malware images and map them back to their corresponding code segments for malicious code localization. Our methods creates a heatmap that highlights key regions in the visualized code, allowing for the identification of code segments that contribute to the classification of malware. Experimental results on real-world malware demonstrate that the proposed approach effectively identifies malicious code at the class level.