Advanced malware detection through dynamic features and attention-powered deep models
Shreeya Prasad, Gunendra M. Chaturvedi, Manish Kumar · 2025
Android is the leading smartphone operating system and a desirable target for malware attacks due to its openness, where applications may be installed outside the official marketplace. Traditional detection methods are ineffective against the sheer number of new applications and sophisticated evasion strategies. Therefore, this paper proposes a new deep learning architecture integrating Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Multi-Head Attention mechanisms for multi-class malware detection. Take advantage of the dynamic analysis and end-to-end preprocessing pipeline—dataset fusion, feature normalization, and reshaping—the model takes on the CCCS-CIC-AndMal-2020 dataset, which contains static and dynamic features, to effectively detect and distinguish malware variants. The model with the proposed method has an average accuracy of 99.77%, the classification accuracies of 99.92% (2 classes), 99.18% (4 classes), 99.34% (6 classes), 99.53% (8 classes), 96.76% (10 classes), 98% (12 classes), and 96.01% (14 classes). It is a scalable and efficient solution for adaptive cybersecurity attacks.