Malware remains a significant threat to modern computing systems and networks worldwide. Evolving malware utilises polymorphism, metamorphism, and also zero-day exploits to bypass defenses. Traditional signature-based and heuristic detection methods are now struggling with the increasing complexity of malware. In this paper, a hybrid neural network model is proposed that combines a convolutional neural network (CNN) to detect spatial malware patterns, recurrent neural networks (RNNs) to analyses temporal behaviors, and also attention mechanisms to select crucial features for accurate and reliable threat classification. Multi-scale convolutional layers and residual connections improve dataset generalization and reduce overfitting. Focal loss functionality addresses the class imbalance in real-world malware detection scenarios. Experimental results on EMBER, EMBER Sim, and SoReL-20M datasets show exceptional accuracy and precision. This interpretable, scalable deep learning (DL) model bridges traditional methods with modern cybersecurity challenges. The model excels in zero-day detection and produces a few false positives, achieving 96.7% accuracy, 96.1% precision, 96.8% recall, and 96.4% F1-score. Additionally, the findings demonstrate clear improvements over previous methods, achieving a 1.1–2.6% increase in accuracy, confirming the model’s superior detection capability. This advanced deep-learning approach sets a new benchmark in cybersecurity.
Muthana S. Mahdi (Sun,) studied this question.