A Comprehensive Survey of Malware Detection Using Deep Learning: Analysis Methodologies

Background: Modern obfuscation, polymorphic, and behavioral evasion techniques can hide or transform malware enough that signature-based and hand-created feature detectors are no longer effective against either new or modified malware. These issues have required the development of advanced detection schemes based on adaptive and resilient learning. Deep learning (DL) has garnered important attention because of its automated feature extraction, ability to model complex behaviors, and effective generalization across diverse malware. Objective: This narrative review summarizes the current state of the art of detecting malware on Android, Windows, and different Internet of Things (IoT) platforms through Deep Learning. Given the volume of literature, this review provides an accounting of the current prominent methods, including architectural differences, platform constraints, and strengths and weaknesses. Methods: Researchers systematically reviewed Scopus, Web of Science, Google Scholar, and PubMed to locate any ۲۰۲۳-۲۰۲۵ deep-learning malware studies that detected with sufficient methodological details. The selected studies were organized according to their analysis approaches such as static, dynamic, or hybrid; data representations like API sequences, images, graphs, or permissions; and model architectures such as CNN, RNN, GNN, or ensemble models. The study characteristics were the datasets, feature sets, model architecture, computational overheads, interpretability, and performance. Results: Although DL-based systems outperform classical ones in most tasks, hybrid systems are stronger to obfuscation. Many graph-based approaches extract structural information often ignored by classical systems. However, hardware-efficient lightweight and ensemble models proposed in IoT studies achieved accuracy while not exceeding hardware restrictions and power consumption. Difficulties remain with diversifying datasets, explaining models, adapting over time, and resisting adversarial attacks. Conclusion: Deep learning has the potential to continue transforming malware detection if three challenges are overcome: interpretability, datasets, and robustness of multi-modal detection pipelines.

کلیدواژه ها:

Malware Detection (MD) ، Deep Learning (DL) ، Dynamic Analysis ، Static Analysis ، Graph Neural Networks (GNN)