Deep Learning for Malware Detection: TransformerBased Analysis of Windows Executables

Abstract

Malware detection is challenged by unprecedented threats exceeding 560,000 new variants daily. A Transformerbased deep learning approach for detecting malicious Windows Portable Executable files is presented. Raw byte sequences are processed using self-attention mechanisms, eliminating manual feature engineering. A multi-modal framework combining raw bytes, PE metadata, and entropy features is implemented. Through experimental evaluation on EMBER2024 (3.2M samples) and SOREL-20M (20M samples), 95.1% accuracy with 0.4% false positive rate is demonstrated, significantly outperforming baseline methods including LightGBM (92.7%), MalConv (93.4%), and LSTM approaches (91.8%). Explainability for forensic analysis is provided through attention visualization.