Publication

Abstract : Malware remains a critical cybersecurity challenge, with increasing diversity and sophistication, further complicated by adaptive obfuscation and evasion techniques. Although static and dynamic detection techniques have proved useful, they meet considerable hurdles in resistance, universality, and resilience against contemporary threats. Recent studies have explored data-driven and image-based malware analysis; however, performance limitations persist in handling class imbalance and detecting unknown malware families. This article suggests an end-to-end deep learning pipeline that converts malware binaries to grayscale images and utilizes state-of-the-art convolutional neural networks with focal loss and focused data augmentation. Results from a 25-family, 20,000-sample benchmark show a best classification accuracy of 99.35%, surpassing previous static and hybrid solutions. The system demonstrates strong generalization, validated through consistent performance on unseen malware families, and shows scalability by successfully training multiple deep architectures without performance degradation. The approach remains constrained by its reliance on static analysis and limited effectiveness against highly obfuscated malware.