Packing Induced Bias in Deep Learning Malware Classifiers: A Systematic Experimental Study

Packing is a highly used malware evasion technique that compresses, encrypts, or obfuscates executable content, significantly altering the structural characteristics that static malware detectors rely on. While prior work has studied the impact of packing on traditional feature based classifiers, limited attention has been given to understanding how modern deep learning based static detectors behave under realistic packing conditions. This paper investigates the robustness and generalization capabilities of a CNN based malware detector trained on image representations of Windows Portable Executable (PE) binaries by conducting two controlled experiments. The first evaluates how increasing exposure to packed benign ware affects a model's ability to correctly distinguish packed malware from packed benign binaries. Results show initial improvement in true negative rates, followed by instability as packing artifacts dominate learned representations. The second experiment analyzes cross packer generalization and demonstrates strong packer dependency, where models perform well only on packers seen during training and collapse when confronted with unseen packers. Overall, our findings demonstrate that packing significantly undermines the reliability of static deep learning based malware detectors, highlighting the need for packing aware training strategies and more resilient detection models.