Optimized Capsule Network Model for Malware Classification

Over the last few years, the malware industry has grown rapidly, becoming more structured and more complex. Modern malware uses sophisticated methods like code obfuscation, packing, and signature evasion to get around antivirus software. Polymorphic and metamorphic malware together with other evasive strategies, present a significant obstacle to cybersecurity. The development of highly advanced AI models to tackle these problems is essential. This paper presents a deep capsule neural network architecture for detection and classification of malware from an image dataset. In the first place, malware executables are provided as grayscale images, which are then subjected to feature extraction using EfficientNetB0 models. Capsule networks are then applied to catch spatial relationships within the images for classification. Separated from other advanced deep learning architectures, the proposed model exhibits superior performance, attaining a classification test accuracy of 98.25%. Moreover, it has high specificity and sensitivity with scores of 99.18% and 97.95% respectively, which were tested using the Microsoft Malware Classification Challenge (MMCC) dataset.