Evasive PDF Malware Detection using Deep Learning with EffiecientNet

Abstract

This study presents a ground-breaking deep learning methodology that harnesses transfer learning in conjunction with EfficientNet, a leading-edge convolutional neural network model, to detect evasive Portable Document Format (PDF) malware. The model is trained on an extensive dataset of 31,006 raw PDF files, a resource uncharted in prior studies, and these PDFs are uniquely transfigured into images to amplify the discriminatory prowess of the deep learning model. Despite the process being laborious, computationally demanding, and seemingly opaque, it delivers stellar results, showcasing an accuracy of 98.87% and an F1-score of 99.20% on the test set, which is reflective of the model's exceptional precision and recall capabilities. The striking performance consistently surpasses that of traditional malware detection techniques, which de-pend on hand-crafted features and are more prone to manipulation by sophisticated malware. The proposed method, by obviating the manual extraction of features, simplifies the detection procedure and diminishes potential human error. The remarkable results of this study attest to the potential and supremacy of transfer learning and convolutional neural networks in malware detection, with the success of the model credited to EfficientNet's depth-wise separable convolu-tions and compound scaling methodology that enhance efficiency and curtail computational costs. The technique of transfer learning, a central part of the approach, endows the model with the ability to leverage pre-existing knowledge, thereby accelerating the training process. The ground-breaking methodology introduced in this study sets a fresh precedent in the realm of ma-chine learning-based malware detection and opens up new vistas for future research, while its implications extend beyond malware detection to encompass broader cybersecurity concerns like intrusion detection, phishing detection and ransomware identification.