Efficient Malware Classification using Transfer Learning and Stacked Ensemble Techniques

The exponential growth of internet usage and communication devices has led to heightened security vulnerabilities, including the proliferation of malware such as viruses, ransomware, trojans, and spyware. These increasingly sophisticated malware variants pose significant challenges in their detection and classification. The existing visualization-based deep learning approach addresses some of these challenges but often requires extensive computational resources and prolonged training times and is prone to overfitting. This work proposed a transfer learning-based stacked ensemble technique to enhance the efficiency and accuracy of a malware classification model. The six scaled variants of the EfficientNetB0 architecture are selected for their performance on the ImageNet dataset. Their scalability was trained on the Malimg dataset, which comprises 9,339 malware images across 25 categories. Leveraging transfer learning for feature extraction significantly reduced training time and achieved a competitive accuracy of 99.10% within fewer epochs. To further enhance performance, the study employed a stacked ensemble approach by combining the strengths of three high-performing transfer learning models into two ensemble configurations: an average ensemble and a weighted average ensemble. The weighted average ensemble model demonstrated superior performance, achieving a remarkable training accuracy of 99.84% and a validation accuracy of 99.25%. These results underscore the effectiveness of the proposed approach in addressing modern malware classification challenges efficiently.