A Novel Image-Based Malware Classification Model Using Deep Learning

Nowadays, the vast volume of data which needs to be evaluated potentially malicious is becoming one of the major challenges of antivirus products. In this paper, we propose a novel image-based mal-ware classification model using deep learning to counter large-scale mal-ware analysis. The model includes a malware embedding method called YongImage which maps instruction-level information and disassembly metadata generated by IDA disassembler tool into an image vector, and a deep neural network named malVecNet which has simpler structure and faster convergence rate. Our proposed YongImage converts malware analysis tasks into image classification problems, which do not rely on domain knowledge and complex feature extraction. Meanwhile, we use the thought of sentence-level classification in Natural Language Processing to establish and optimize our malVecNet. Compared to previous work, malVecNet has better theoretical interpretability and can be trained more effectively. We use 10-fold cross-validation on Microsoft malware classification challenge dataset to evaluate our model. The results demonstrate that our model can achieve 99.49% accuracy with 0.022 log loss. Although our scheme is less precise than the winner's, it makes an orders-of-magnitude performance boost. Compared with other related work, our model also outperforms most of them.