Malicious Code Family Classification Method Based on Vision Transformer

At present, the classification model based on malicious code gray image mainly takes convolutional neural network as the main framework. These methods will inevitably introduce noise in the process of image visualization due to truncation or filling operation, and the convolutional neural network itself has disadvantages such as easy overfitting and poor noise resistance. In this paper, a malicious code family classification method based on vision Transformer model is proposed, and the improvement of visual mapping is proposed according to the generation process of malicious code image. Lambda attention mechanism is applied to learn the texture position relationship of malicious code image, and generalization verification is carried out on two data sets. The proposed method achieves 99.30% accuracy on Microsoft dataset.