Research on image malware classification method based on AlexNet transfer learning and MHSA attention mechanism

In view of the limitations of traditional deep convolutional neural network models in malware classification, such as low accuracy and low precision, this paper proposes an image malware classification method based on AlexNet transfer learning and MHSA attention mechanism. AlexNet is used as the base network, and ImageNet pre-trained image classification model is initialized. By freezing the underlying weights, fine-tuning the output of the last fully connected layer, and introducing transfer learning as training parameters, the overfitting of the model is prevented. Add an MHSA module to the 10th layer (the 5th convolutional layer) of features to capture long-distance dependencies in malware images. Experiments are conducted on the public malware classification dataset maleVis. Compare the information such as accuracy rate, 2-D dimensionality reduction visualization of semantic features in t-SNE, and image regions of interest in Grad-CAM heat map. The results show that the highest classification accuracy of this method on the dataset maleVis reaches 87.3%, which can effectively improve the classification accuracy and precision of malicious software, verifying its feasibility and effectiveness.