Aiming at the problems of insufficient accuracy， high prediction time cost and weak ability against obfuscation of the traditional malicious code classification methods， a malicious code classification method based on improved MobileNetV2 was proposed. Firstly， aiming at the problems of malicious code encryption and obfuscation， the Coordinate Attention （CA） method was used to introduce a wider range of spatial locations to enhance malicious code image features. Then， aiming at the problem of high training cost caused by training from scratch， the Transfer Learning （TL） was used to improve the MobileNetV2 learning method to increase the ability against obfuscation. Finally， aiming at the large computational load and slow convergence of traditional deep learning networks， the MobileNetV2 lightweight convolutional network model was used， and Ranger21 was combined to improve the training method to promote rapid convergence. Experimental results show that the above-mentioned method has the accuracy achieved 99.26% and 96.98% for Malimg dataset and DataCon dataset. The method has the accuracy increased by 1.49% and the detection efficiency increased by 45.31% on the Malimg dataset compared with the AlexNet method， and has the accuracy increased by 1.14% on the DataCon dataset compared with the ensemble learning method. It can be seen that the improved MobileNetV2 based malicious code classification method can improve the generalization ability， ability against obfuscation and classification efficiency of the model.