Deep learning algorithms are widely used in the field of network traffic classification and have achieved good results. However, the emergence of adversarial attacks has brought a serious threat to its security, and the accuracy of the current mainstream classification algorithms based on convolutional neural network models has been seriously reduced. In response to this, this paper proposes an encrypted traffic classification method that resists gray-scale adversarial attacks in traffic classification. The proposed method constructs a topology graph by extracting traffic interaction information such as packet load length, sending order, direction, and cluster, and transforms the encrypted traffic classification problem into a graph classification problem. Then, this paper uses the classification method based on graph convolutional neural network to learn and classify features. The graph convolutional neural network model can automatically extract features from the input topology and map features to different representations in the embedding space to distinguish different graph structures. The experimental results show that the proposed method can not only avoid adversarial attacks, but also improve the classification performance on public datasets by more than 5% compared with the existing typical methods.