Structural-MRI based autism spectrum disorder classification is important nowadays for early-stage screening and accurate diagnosis of autism spectrum disorder. However, due to the limitation of large data noise and small data volume, the predictive accuracy of structural-MRI based autism classification is not ideal. In this study, a new data augmentation model is proposed to improve classification accuracy without increasing training data volume. The UM1 dataset of ABIDE I for autism study was used for training and evaluating the proposed model. 78 samples of the UM1 dataset was selected randomly for performing the experiment, and the performance improved using proposed data augmentation algorithm for ASD classification task was calculated. Such a procedure was repeated for 500 times to collect results that are statistically significant. Based on the results, this method can stably improve classification accuracy between 10% and 20% in 494 out of 500 experiments (over 98% of the experiments). Through a comparative study of accuracy improvement and label change ratio, the problem of data label noise was explored.