In recent years, the classification of medical images has become more and more important. With the help of the deep learning approach, the classification accuracy of diabetic retinopathy has been greatly improved, and it has brought great benefits to the residents living in the suburbs. However, the researchers found that these neural network models became extremely vulnerable when confronted with an adversarial example. Specifically, some minor changes to the samples can fail the model immediately. In this paper, we use adversarial training methods instead of traditional training methods to improve the model robustness against adversarial example. Experiments showed that using this method in APTOS data set the adversarial accuracy increased from 43% to 83%.

No abstract

No abstract