Spatially Transformed Text-Based CAPTCHAs

In recent years, deep learning technology has achieved great achievements in the fields of text, image, and speech recognition. As long as there is enough data, deep learning techniques often achieve good results. In the field of text-based CAPTCHA recognition, facing the anti-segmentation technology, using deep learning without using character segmentation techniques to end-to-end recognize CAPTCHAs, can also achieve good accuracy. However, some recent studies have found that deep neural networks are vulnerable to adversarial examples. By adding a very small perceptible disturbance to the input samples, the input of the disturbance will cause the model to output incorrect prediction results with high confidence. In this paper, we propose an adversarial text-based CAPTCHAs based on spatial transformation. And we use four state-of-the-art CNN models to recognize such adversarial CAPTCHAs with and without preprocessing. Experiments show that this type of CAPTCHAs can effectively reduce the recognition rate of the attack models.