Designing a Text-based CAPTCHA Breaker and Solver by using Deep Learning Techniques

Text-based CAPTCHAs are most commonly used by various websites to distinguish between humans and computers. It is used as a security measure. This work consists of a dynamic approach that is proposed to predict Text-based CAPTCHAs that challenges the supposition that they cannot be solved by computers. Three types of CAPTCHAs namely Rotated, Noisy Arc, Complicated Background have been taken. The CAPTCHAs are pre-processed based on their type. Different pre-processing techniques like Erosion, Dilation, Binarization are used to remove the noise from the CAPTCHA. The pre-processed CAPTCHAs are then fed to the Convolutional neural network (CNN) which generates a feature vector. This feature vector is then passed to the long short term memory (LSTM) which generates a sequence of characters. This sequence is displayed as outcome to the user. The dataset for Rotated, Noisy Arc, Complicated Background CAPTCHAs consisted of 9,955, 1,070 and 1,000 images respectively. The model was also tested for CAPTCHAs involving a combination of different resistance mechanisms. The model was able to predict Rotated CAPTCHAs with an accuracy of 85.97%, Noisy Arc CAPTCHAs with an accuracy of 84.52% and Complicated BackgroundCAPTCHAswithanaccuracyof82.91 %.