Security analysis of image CAPTCHA using a mask R-CNN-based attack model

Abstract

The CAPTCHAs are attacked by automated programs to break their underlying design principle. Therefore, analysing the robustness of CAPTCHA is the critical requirement. Recently, a neural style transfer-based image CAPTCHA called style area CAPTCHA (SACAPTCHA) has been reported. Though the security of SACAPTCHA is evaluated on R-CNN and FCN attack models using accuracy metrics, they are ineffective to analyse its robustness. Therefore, we propose a mask R-CNN-based attack model to critically analyse the robustness of SACAPTCHA. The proposed model performs a shape-wise analysis to test the usability of different shapes and quantifies the model performance using the F1-score. The simulation results show the highest F1 score of 0.962 and 0.828 for star and circle shapes in dataset-1 and dataset-2 respectively. The results show that model prediction is independent of the regularities of the shape. The observations prove that SACAPTCHA is vulnerable to object detection attack even after using irregular shapes.