Comparative Study on Adversarial Captcha with Three Target Detection Models

Abstract

In order to distinguish whether the user is a human or a robot, prevent decryption of password, accounts and so on, three kinds of target detection model training for Adversarial Captcha are proposed. The process of training and evaluate-on of adversarial samples are described in detail in this paper. Depth target detection model is selected as attack model, and Projected Gradient Descent (PGD) algorithm is applied to generate adversarial samples based on three models: YOLOv4, Single Shot MultiBox Detector (SSD) and Faster Region-CNN (R-CNN). The samples are compared with clean samples under undirected attack. The results show that the captcha samples have successfully misled the neural network with high confidence, making them output incorrect results. The algorithm proposed in this paper can be used in many fields in the Internet to verify whether it is human or not.