
Research Article
VRC-GraphNet: A Graph Neural Network-Based Reasoning Framework for Attacking Visual Reasoning Captchas
@INPROCEEDINGS{10.1007/978-3-031-64948-6_9, author={Botao Xu and Haizhou Wang}, title={VRC-GraphNet: A Graph Neural Network-Based Reasoning Framework for Attacking Visual Reasoning Captchas}, proceedings={Security and Privacy in Communication Networks. 19th EAI International Conference, SecureComm 2023, Hong Kong, China, October 19-21, 2023, Proceedings, Part I}, proceedings_a={SECURECOMM}, year={2024}, month={10}, keywords={captchas visual reasoning object detection graph neural network features extraction}, doi={10.1007/978-3-031-64948-6_9} }
- Botao Xu
Haizhou Wang
Year: 2024
VRC-GraphNet: A Graph Neural Network-Based Reasoning Framework for Attacking Visual Reasoning Captchas
SECURECOMM
Springer
DOI: 10.1007/978-3-031-64948-6_9
Abstract
Captchas are widely used to distinguish between human and machine programs and protect computers from malicious attacks. However, with the development of image recognition and deep learning techniques, the attack success rate of traditional text-based and image-based captcha is getting higher. This leads to increasing demand for more secure captcha. In recent years, some captcha service providers such as Tencent, NetEase, Geetest, etc. put forward novel visual reasoning captchas to improve the safety level, and reduce the risk of attacks. There has been little research on this kind of novel captcha. Existing method mainly uses modular method to break it, but has to train separately and is still insufficient for reasoning task. In order to solve above challenges for visual reasoning captcha, this paper introduces a novel end-to-end graph reasoning network to crack the visual reasoning captcha for the first time. We use object detection model to identify all the objects in the captcha. Then, we extract the distribution of question attention and image features to build a graph neural network. Finally, an end-to-end reasoning framework for attacking visual reasoning captcha is constructed by using reasoning module to integrate multi-modal. We achieve a higher success rate of attack on some very popular visual reasoning captchas. The results will provide technical and theory support for the security evaluation of captchas, and promote research on more secure captchas.