【作者】 马巍；

【导师】 杜瑞颖；

【作者基本信息】 武汉大学 ， 网络空间安全， 2020， 硕士

【摘要】 文本验证码作为一种保护网站免受自动爬虫或恶意流量攻击的安全措施,被广泛应用于互联网中,是网站的重要防护手段之一。而近年来随着神经网络技术的快速发展,将卷积神经网络应用于文本验证码的非分割识别成为一大研究热点。文本验证码防御性能受到的挑战也因此更加严峻。如何在基于卷积神经网络的识别场景下保护文本验证码成为亟待解决的问题。对抗扰动作为一种精心构造的细微扰动,添加到原始样本中,能够使得神经网络将之前能够正确分类的样本错误分类。为了提高验证码的防御能力,本文将对抗样本应用于验证码防御中。但是仅仅简单的将对抗样本生成算法应用于文本验证码中,生成的对抗文本验证码不仅泛化对抗能力差,能够欺骗的模型只有生成模型,同时抗预处理能力也存在不足。因此本文针对该问题设计了一种迭代的对抗生成框架,将多子卷积神经网络与预处理模块结合形成判别模块,在生成模块中本文应用权重显著性图筛选待修改特征点,通过框架内生成模块和判别模块间的博弈,增强验证码的对抗能力。在该框架下生成的对抗验证码,具备较强的对抗泛化能力,可广泛降低多种主流卷积神经网络的识别能力,同时能够抵抗预处理,并不显著降低用户体验,在保证可用性的同时提高验证码对多种主流卷积神经网络自动识别的防御能力。实验结果表明本文提出的框架能够大幅度提高文本验证码针对基于卷积神经网络非分割式识别的防御能力。更多还原

【Abstract】 Text CAPTCHA is widely used in the Internet as a security measure to protect the website from automatic crawler or malicious traffic attack.In recent years,with the rapid development of neural network,the application of convolutional neural network in the non-segmentation recognition of text verification code has become a hot topic.This makes the performance of text captcha defense more challenging.How to protect text CAPTCHA in the context of recognition based on convolutional neural network has become an urgent problem.As a subtle disturbance carefully constructed,the adversarial noise can be added to the original sample to lead the neural network to misclassify the previously correctly classified samples.In order to improve the defense ability of captcha,this paper applies the adversarial examples to text CAPTCHA.However,if the adversarial example generation algorithm is simply applied to the text captcha,the adversarial text CAPTCHA not only has poor generalization ability,but also has insufficient anti-preprocessing ability.So this paper designs an iterative adversarial generative framework,convolution neural networks combined with pretreatment module to form a discriminant module,the weight saliency map are used in generating module to filter which feature point should be modified.By the game of generative module and discriminative modules within the framework,the defense capability of verification code can be enhanced.The adversarial text CAPTCHA generated by the framework have stronger ability of generalization,which can widely reduce the variety of mainstream convolution neural network recognition ability.At the same time,adversarial text CAPTCHA is able to resist pretreatment without significantly reducing the user experience,which keeps human users availability with improving CAPTCHA for a variety of mainstream convolution neural network automatically recognition defenses.The experimental results show that text CAPTCHA can be greatly enhanced by the framework against automatic recognition accuracy based on convolutional neural network.更多还原