Adversarial Attacks on CAPTCHA | Towards AI

Breaking CAPTCHA Using Machine Learning in 0.05 Seconds

Machine learning model breaks CAPTCHA systems on 33 highly visited websites, concept is based on GANs

December 19, 2018 by Roberto Iriondo

Image Source: Machine Learning Memoirs

Everyone despises CAPTCHAs (humans, since bots do not have emotions) — Those annoying images containing hard to read text, which you have to type in before you can access or do “something” online. CAPTCHAs (Completely Automated Public Turing tests to tell Computers and Humans Apart) were developed to prevent automatized programs from being mischievous (filling out online forms, accessing restricted files, accessing a website an incredible amount of times, etc.) on the world wide web, by verifying that the end-user is in fact “human” and not a bot. Nevertheless, several attacks on CAPTCHAs have been proposed in the past, but none has been as accurate and fast as the machine learning algorithm presented by a group of researchers from Lancaster University, Northwest University and Peking University showed below.

Figure 1: Overview of the approach. The researchers first use a small set of non-synthesized CAPTCHAs to train a CAPTCHA synthesizer. (1) the CAPTCHA synthesizer is then used to generate synthetic CAPTCHAs, which at the same time (2) the synthetic CAPTCHAs are used to train a machine learning base solver, (3) which is refined to build a fine-tuned solver of non-synthesized CAPTCHAs. | [1]

One of the first known people to break CAPTCHAs was Adrian Rosebrock, whom in his book “Deep Learning for Computer Vision with Python,” [4] Adrian goes through how he bypassed the CAPTCHA systems on the E-ZPass New York website using machine learning. Where he used deep learning to train his model by downloading a large image dataset of CAPTCHA examples in order to break the CAPTCHA systems.

The main difference between Adrian’s solution and the solution from the research scientists from Lancaster, Northwest and Peking, is that the researchers did not have a need to download a large dataset of images in order to break the CAPTCHAs system, au contraire, they used the concept of a generative adversarial network (GAN) in order to create synthesized CAPTCHAs, along a small dataset of real CAPTCHAs in order to create an extremely fast and accurate CAPTCHA solver.

Generative adversarial networks, introduced by Ian Goodfellow along other researchers [2], are deep neural net architectures comprised of two neural networks, which compete against the other in a zero-sum game [3] in order to synthesize superficially authentic samples. These are especially useful in scenarios where the model does not have access to a large dataset.

Figure 2: Targeted CAPTCHA security features. Sample examples collected from Baidu, Sina, Microsoft and JD captcha schemes. | [1]

The researchers evaluated their approach by applying 33 text-based CAPTCHA schemes, 11 which are currently being used by 32 of the world’s most popular websites ranked by Alexa. Including CAPTCHA schemes being used by Google, Microsoft, eBay, Wikipedia, Baidu and many others. The machine learning model used to attack these CAPTCHA systems only needed 500 non-synthesized CAPTCHAs instead of millions of examples as other attacks before this one (such as Adrian’s) have proposed.

Figure 3: List of text-based captcha schemes used as training data, along testing of the machine learning CAPTCHA solver. | [1]

Once the model was initialized with the CAPTCHAs security parameters in mind shown in Figure 2, it was used to generate a batch of synthetic CAPTCHAs in order to train the synthesizer with the 500 real CAPTCHAs obtained from the diverse CAPTCHA schemes shown in Figure 3. The researchers used 20,000 CAPTCHAs to train the pre-processing model along 200,000 synthetic CAPTCHAs to train the base solver.

The machine learning prototype was implemented using Python, the pre-processing model is built using the Pix2Pix framework, which was implemented using Tensorflow. The fine-tuned solver was coded using Keras. [1]

Figure 4: Real Google CAPTCHAs and the synthetic versions generated by the researchers’ CAPTCHA synthesizer | [1]

After the generative adversarial networks were trained by using the synthesized and real CAPTCHA samples, the CAPTCHA solver was used then to solve CAPTCHAs from highly visited websites, such as Megaupload, Blizzard, Authorize, Captcha.net, Baidu, QQ, reCaptcha, Wikipedia, etc. The impressive approach of this method is that most of the sites CAPTCHAs were solved with over 80% success rate, exceeding 100% on sites like Blizzard, Megaupload and Authorize.net, an attack method that has proven to have better accuracy on all other prior methods to solve CAPTCHAs, which used large non-synthesized training datasets.

Figure 5: Compares the researchers’ CAPTCHA solver against four prior attack methods to solve CAPTCHAs. | [1]

Other than enhanced accuracy, the researchers mentioned on their paper that their approach was not only more accurate, but also more efficient, and less expensive to implement that other methodologies proposed [1]. Besides from being the first GAN based solved for text-based CAPTCHAs, it is an open door for attackers to use, hence their effectiveness and inexpensiveness to implement.

Nevertheless, the approach has some limitations, such as the use of CAPTCHAs with variable numbers of characters, the current approach uses a fixed number of characters — if it’s extended the prototype would break. Another, is the use of variable characters on the CAPTCHA, while the prototype can be trained to support this change, it currently does not as is.

It is important for highly visited websites to use more robust ways to protect their systems, such as bot-detection measures, cyber-security diagnoses and analytics, along multiple layers of security such as device location, types, browsers, etc. — as they are now, an even easier target to attack.

DISCLAIMER: The views expressed in this article are those of the author(s) and do not represent the views of Carnegie Mellon University, nor other companies (directly or indirectly) associated with the author(s). These writings are not intended to be final products, yet rather a reflection of current thinking, along being a catalyst for discussion and improvement.

You can find me on: My personal website, Medium, Instagram, Twitter, Facebook, LinkedIn or through my web design company.

Recommended Stories

References:

[1] Yet Another Text Captcha Solver: A Generative Adversarial Network Based Approach | Guixin Ye, Zhanyong Tang, Dingyi Fang, Zhanxing Zhu, Yansong Feng, Pengfei Xu, Xiaojiang Chen, Zheng Wang | Lancaster University, Northwest University, Peking University | https://www.lancaster.ac.uk/staff/wangz3/publications/ccs18.pdf

[2] Generative Adversarial Networks | Ian J. Goodfellow, Jean Pouget-Abadie , Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair , Aaron Courville, Yoshua Bengio | Department of Computer Science and Operations Research, University of Montreal | https://arxiv.org/pdf/1406.2661.pdf

[3] Zero-Sum Games | Game Theory | Stanford University | https://cs.stanford.edu/people/eroberts/courses/soco/projects/1998-99/game-theory/zero.html

[4] Deep Learning for Computer Vision with Python | Adrian Rosebrock | https://www.pyimagesearch.com/deep-learning-computer-vision-python-book/

[5] Gao, H., Tang, M., Liu, Y., Zhang, P., and Liu, X. Research on the security of Microsoft’s two-layer captcha. IEEE Transactions on Information Forensics & Security 12, 7 (2017), 1671–1685

[6] Gao, H., Wei, W., Wang, X., Liu, X., and Yan, J. The robustness of hollow captchas. In ACM Sig | https://www.lancaster.ac.uk/staff/yanj2/ccs13.pdf

[7] Mohamed, M., Sachdeva, N., Georgescu, M., Gao, S., Saxena, N., Zhang, C., Kumaraguru, P., Oorschot, P. C. V., and Chen, W. B. A three-way investigation of a game-captcha:automated attacks, relay attacks and usability. In ACM Symposium on Information, Computer and Communications Security (2014), pp. 195–206

[8] Yan, J., and Ahmad, A. S. E. A low-cost attack on a microsoft captcha. In ACM Conference on Computer and Communications | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.295.9469&rep=rep1&type=pdf