Breaking CAPTCHA
CS-134 Project Proposal
Robin Chhetri and Radhika Bhasin
INTRODUCTION
A CAPTCHA, or the Completely Automated Public Turing test to tell Computers and Humans Apart, is defined as a program that can create and grade tests that most humans can pass, but computers cannot[2] . CAPTCHAs are widely used by websites to distinguish between bots and human users. Captchas present a user with a simple test such as matching characters or digits with what they type or hear from the website. The image or sound is usually distorted to make it difficult for a machine to perform the test. When successful, Captchas can prevent a wide variety of abuses, such as invalid account creation and spam comments on blogs and forums. Captchas are intended to be easy for humans but not so easy for machines.
MOTIVATION
By studying the necessary skills required to break CAPTCHAs, we can better understand how to build stronger CAPTCHAs in the future. Our goal is to use machine learning techniques to interpret or read text from CAPTCHA images. From the time Louis von Ahn et al. [3] proposed CAPTCHAs, there has been a cat and mouse game between those who want to break CAPTCHAs and those who want to create better CAPTCHAs. Recent CAPTCHAs include non-dictionary words, complex words, images, and sounds. There are many research groups [4] working in either defeating or strengthening CAPTCHAs. We will first implement the method proposed by Mori et al. [8] After that, we will modify his method by using probabilistic models. We will be investigating Relevance Vector Machines (RVM), which is a Bayesian Learning method, as the classifier to interpret Captchas. We will also search for other methods including SVM and HMM to improve on their output.
METHOD
We will build a training set of images of characters, and use canny edge detection to get edges of the characters. This creates a outline of the CAPTHCA image.Each of the characters in the alphabet will be represented by shape context descriptors.We will determine the candidate location based on the method proposed by Mori et al. [8] using voting methods and we will be using RVM as the classification method. At last, the final word is determined by combining the guesses for each of the character in test image.
DATA SETS
We will be using Captchas data set from below resources:
PROJECT PLAN
RELATED WORK
References:
[1] Title image, http://www.e-ignite.co.uk/blog/wp-content/uploads/2008/03/captcha_banner.jpg
[2]Von Ahn, L., Blum M., and Langford J.. Telling Humans and Computers Apart (Automatically). CMU Tech Report CMUCS, 2002
[3] Von Ahn, L. and Blum, M. and Hopper, N. and Langford. J., CAPTCHA: Using Hard AI Problems for Security. Advances in Cryptology, 2003
[5] http://www.cs.cmu.edu/~guestrin/Class/10701/projects.html#image
[6] http://jcaptcha.sourceforge.net/
[7] Chellapilla, K. and Simard, P., Using Machine Learning to Break Visual Human Interaction Proofs (HIPs), Advances in Neural Information Processing Systems, 2004
[8] Mori, G. and Malik, J., Recognizing Objects in Adversarial Clutter: Breaking a visual CAPTCHA, In Proceedings: IEEE Computer Society Conference on Computer Vision and Pattern Recognition,2003
[9] Moy, G. and Jones, N. and Harkless, C. and Potter, R., Distortion estimation techniques in solving visual CAPTCHAs,Computer Vision and Pattern Recognition, 2004
[10] Yan, J. and El Ahmad, A.S., Breaking Visual CAPTCHAs with Naive Pattern Recognition Algorithms, In Proceedings: Computer Security Applications Conference, pages 279-291, 2007
[11] Yan, J. and El Ahmad, A.S., A Low-cost Attack on a Microsoft CAPTCHA, In Proceedings: ACM Conference on Computer and Communications Security,pages 543-554, 2008
