Recognizing some of the modern CAPTCHAs Dmitry Nikulin LCME, Saint-Petersburg, 2011

Examples

Stands for Completely Automated Public Turing test to tell Computers and Humans Apart

Turing test Introduced by the mathematician Alan Turing in 1950 Aimed to distinguish between a machine and a human The classic version is carried out by a human Loebner Prize has not been won yet

Reverse Turing Test Carried out by a computer A widespread example is CAPTCHA - Checks for human presence - Protects against spam and automated registrations - Uses human ability to recognize distorted text (Google reCAPTCHA)

Requirements for a CAPTCHA Simple for a human Difficult for a machine Does not require large computational resources Let us call a CAPTCHA efficient if a machine can successfully bypass it in no more than 1% of attempts.

Study the efficience of the widespread CAPTCHAs CAPTCHAs from the largest Russian mobile network operators web sites were chosen Objectives

Reasons of choice Operators have enough money to hire a programmer of any qualification Operators need to minimize the amount of spam in order to safeguard their reputation

Recognition method overview Preprocessing Segmentation Recognition In the following slides details on these stages will be given.

Preprocessing Clearing the noise Removing distortions © Beeline© MTS

Segmentation Extracting characters Post-processing characters

Recognition Classification of characters with a pre-trained neural network

Example Let us consider the following type of CAPTCHA: © Megafon

Analyzing the problem Characters lie on a 3D wireframe The wireframe is rotated and moved The brightness is inconsistent Seems to be quite bad :(

Ideas of the solution Ignore the three-dimensionality and use classic methods The characters are generally darker than the background and can be separated by brightness The upper side of the wireframe is clearly seen – this can be used for the reverse rotation

Estimating the rotation angle

Removing the background

Removing tiny holes

Segmentation

Statistics Total number of images – 100 Recognized successfully – 69 Recognition error – 31 Average error – 0.3 сharacters

Other types of CAPTCHAs Preprocessing varies greatly Segmentation is quite similar Almost identical recognition Conclusion the more transformations are applied to the original image, the more general methods can be used.

Neural network segmentation In Beeline's CAPTCHA, the classic method did not show satisfactory results A new method which combines the segmentation and recognition was developed

Example © Beeline

Conclusion Only preprocessing varies significantly All considered types of CAPTCHAs proved to be inefficient reverse Turing tests

Questions?