Использование доменно-состязательного обучения для распознавания текстовых капч

Abstract

Nowadays the problem of legal regulation of automatic collection of information from sites is being actively solved. This means that interest in tools and programs for automatic data collection is growing and that's why interest in automatic programs for solving CAPTCHA («Completely Automated Public Turing test to tell Computers and Humans Apart») is increasing too. In spite of сreation of more advanced types of captcha, nowadays text captcha is quite common. For instance, such large services as Yandex, Google, Wikipedia, VK continue to use them. There are many methods of breaking text captchas in literature, however, it should be noted that most of them have a limitation to priori know the length of the text on the image, which is not always the case in the real world. Also, many methods are multi-stage, which brings additional inconvenience to their implementation and application. Moreover, some methods use a large number of labeled pictures for training, but in reality, to collect data one has to be able to solve captchas from different sites. Respectively, manually labeling dozens of thousands of examples for training for each new type of captcha is an unrealistically difficult action. In this paper we propose a one-step algorithm of attack on text captchas. This approach does not require a priori knowledge of the text's length on the image. It has been shown experimentally that the usage of this algorithm in conjunction with the adversarial learning method allows one to achieve high quality on real data, using the low number (200-500) of labeled examples for training. An experimental comparison of the developed method with modern analogs showed that using the same number of real examples for training, our algorithm shows a comparable or higher quality, while it has a higher speed of working and training.