SOLVING DISTORTED CAPTCHA-TEXT BY USING CONVOLUTIONAL NEURAL NETWORK

• Main Author: Akbar Yasin - NIM: 23216322 , Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/29112
• Institution: Institut Teknologi Bandung

At the early stage of the development, CAPTCHA-text used distorted which difficult to solve by OCR technology. The development of AI technology, machine learning and image processing year after year makes the task to distinguish between human interactions and "bot" becomes more challenging. Recently, more advanced CAPTCHA types are available to avoid the risk of using CAPTCHA-text that can be considered irrelevant anymore to secure a website. However, until now can be found some websites that still use CAPTCHA-text. This paper contains the design of an intelligent "bot" system that can solve distorted CAPTCHA-text along with the results of the experiments that have been carried out. Convolutional Neural Network was chosen as the approach for this study because its performance has proved excellent for object recognition applications. The CNN architecture used for this research consists of three convolutional layers, three pooling layers and two fully-connected layers. The level of accuracy that needs to be achieved is 75% - which based on our considerations that CAPTCHA usually can tolerate incorrect answers or solution for its challenge so the specified accuracy level is considered enough to carry out a successful attack to breach a CAPTCHA security system. From the result of experiment, the system managed to achieve 75% prediction accuracy in ± 19 hours of program execution for one type of distorted CAPTCHAtext which was chosen as the case example in this study. <br />