Recognizing Arabic Typed Text-based CAPTCHAs Using Deep Learning Algorithm

Abstract

CAPTCHA (Completely Automated Public Turing test to tell Computer and Human Apart) is one of the important security technologies which aimed to distinguish between a human and a robot. Arabic CAPTCHA recognition refers to the identification of Arabic CAPTCHA characters that are typed textbased CAPTCHA. Arabic CAPTCHA recognition has emerged as a new research area in recent years for the ease of access to Arabic websites. Feature extraction and accurate classification help in achieving increased recognition accuracy. This project combines between cyber security and artificial intelligence fields. The purpose of this project is to compare the efficiency of some classification techniques in extracting distinctive features of all forms of Arabic text-based CAPTCHAs characters and recognize them. That is, this project combines segmentation and recognition processes of Arabic text-based CAPTCHAs. This project is an extension of the previous paper "Evaluating Robustness of Arabic CAPTCHAs”. Since the recognition of Arabic text-based CAPTCHA characters could be done using deep learning techniques after segmentation, this project determines the effectiveness of these techniques in capturing useful information, and therefore, achieving more accurate recognition results. In the first phase, we applied a vertical segmentation method on Arabic text-based CAPTCHA samples to evaluate the robustness of these samples. In the second phase, we applied recognition process on images that were segmented during the first phase by performing some of the deep learning methods (like Convolutional Neural Network (CNN) and Multi-Layer Feed-Forward ANN) besides other machine learning methods like Artificial Neural Networks (ANN), k-Nearest Neighbor (kNN), and Support Vector Machines (SVM) to identify various characters in Arabic CAPTCHA. We have got the best result of accuracy for two databases when applying CNN method on dataset#1 and ANN method on dataset#2.