CAPTCHA Recognition using Machine Learning

Abstract

This project studies the feasibility of decoding visual CAPTCHAs using machine learning technique. Specifically, Support Vector Machine (SVM) is employed for classifying segmented CAPTCHA images into characters. To demonstrate the effectiveness of my approach, the login system of EE FYP website is selected as the target. 500 CAPTCHAs from the website were segmented and labeled to form the training data, which was then fed into the system for training and performance evaluation. Although generic CAPTCHA decoding is still unsolved due to the great variations of noises and distortions, this system shows a sufficient high accuracy and can be practically used to bypass the target CAPTCHA system. It is also found that with a different segmentation algorithm, CAPTCHAs from other systems can also be recognized by the classifier without specific training despite the difference in fonts. The results suggest that most CAPTCHAs are vulnerable to targeted attacks. Since most failures were caused by segmentation errors, mature solutions with non-easily segmentable CAPTCHAs such as Google reCAPTCHA are recommended.