Mohammed Korayem, Abdallah A. Mohamed, David Crandall and Roman V. Yampolskiy

Captchas are challenge-response tests used in many online systems to prevent attacks by automated bots. Avatar Captchas are a recently-proposed variant in which users are asked to classify between human faces and computer generated avatar faces, and have been shown to be secure if bots employ random guessing. We test a variety of modern object recognition and machine learning approaches on the problem of avatar versus human face classification. Our results show that using these techniques, a bot can successfully solve Avatar Captchas as often as humans can. These experiments suggest that this high performance is caused more by biases in the facial datasets used by Avatar Captchas and not by a fundamental flaw in the concept itself, but nevertheless our results highlight the difficulty in creating Captcha tasks that are immune to automatic solution.

ROC curves for the human versus avatar classification task. Top left: Naive Bayes classifiers, Top right: feature selection and Naive Bayes, Bottom row: LibLinear classifiers.

[papersandpresentations proj=other:avatar]