StyleCAPTCHA: CAPTCHA Based on Stylized Images to Defend against Deep Networks

Abstract

CAPTCHAs are widely deployed for bot detection. Many CAPTCHAs are based on visual perception tasks such as text and objection classification. However, they are under serious threat from advanced visual perception technologies based on deep convolutional networks (DCNs). We propose a novel CAPTCHA, called StyleCAPTCHA, that asks a user to classify stylized human versus animal face images. StyleCAPTCHA creates each stylized image by combining the content representations of a human or animal face image and the style representations of a reference image. Both the original face image and the style reference image are hidden from the user. To defend against attacks using DCNs, the StyleCAPTCHA service changes the style regularly. To adapt to the new styles, the attacker has to repeatedly train or retrain her DCNs, but since the attacker has insufficient training examples, she cannot train her DCNs well. We also propose Classifier Cross-task Transferability to measure the transferability of a classifier from its original task to another task. This metric allows us to arrange the schedule of styles and to limit the transferability of attackers' DCNs across classification tasks using different styles. Our evaluation shows that StyleCAPTCHA defends against state-of-the-art face detectors and against general DCN classifiers effectively.