Hristo Paskov : 2012 Security Session

CAPTCHAs, automated tests that distinguish humans from computers, have become the norm for reducing bot-based abuse in online services. Their popularity has inevitably led to a number of attacks, some of the most effective of which are based on machine-learning. We survey several recent papers and demonstrate that attacks are generally based on a two-stage segment and classify paradigm that is tailored to the particular distortions used in the CAPTCHA. The status quo is therefore a game of cat and mouse in which designers are constantly trying to come up with new distortions that crackers have not adapted to. What will be the ultimate result of this game? Can we gain a more general understanding of the differences between humans and machines that can answer this question? I will discuss several deep ideas in machine learning that hold promise in giving us this very insight.

Hristo Paskov is a PhD student at Stanford's Security Lab, advised by Professor John Mitchell and Professor Trevor Hastie. He graduated from MIT in 2010 with a BS and MS in Computer Science. Hristo's main research focuses on regularization networks, deep learning, and large-scale learning problems. He uses CAPTCHAs to guide new machine learning algorithms and to characterize the differences between humans and computers on perceptual tasks.