"Human or not?  How to tell the difference?"
Dr. Monica Chew
University of California at Berkeley

CAPTCHAs are tests to distinguish humans from machines in an online
environment.  CAPTCHAs mitigate abuse of internet services. For example,
adversaries write software robots to register free email accounts and 
use the
accounts to send spam. CAPTCHAs allow us to distinguish legitimate 
requests
from automated requests.  I implement and measure CAPTCHAs based on text
recognition, image recognition, and collaborative filtering.

I present the best existing text recognition CAPTCHA, BaffleText.  
BaffleText
images includes novel Gestalt-motivated masking degradations to defend 
against
image restoration attacks.  BaffleText is the first CAPTCHA whose 
difficulty is
explicitly parameterizable in order to avoid challenges that are too 
difficult
for human users.  I evaluate BaffleText with a user study to validate 
its
design, and also apply the Mori-Malik text-recognition attack to 
BaffleText
images.  BaffleText proves to be more resistant to this type of attack 
than all
other text-recognition CAPTCHAs tested.

I propose and implement three CAPTCHAs based on naming images, 
distinguishing
images, and identifying an anomalous image out of a set.  Two of these
CAPTCHAs, the anomaly and distinguishing CAPTCHAs, are novel.  I invent 
a new
metric for evaluating CAPTCHAs, implement all three approaches, 
evaluate them
both theoretically and in user studies, and find that anomaly 
identification is
the most promising approach.  I also help expand the literature on the
intersection between HCI and computer security by discussing a number 
of unique
HCI issues that arise in the CAPTCHA evaluation, such as language 
dependence
and cultural knowledge.

I propose a novel class of CAPTCHAs based on collaborative filtering, 
and
outline a general collaborative filtering CAPTCHA using singular value
decomposition.  I test this CAPTCHA scheme using two sources of data 
(jokes and
visual images) and discover several requirements for the security of
collaborative filtering based CAPTCHAs.  Though the results of the 
experiments
are inconclusive, collaborative filtering CAPTCHAs are worthy of further
research.

Bio: Monica Chew finished her PhD in Computer Science at the University 
of California at Berkeley in
December 2004 (http://www.cs.berkeley.edu/~mmc/) .  Her thesis was 
entitled "Automatically Distinguishing Humans and Computers in an 
Online Environment";  her  advisor was Doug Tygar, and her main 
research area is computer security.  Monica also has a B.S. in Computer 
Science and Mathematics and a B.M. in Piano Performance, both  from the 
University of North Carolina at Greensboro as well as  an M.M  in Piano 
Performance from the San Francisco Conservatory.