Improving Abilityy to verify Audio CAPTCHA's (Semester Unknwon) IPRO 316
MetadataShow full item record
CAPTCHAs (Completely Automated Public Turing Test to Tell Computers and Humans Apart) are used to prevent automated access to sensitive information online. In its usual format, users are presented with distorted text and asked to enter the displayed text in an answer box. If successful, humans, but not computers, will be able to interpret the distorted text. Another format of CAPTCHAs asks users to identify audio information (usually a string of digits or phrase of words) that has been distorted or placed against a background of noise (\white" noise, reversed speech, etc.). Users type the words they hear into an answer box. The audio format is intended to be accessible to blind and low-vision users who cannot use the visually-based format. Unfortunately, audio CAPTCHAs are di cult for humans to use (Bigham and Cavendar 2009) but relatively easy for computers to solve (Tam et al. 2008), which is exactly the opposite outcome desired. To take two extreme examples, in one recent study (Sauer et al. 2008), users were able to solve only 46 percent of audio CAPTCHAs, while in another study (Burztein and Bethard 2009), a computer program was able to break 75 percent of audio CAPTCHAs. At issue is whether audio CAPTCHAs can be designed so that users can easily solve them but computers cannot. Yan and Ahmad (2008) propose testing di erent kinds of background noise to determine which is the most e ective at blocking computers but admitting humans. To this end, Tam et al. (2008) suggest using other human voices as background noise (to thwart computers) but familiar phrases as the string to decode (to aid listeners). The current project will focus on selecting from a set of potential solutions to test.