Evaluation of captchas.
| Title | Evaluation of captchas. |
|---|---|
| Summary | Evaluation of liveness detection provided by ocr captchas |
| Keywords | |
| TimeFrame | |
| References | |
| Prerequisites | |
| Author | |
| Supervisor | Josef Bigun, Kevin Hernandez Diaz |
| Level | Master |
| Status |
Intelligent vision systems provide automatic
services in an increasing number of domains. One of the oldest is
automatic text recognition, also known as Optical Character
Recognition, OCR, was introduced in 1970'ies. For many people it became synonym with "scanning"
as it has been available as standard with most photocopy machines
since two decades and with widely available software, e.g. Adobe Acrobat.
Captchas, on the other hand have been relatively recently deployed as part of automatic identity management systems in a variety of applications such as automatic Schengen Visa appointment systems in embassies, to make sure that valuable resources are not abused. An important subset of captchas, called here ocr captchas, rely on recognition of noise corrupted digit and letter sequences in the latin alphabet. Here it is assumed that humans are superior to machines to read them to the effect that such captchas are in fact used to decide if someone who requests a valuable service is effectively a human or a machine, also called liveness detection in biometric identification. In the latter case, the system automatically denies the requested service.
The thesis work will evaluate ocr captchas used in liveness detection for web services. It will study how well modern CNN algorithms can be used to detect purposely corrupted digit and letter sequences used in ocr captchas, since such an algorithm will then be a threat to a valuable service protected by captchas.
The problem is suggested by Bigsafe Technology AB.
