Date of Graduation

2018

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Afzel Noore

Committee Co-Chair

Edgar J Fuller

Committee Member

James D Mooney

Committee Member

George E Trapp

Committee Member

Frances L Van Scoy

Abstract

Given the data loss, productivity, and financial risks posed by security breaches, there is a great need to protect online systems from automated attacks. Completely Automated Public Turing Tests to Tell Computers and Humans Apart, known as CAPTCHAs, are commonly used as one layer in providing online security. These tests are intended to be easily solvable by legitimate human users while being challenging for automated attackers to successfully complete. Traditionally, CAPTCHAs have asked users to perform tasks based on text recognition or categorization of discrete images to prove whether or not they are legitimate human users. Over time, the efficacy of these CAPTCHAs has been eroded by improved optical character recognition, image classification, and machine learning techniques that can accurately solve many CAPTCHAs at rates approaching those of humans. These CAPTCHAs can also be difficult to complete using the touch-based input methods found on widely used tablets and smartphones.;This research proposes the design of CAPTCHAs that address the shortcomings of existing implementations. These CAPTCHAs require users to perform different image-based tasks including face detection, face recognition, multimodal biometrics recognition, and object recognition to prove they are human. These are tasks that humans excel at but which remain difficult for computers to complete successfully. They can also be readily performed using click- or touch-based input methods, facilitating their use on both traditional computers and mobile devices.;Several strategies are utilized by the CAPTCHAs developed in this research to enable high human success rates while ensuring negligible automated attack success rates. One such technique, used by fgCAPTCHA, employs image quality metrics and face detection algorithms to calculate a fitness value representing the simulated performance of human users and automated attackers, respectively, at solving each generated CAPTCHA image. A genetic learning algorithm uses these fitness values to determine customized generation parameters for each CAPTCHA image. Other approaches, including gradient descent learning, artificial immune systems, and multi-stage performance-based filtering processes, are also proposed in this research to optimize the generated CAPTCHA images.;An extensive RESTful web service-based evaluation platform was developed to facilitate the testing and analysis of the CAPTCHAs developed in this research. Users recorded over 180,000 attempts at solving these CAPTCHAs using a variety of devices. The results show the designs created in this research offer high human success rates, up to 94.6\% in the case of aiCAPTCHA, while ensuring resilience against automated attacks.

Share

COinS