Gerit Wagner

Development of two novel face-recognition CAPTCHAs

CAPTCHAs are challenge-response tests that aim at preventing unwanted machines, including bots, from accessing web services while providing easy access for humans. Recent advances in artificial-intelligence based attacks show that the level of security provided by many state-of-the-art text-based CAPTCHAs is declining. At the same time, techniques for distorting and obscuring the text, which are used to maintain the level of security, make text-based CAPTCHAs difficult to solve for humans, and thereby further degrade usability. The need for developing alternative types of CAPTCHAs that improve both the current security and the usability levels has been emphasized widely.With this study, we contribute to research through (1) the development of two new face recognition CAPTCHAs (Farett-Gender and Farett-Gender&Age), (2) the security analysis of both procedures, and (3) the provision of empirical evidence that one of the suggested CAPTCHAs (Farett-Gender) is similar to Googles reCAPTCHA and better than KCAPTCHA concerning effectiveness (error rates), superior to both regarding learnability and satisfaction but not efficiency.

Assessing Data Quality - A Probability-based Metric for Semantic Consistency

Abstract We present a probability-based metric for semantic consistency using a set of uncertain rules. As opposed to existing metrics for semantic consistency, our metric allows to consider rules that are expected to be fulfilled with specific probabilities. The resulting metric values represent the probability that the assessed dataset is free of internal contradictions with regard to the uncertain rules and thus have a clear interpretation. The theoretical basis for determining the metric values are statistical tests and the concept of the p-value, allowing the interpretation of the metric value as a probability. We demonstrate the practical applicability and effectiveness of the metric in a real-world setting by analyzing a customer dataset of an insurance company. Here, the metric was applied to identify semantic consistency problems in the data and to support decision-making, for instance, when offering individual products to customers.