Enhancing the security of image CAPTCHAs through noise addition

Abstract

Text based CAPTCHAs are the de facto method of choice to ensure that humans (rather than automated bots) are interacting with websites. Unfortunately, users often find it inconvenient to read characters and type them in. Image CAPTCHAs provide an alternative that is often preferred to text-based implementations. However, Image CAPTCHAs have their own set of security and usability problems. A key issue is their susceptibility to Reverse Image Search (RIS) and Computer Vision (CV) attacks. In this paper, we present a generalized methodology to transform existing images by applying various noise generation algorithms into variants that are resilient to such attacks. To evaluate the usability/security tradeoff, we conduct a user study to determine if the method can provide “usable” images that meet our security requirements – thus improving the overall security provided by Image CAPTCHAs.