Fraudulent web sites are designed to fool users into believing they represent genuine brands, such as banks or e-commerce sites. In doing so, these sites inevitably make use of branded logos which are among the strongest visual marks for a brand and well-established as key "trust cues" to typical users. This thesis explores the efficacy of image-based logo recognition as a tool for detecting or classifying such sites. We show that modern vision algorithms can provide accuracy above 99%, offering a significant new capability for improving response time and reducing overhead in combating such scams. |