Used by hundreds of millions of people every day, online services are central to everyday life. Their popularity and impact make them targets of public opinion skewing attacks, in which those with malicious intent manipulate the image of businesses, mobile applications and products. Website owners often turn to crowdsourcing sites to hire an army of professional fraudsters to paint a fake flattering image for mediocre subjects or trick people into downloading malicious software. This research aims to disrupt fraudulent job markets, identify behavioral differences between fraudsters and honest users, and design fraud detection methods for popular crowdsourcing sites such as Yelp and Google Play. Detecting fraudulent information and malicious behavior serves to improve the quality of life of online service users, help reduce the distribution and impact of malware, and protect the credibility of online services.The overarching goal of this project is to significantly increase the complexity and cost of attacks that seek to undetectably influence the ranking and public image of online service subjects. The project consists of two thrusts: modeling and detecting search rank fraud. The modeling thrust dissects fraudulent job markets supported by crowdsourcing sites, documents and models their employer-to-worker interactions and equilibria, and identifies the subjects they target. The researchers will monitor subjects and leverage community expertise and malware detection tools to create large-scale, ground truth datasets in order to assist the effort to identify temporal, spatial and social indicators as well as patterns that differentiate fraudsters from legitimate users. The detection thrust builds upon game theory to develop solutions that disrupt detected, active fraudulent jobs. The researchers will also design scalable search rank fraud detection techniques and algorithms - built on the indicators and patterns identified in the modeling thrust to detect fraudulent, search rank altering behaviors - for online services. The researchers will devise graph-based methods that identify suspicious connections among users and flag multiple colluding accounts. This final aspect uncovers not only targets of large-scale review campaigns but also less popular subjects, whose rank is impacted by even a few fraudulent reviews.
