Yuanshun Yao, Bimal Viswanath, Jenna Cryan, Haitao Zheng, Ben Y. Zhao

Using Yelp reviews as an example platform, we show how a two phased reviewgeneration and customization attack can produce reviews that areindistinguishable by state-of-the-art statistical detectors. We conduct asurvey-based user study to show these reviews not only evade human detection,but also score high on “usefulness” metrics by users. Finally, we develop novelautomated defenses against these attacks, by leveraging the lossytransformation introduced by the RNN training and generation cycle. We considercountermeasures against our mechanisms, show that they produce unattractivecost-benefit tradeoffs for attackers, and that they can be further curtailed bysimple constraints imposed by online service providers.