A New Dynamically Changing Attack on Review Fraud Systems and a Dynamically Changing Ensemble Defense

Youzhi Zhang, Sayak Chakrabarty, Rui Liu, Andrea Pugliese, V. S. Subrahmanian

Abstract: A number of companies sell "fake" reviews of products for a fee to unscrupulous customers. To do this, they operate a set of sockpuppet accounts. We use the term SockFarm to refer to such companies. We propose SockAttack, a way that such shady companies can judiciously create fake accounts in order to minimize detection. We show that SockAttack compromises the F1-score of 4 well known review fraud detectors more than baselines on real world datasets (up to 28.8%). We then propose SockDef, an algorithm to mitigate the efficacy of SockAttack and show that SockDef mitigates the impact of SockAttack (up to 63.6% w.r.t. F1-score).