TomCaT: Target-oriented crowd review attacks and countermeasures

Abstract

Online platforms like Amazon, Yelp, and Regulations.gov give a voice to masses of users through reviews, comments, and ratings. However, this crowd-based feedback is susceptible to manipulation. To tackle this problem, most previous efforts have only indirectly sought to uncover targets of attacks by focusing on manipulation at the review or user level. Instead, this paper focuses on the challenge of countering target-oriented crowd attacks. We introduce a unique ground truth dataset of Amazon products that have been targeted for attack and identify two target-oriented attack patterns: (i) promotion attacks and (ii) restoration attacks. With these attacks in mind, we propose the TOmCAT detection framework based only on the timing and sequencing of product ratings. Although TOmCAT succeeds in uncovering targets of manipulation with high accuracy by addressing existing attacks, strategic attackers potentially can create hard-to-detect behavioral patterns by undermining timing-based footprints. Hence, we further propose a complementary approach to TOmCAT called TOmCATSeq which is resistant against strategic manipulation.