On the User Behavior Leakage from Recommender System Exposure

Modern recommender systems are trained to predict users’ potential future interactions from historical behavior data. During the interaction process, despite data coming user side, also generate exposure provide users with personalized recommendation slates. Compared sparse data, system much larger in volume since only very few exposed items would be clicked by user. In addition, privacy sensitive and commonly protected careful access authorization. However, large of generated service provider itself usually receives less attention could accessed within a relatively scope various information seekers or even adversaries. this article, we investigate problem leakage field systems. We show that privacy-sensitive past can inferred through modeling exposure. other words, one infer which has just observation current for . Given fact widely scope, believe high risk More precisely, conduct an attack model whose input is recommended item slate (i.e., exposure) while output user’s behavior. Specifically, exploit encoder-decoder structure construct apply different encoding decoding strategies verify performance. Experimental results on two real-world datasets indicate great danger leakage. To address risk, propose two-stage privacy-protection mechanism first selects subset then replaces selected uniform popularity-based evaluation reveals trade-off effect between accuracy disclosure interesting important topic concerns