This paper proposes a method to quantify the privacy risk of training data for recommender systems (RecSys). Existing RecSys utilizes sensitive user-item interaction data, but lacks privacy protection. While users have the right to withhold sensitive interaction information, the problem is that it is difficult to determine which interactions are more sensitive. Therefore, this paper proposes RecPS, a privacy score measurement method based on membership inference attacks (MIA). RecPS measures privacy risk at both the interaction and user levels, and the interaction-level score is derived from the concept of differential privacy. Its core component is RecLiRA, an interaction-level MIA method that provides high-quality membership estimation. Experimental results demonstrate that the RecPS score is effective for risk assessment and untraining RecSys models.
