PrivNet: Safeguarding Private Attributes in Transfer Learning for Recommendation

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4506–4516 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 4506 PrivNet: Safeguarding Private Attributes in Transfer Learning for Recommendation Guangneng Hu HKUST / Hong Kong, China njuhgn@gmail.com Qiang Yang HKUST / Hong Kong, China qyang@cse.ust.hk Abstract Transfer learning is an effective technique to improve a target recommender system with the knowledge from a source domain. Existing research focuses on the recommendation per- formance of the target domain while ignores the privacy leakage of the source domain. The transferred knowledge, however, may unin- tendedly leak private information of the source domain. For example, an attacker can accu- rately infer user demographics from their his- torical purchase provided by a source domain data owner. This paper addresses the above privacy-preserving issue by learning a privacy- aware neural representation by improving tar- get performance while protecting source pri- vacy. The key idea is to simulate the attacks during the training for protecting unseen users’ privacy in the future, modeled by an adversar- ial game, so that the transfer learning model becomes robust to attacks. Experiments show that the proposed PrivNet model can success- fully disentangle the knowledge benefitting the transfer from leaking the privacy. 1 Introduction Recommender systems (RSs) are widely used in everyday life ranging from Amazon product- s (Zhou et al., 2018; Wan et al., 2020) and YouTube videos (Gao et al., 2010; Cheng et al., 2016) to Twitter microblogs (Huang et al., 2016) and news feeds (Wang et al., 2018a; Ma et al., 2019b). RSs estimate user preferences on items from their his- torical interactions. RSs, however, cannot learn a reliable preference model if there are too few in- teractions in the case of new users and items, i.e., suffering from the data sparsity issues. Transfer learning is an effective technique for alleviating the issues of data sparsity by exploiting the knowledge from related domains (Pan et al., 2010; Liu et al., 2018). We may infer user pref- erences on videos from their Tweet texts (Huang and Lin, 2016), from movies to books (Li et al., Figure 1: t-SNE projection of transferred representa- tions of users with (left) and without (right) training of PrivNet on the MovieLens-Gender dataset. (see Sec- tion 5.5.1 for details) 2009), and from news to apps (Hu et al., 2018, 2019). These behaviors across domains are differ- ent views of the same user and may be driven by some inherent user interests (Elkahky et al., 2015). There is a privacy concern when the source do- main shares their data to the target domain due to the ever-increasing user data abuse and priva- cy regulations (Ramakrishnan et al., 2001; Yang et al., 2019b). Private information contains those attributes that users do not want to disclose, such as gender and age (Jia and Gong, 2018). They can be used to train better recommendation models by alleviating the data sparsity issues to build better user profiles (Zhao et al., 2014; Cheng et al., 2016). Previous work (Weinsberg …

Content truncated to 500 words