[4Xin1] Poster session 2

[4Xin1-55] Vulnerability Verification using Differential Information for Federated Learning-based POI Recommender System

〇Chihiro Yamasaki¹, Kazushi Okamoto¹ (1.The University of Electro-Communications)

Keywords:Privacy, Information Recommend, Point-of-Interest, Federated Learning, Embedding

Point-of-Interest (POI) recommender systems recommend unvisited POIs that are attractive to users based on their check-in (C/I) histories. However, centralized POI recommender systems have privacy problems. PREFER, a federated learning-based POI recommender framework, ensures user privacy by only exchanging user-independent parameters with the server, but it has not been verified that C/I information is not leaked from the exchange parameters. This study verifies the vulnerability of PREFER applied PRME-G as a model by using observed exchange parameters during the model learning. Four experimental datasets, similar in sizes to PREFER's study, were constructed from the Foursquare dataset and were applied to the evaluation experiment. The results indicate that the server can identify which POIs a user has visited from the observed differential information in the parameter exchange, and that much of the user's C/I information is leaked to the server.