Differentially Private Recommender System with Autoencoders
Xiaoqian Liu, Qianmu Li, Zhen Ni, Jun Hou · 2019
In recent years, deep learning has achieved remarkable fruits in a wide variety of domains, such as recommender systems. In addition, privacy preservation is unprecedentedly necessary in today's era. In this paper, we leverage the privacy preservation problem in recommendation with the deep learning model, i.e., autoencoders. In order to predict user preferences in the collaborative filtering way, the work reconstructs user sparse ratings with autoencoders to estimate unobserved user preferences. To further protect user privacy, the Gaussian mechanism is combined in the stochastic gradient descent process to ensure that the training process meets the requirements of approximate differential privacy.