Skip to content

Latest commit

 

History

History
74 lines (39 loc) · 5.25 KB

2018-11-21-privsr.md

File metadata and controls

74 lines (39 loc) · 5.25 KB
Raw

PrivSR

Title: Personalized Privacy-Preserving Social Recommendation

Authors: Xuying Meng, Suhang Wang, Kai Shu, Jundong Li, Bo Chen, Huan Liu, Yujun Zhang

Institution: Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China; University of Chinese Academy of Sciences, Beijing, China; Computer Science and Engineering, Arizona State University, Tempe, USA; Department of Computer Science, Michigan Technological University, Houghton, USA

Conference: AAAI Conference on Artificial Intelligence (AAAI)

Year: 2018

Topic

Differential privacy; Social recommendation

Motivation

A concrete example

Potential privacy leakage:

  • To update , the recommender requires user , who has rated item , to share being calculated from rating and user latent vector . However, when there is an item , on which regards its as non-sensitive and publishes it, the recommender can obtain and , compute , and further obtain sensitive ratings and ;

  • With the exposed non-sensitive ratings, the recommender may conduct reconstruction attack to infer an approximation latent vector , by which 's all ratings may be disclosed;

  • The malicious friend requires user latent vector for social regularization, by which may learn 's ratings by computing .

Privacy-preserving social recommendation problem:

Given the observed values in , the set of friends , a set of sensitive ratings , as well as a set of non-sensitive ratings , we want to infer the missing values in without privacy leakage of .

Approach

The proposed framework

  • Divide the learning process of user latent vectors into small components for each specific user, and utilize objective perturbation to provide privacy guarantee under differential privacy.

  • Divide the ratings into sensitive and non-sensitive ratings, and only attach sensitive ratings with small privacy budgets, i.e. big magnitude noises. In this way, the non-sensitive ratings' modeling will not be significantly affected, which can help retain recommendation effectiveness.

  • Decouple the components of noise perturbation into small pieces each of which can be independently processed by individual users. In this way, each user can decide his/her own noise magnitude locally. The entire process can still satisfy the requirement of differential privacy.

Contribution

  • The first to study the problem of privacy-preserving social recommendation with personalized privacy.

Performance

Dataset:

Comparative approach:

  • MF

  • SoReg

  • DPMF

Metric:

  • Mean Absolute Error (MAE)
    • Recommendation effectiveness comparison
    • Privacy protection comparison
    • Impact of parameters and

Question:

How to set the privacy budget of the derived item latent matrix?