A NOVEL PRIVACY-PRESERVED RECOMMENDER SYSTEM FRAMEWORK BASED ON FEDERATED LEARNING

A NOVEL PRIVACY-PRESERVED RECOMMENDER SYSTEM FRAMEWORK BASED ON FEDERATED LEARNING

Anonymous authors
Paper under double-blind review
communication efficient fair federated recommender system
ICLR 2023

1. What does literature study?

1. proposed a novel privacy-preserved recommender system framework（PPRSF）to reduce the privacy leakage risk, allow various recommendation algorithms to be applied.

2. What’s the innovation?

1. Past shortcomings

1. a candidate architecture, it cannot avoid collecting open data by curious service providers, while users often choose to open data to obtain better services.
2. a pure P2P system, remove the role of the central server to avoid individuals Centralized storage of information. still leak user data to other users and impose requirements on local devices’ computing power.

2.Innovation:

3. What was the methodology?

1. Recall Layer - Server: sort, generate relatively small candidate sets $Recall\_I_{ik}$for each user, send candidate set to each user.
2. Ranking Layer - Client: local private data and $Recall\_I_{ik}$ 生成fine-sorted candidate item list $Rank\_I_{ik}$符合本地私有数据隐含的偏好，$Ranked\_I_{ik}$(Top-T candicate item list)send back to the server。
3. Re-Rank Layer - Server: 接收$Ranked\_I_{ik}$。
4. Service Layer - Client: 为每个用户呈现最后的推荐结果，同时在这一层记录用户交互行为，为服务器端收集用户公共数据，保存用户相关隐私数据。
   Global Recall Model and Local Ranking Model:
   
   Algorithm 1 Local Ranking Model Training -Client:
   
   Algorithm 2 Global Model Aggregation -Server:
   

4. What are the conclusions?

1. 结论： 将联邦学习范式与推荐系统结合，提出了一种新颖的隐私保护推荐系统框架，可以在保护用户隐私的前提下实现训练和在线推断

5. others

1. 基于算法的隐私保护: the main methods include differential priacy(DP) and encryption algorithms(homomorphic encryption-HE).
2. DP 不能同时保证推荐系统的隐私和推荐准确度。
3. HE 需要消耗大量计算时间、空间以及通信成本。