Reading notes: Secure federated matrix factorization

Secure Federated Matrix Factorization

August 2020, Intelligent Systems, IEEE 2021/10/3
The article explain ：
Security matrix decomposition under federated learning | 2021 WAIC Paper sharing (qq.com)
https://mp.weixin.qq.com/s/CijeQBouFEQbUVB6cwVjTw.

1. What does literature study?

1. Security federation matrix decomposition , Is the gradient information uploaded by the client easy to leak? Turn over the user's private scoring information .

2. What’s the innovation?

1. Simultaneous interpreting and encryption methods based on different traditional methods , Using homomorphic encryption to increase gradient security, a user level distributed matrix decomposition framework is designed .

3. What was the methodology?

1. utilize SGD Matrix decomposition optimization
2. Prototype implementation , utilize Paillier A prototype of encryption construction .
3. Two experimental settings ：
   1. Upload all item Gradient of .
   2. Still upload score Items Gradient of

4. What are the conclusions?

1. Only the two experimental settings are compared, and the experimental results of encryption are not seen .
2. item Increased consumption of time , Most of it is consumed on the server side , Still stay in theoretical analysis .
3. Need to be in FullText(users upload gradients for all the items) And PartText(user only upload the gradients of the rated items) Find a balance between .

5. others

1. they utilize an open source python package, python-paillier to accomplish the ecryption part in prototype system: https://github.com/n1analytics/python-paillier.
2. By analyzing the code , Can answer my doubts .
   Namely gradient and local Update Information used