Skip to content

Latest commit

 

History

History
61 lines (41 loc) · 2.61 KB

README.md

File metadata and controls

61 lines (41 loc) · 2.61 KB
Raw

Federated Private Set Intersection

In this example, we will demonstrate the built-in operator for Private Set Intersection (PSI).

What is PSI?

According to Wikipedia: The Private set intersection is a secure multiparty computation cryptographic technique that allows two parties holding sets to compare encrypted versions of these sets in order to compute the intersection. In this scenario, neither party reveals anything to the counterparty except for the elements in the intersection.

psi.png

What's the use cases for PSI in federated learning?

There are many use cases for PSI, in terms of federated machine learning, we are particularly interested in the following use cases:

  • Vertical Learning -- User ids matching user_id_match.png

  • Vertical Learning -- feature overlapping discovery Site-1 : Feature A, B, C, D Site-2: Feature E, A, C, F, X, Y, Z Overlapping features: A, C

  • Federated Statistics -- categorical feature distinct values count feature = email address discover : how many distinct emails in the email addresses feature = country discover: how many distinct countries

    site-1: features: country. total distinct country = 20 site-2: features: country, total distinct country = 100 site-1 and site2 overlapping distinct country = 10
    Total distinct countries = 20 + 100 - Overlapping countries = 120-10 = 110

In federated statistics use case, the PSI will be used inside the Federated Statistics operations. For Vertical Learning user id matching use case, we can directly do PSI calculation

PSI Protocol

There are many protocols that can be used for PSI.

For our implementation in nvflare/app_opt/psi, the Private Set Intersection (PSI) protocol is based on ECDH, Bloom Filters, and Golomb Compressed Sets PSI algorithm.

The algorithm is developed by openmined PSI for two-party PSI.

We took the two-party direct communication PSI protocol and extended to Federated Computing setting where all exchanges are funneled via a central FL server. We supported multi-party PSI via pair-wise approach.

In current implementation, we only support the vertical learning use cases. The federated statistics use case will be addressed in the future release.

Please refer to psi protocol

Examples

Please make sure you set up virtual environment follows example root readme

User email match