2-party Asymmetric learning

[TTitcombe]

Feature Description

Implement an asymmetric learning protocol when calculating the ID intersection between parties. See this paper for more information

Is your feature request related to a problem?

Asymmetric learning is the case where one of the parties in vertical federated learning has the majority of data IDs. The major party can learn a great deal about the individuals/entities the minor party holds data on, but the minor party learns almost nothing about the major party's dataset.

Protocols to protect both parties in this scenario include obscuring the intersection of data IDs by adding random IDs to the set sent to each party.

What alternatives have you considered?

None

Additional Context

This may need to implemented upstream by the PSI team.

This issue is should be worked on after

• Integration with syft (i.e. we have worker-to-worker communication in place)
• Robust PSI strategy ( securely sending IDs to and from computational server)

Open questions:

• Should we always do an obscuring method?
• If not, what is the determining factor?
• Should workers be able to agree on using/not using an asymmetric protocol?