Flower makes no sense with our architecture

We noticed that the rules for uploading the solution require us to use Flower. Our solution is based on HPE Swarm Learning. Swarm Learning is a commercial product that competes with Flower. Its design philosophy differs from Flower - it uses a decentralized approach, unlike Flower’s client-server approach. As such, it is not easy to integrate Swarm Learning with Flower. So, can this requirement be removed?

3 Likes

Thank you for this note as we are in the same situation … our solution is based on an approach that does not need Flower and “fitting” our solution “into” Flower is a non-trivial question … we echo that we would like to see if this particular requirement can be removed

1 Like

On Slack there are more people with the same problem. The current advice is to describe your problem specifically in a private message to the organizer. If enough people will propose this, hopefully this can change the mind of the organizers.

Hi all,

Here is the guidance provided in the NIST Slack as sorrge mentioned:

Thanks for providing your feedback. If you have concerns or questions about your solution being compatible with the Flower framework, please provide specific and concrete explanations of the problems. If your explanations involve sensitive details about your solutions, please email it to info@drivendata.org. Specific and concrete details will help us work with you to get your solution implemented successfully.

We plan to start an FAQ to address common concerns. While there are currently no plans to revise the API specifications for submissions, specific and concrete feedback is the best way to provide the challenge organizers with the necessary information to make sure the challenge design meets its objectives.


With respect to centralized vs. decentralized federated learning, here is the response to another participant that was made in the NIST Slack:

We are requiring all solutions to be implemented using the centralized federation structure with a server (that is running the strategy). Direct peer-to-peer communication is not supported between clients. However, you can route client-to-client communications through the server. If routing messages through the server has privacy differences compared to peer-to-peer communications, please discuss it in your technical paper.

If you have additional questions, concerns, or feedback, the challenge organizers are interested in hearing from you. Please follow the above request for specific and concrete explanations.

It doesn’t make sense to force all teams to use Flower, which is not friendly to new algorithmic customization. Please remove such a requirement, otherwise we have to quit the challenge. Sorry about this.

1 Like

Dear Organizers,

Forcing all participants to use Flower is not sensible for this competition because there was no statement on which library to use while we were proposing our ideas in Phase 1. Just as in Kaggle, where participants have no restrictions over Python libraries, PETS should not have regulations over federated learning libraries. Besides, with Flower, implementing our algorithms is hard and nearly impossible, especially for the pandemic forecasting challenge. Plus, it requires time for people who do not know Flower which creates inequalities between the participants.

Summary

Please remove this constraint; otherwise, we have to quit the challenge.

1 Like

Our team has already dropped out because of the Flower requirement. We joined the challenge because it aligned with our goal of expanding the privacy-preserving and security features in our current Swarm Learning product. We found no value in the recoding effort to meet the Flower requirement.
We appreciate the opportunity and will look for other challenges that complement our goals. Best of luck to all!

1 Like