Federated aspects for Track B

Had a quick question about Track B. In the code provided, there is both a graph network model and a logistic regression model. When proposing a federated architecture for Track B, do we need to build a federated architecture for both the graph network and the logistic regression? Or does each local site create its own graph network and the logistic regression part is the aspect that needs to be done in a federated way? Was having some difficulty deciphering this from the readme and description pages.

Thanks!

Hi @surajraj99. This challenge is looking for end-to-end privacy-preserving solutions. That means that the choice and design of machine learning model is part of what you should be developing in your solution. The provided graph neural network and logistic regression models are meant as examples for getting started. You may choose to incorporate them, build off of them, or go in entirely different directions for the machine learning component of your solution.

Hi Jay, I had a follow up question.
Based on the challenge prompt and the webinars I had a slightly different understanding and wanted to confirm something. My understanding was that if the proposed solution was a preliminary data transformation step (a new kind of differential privacy) and as of now can be applied with minimal constraints to various different models, we would not have to also try and design the most effective model architecture. Instead the goal would be to show that the performance degradation for proposed solution would be as minimal as possible while also guaranteeing privacy under specific threat profiles.

However, based off your statement “that the choice and design of machine learning model is part of what you should be developing in your solution”, will we be docked points if we don’t focus much on the model architecture in our white paper but instead use the provided template models?

Hi @prathic, the considerations you mentioned are all things that will be considered in the evaluation. The strength of the privacy aspects of your solution will make up the largest part of how your solution is evaluated, but the accuracy of the solution is also considered. In evaluating the accuracy, both the privacy–accuracy trade-off (how does accuracy degrade with more privacy) as well as the absolute level of accuracy will be part of the evaluation.

I encourage you to review these two areas of documentation on evaluation criteria:

1 Like