PETs challenge - model personalization solutions


May I please check if model personalization algorithms are allowed for the competition? That is, instead of having a single shared FL model, can each participating client (federation unit, e.g. a bank) maintain its own customized model which is learned as part of the FL procedure?

Both tracks currently seem to suggest that the final solution should be a single “privacy-preserving federated model trained using your privacy solution” (e.g. Competition: PETs Prize Challenge: Phase 1); I wanted to confirm with you since there has been past research suggesting that model personalization can be significantly better than global methods (i.e. methods with a single shared model), particularly when the clients may be stateful.


Hi @kzliu,

The challenge datasets are not designed for model personalization. However, it is allowed under the rules of the competition and you are welcome to try it as part of your solution.