TEE based model training service

[pstorm]

Will there be TEE based training service for training ML model over private features & non private features? It logically looks like an aggregation service that acts on reports to generate a summary report. However, for a TEE training service, it operates on both user features and non-private features, to generate a model (summary).

B&A Inference explainer pointed out various details about TEE inference service periodically pulling models from currently supported cloud providers (AWS & GCP), providing inference over the input request (composed of private features) coming from the same machine, for larger models.

[chatterjee-priyanka]

Thanks for the question.

yarongmu-google is the lead in Privacy Sandbox, who may be able to help. Thanks.

[yarongmu-google]

Thanks for the question. Yes, TEE based training services is something we are interested in exploring. Used with Differential Privacy, TEE based training has the potential to bring real value here. There are multiple algorithm candidates in this space, for different problem settings.

pstorm, do you mind elaborate your problem setting a little bit more?