Hi all,

I am currently researching the topic of Federated ML in industrial applications. While I understand the idea of keeping all the data in a silo private while model training can be federated across several silos, and their updates aggregated in a central server, I do not see how such models can be developed in practice. Model development require steps such as EDA, feature engineering, data quality assessment, etc, which all require some access to the data. Unless each silo has the required skills in-house, how do you go about engineering models without access to the data? Do you simply use anonymised samples? Even those can be subjected to significant regulatory constraints.

I'd be interested in hearing your thoughts.