This is where zero knowledge federated learning comes in. Unfortunately, this is very much a tomorrow technology (it needs the infrastructure to support it). Why invest in privacy-preserving methods for training machine learning models tomorrow when you can steal users private information today (or even better, bully them into doing so by being the defacto VC that everyone needs to use because of network effects).
Perhaps zero knowledge is a poor choice of terms on my part as it is used in ZKP (as you pointed out). What I meant is “privacy-preserving”.