There are numerous issues with the 'privacy-preserving' federated learning models that are used by Google and Apple. They have constantly been called out by privacy researchers and cryptographers due to their misuse of differential privacy. There is very little trust in the privacy-preserving measures that are being used by these companies. This is not to say that these methods are bad in general but in terms of how they are used by Google and Apple, they amount to little more than a PR stunt. There are numerous reconstruction attacks on their privacy protocols as used by things like GBoard.