There is an even more advanced solution: homomorphic encryption lets you operate over fully-encrypted, undecryptable data, and compute new derived data from it

There is an even more advanced solution: homomorphic encryption lets you operate over fully-encrypted, undecryptable data, and compute new derived data from it that -- if decrypted with the same key -- is as if the computation was done over the decrypted version to begin with. This can be used to create 'blind' machine learning computer vision models that can detect objects in encrypted data but still can't actually 'see' what's going on.