I think we could use a combination of zero knowledge proofs (zkp), fully homomorphic encryption (FHE), trusted execution environment (TEE) if done very carefully, and/or secure multi-party computation (MPC)

I'm probably forgetting other techniques that could be used to keep inputs, outputs and computation privacy preserving while still being unable to abuse it or deanonymize the dataset