But what is meant by “integrity of the model, inputs and outputs”? I guess I don’t understand the attack vector, what’s the threat here? Someone messes with the model file or refines a model towards a specific malicious bias?