By co-training, using split training, only those abstractions, we can compromise no privacy, no personally-protected data, but then together, learn something about potentially illicit behavior. This is maybe one of the best cases for this kind of thing, because when done wrong, it could justify very easily surveillance statism.