It's feasible to design a reinforcement learning based network that uses output from a non-differentiable deepfake detector as a component of the loss function.