Towards Reverse-Engineering Black-Box Neural Networks