On the complexity of verifying differential privacy