The rising complexity of deep neural networks has raised rigorous demands for computational hardware and deployment expertise. As an alternative, outsourcing a pre-trained model to a third party server has been increasingly prevalent. However, it creates opportunities for attackers to interfere with the prediction outcomes of the deep neural network. In this paper, we focus on integrity verification of the prediction results from outsourced deep neural models and make a thread of contributions. We propose a new attack based on steganography that enables the server to generate wrong prediction results in a command-and-control fashion. Following that, we design a homomorphic encryption-based authentication scheme to detect wrong predictions made by any attack. Our extensive experiments on benchmark datasets demonstrate the invisibility of the attack and the effectiveness of our authentication approach.