Communication-Efficient Secure Three-Party Neural Network Inference with dropped server

With the widespread application of deep learning, privacy concerns during neural network inference are gradually gaining attention. Secure inference in deep-learning-as-a-service (DLaaS) environments can effectively protect model parameters of model providers and the sensitive data of clients. However, nonlinear computations in secure neural network inference introduce significant communication overhead. In this paper, we present Ctnid, a secure outsourced computing framework that supports lightweight secure neural network inference. Specifically, we design a new secure computation protocol using (2,3)-secret sharing and additive secret sharing to address the dropped server problem. The protocol is then applied to develop secure nonlinear (ReLU and Maxpool) operations in neural networks. Experimental results show that our neural network inference scheme is secure and communication efficient.