GFS-CNN: A GPU-friendly Secure Computation Platform for Convolutional Neural Networks

Outsourcing convolutional neural network (CNN) inference services to the cloud is extremely beneficial, yet raises critical privacy concerns on the proprietary model parameters of the model provider and the private input data of the user. Previous studies have indicated that some cryptographic tools such as secure multi-party computation (MPC) can be used to achieve secure outsourced inferences. However, MPC-based approaches often require a large number of communication rounds across two or more non-colluding servers, which make them hard to exploit GPU acceleration. In this paper, we propose GFS-CNN, a GPU-friendly secure computation platform for convolutional neural networks. The following two specific efforts of GFS-CNN have been made by combining machine learning and cryptography techniques. Firstly, We use quadratic activation functions to replace most of the ReLU functions without losing much accuracy, so as to create a mixed linear layer for better efficiency by integrating convolution, batch normalization, and quadratic activation. Secondly, for the rest ReLU functions, we implement the secure ReLU protocol using function secret sharing, enabling GFS-CNN to evaluate the secure comparison function via a single interaction during the online phase. Extensive experiments demonstrate that GFS-CNN is accuracy-preserving and reduces online inference time by 16.4% on VGG-16 models compared to Delphi (USENIX Security’20).