PrivyNet: A Flexible Framework for Privacy-Preserving Deep Neural Network Training with A Fine-Grained Privacy Control

Massive data exist among user local platforms that usually cannot support deep neural network (DNN) training due to computation and storage resource constraints. Cloud-based training schemes can provide beneficial services, but rely on excessive user data collection, which can lead to potential privacy risks and violations. In this paper, we propose PrivyNet, a flexible framework to enable DNN training on the cloud while protecting the data privacy simultaneously. We propose to split the DNNs into two parts and deploy them separately onto the local platforms and the cloud. The local neural network (NN) is used for feature extraction. To avoid local training, we rely on the idea of transfer learning and derive the local NNs by extracting the initial layers from pre-trained NNs. We identify and compare three factors that determine the topology of the local NN, including the number of layers, the depth of output channels, and the subset of selected channels. We also propose a hierarchical strategy to determine the local NN topology, which is flexible to optimize the accuracy of the target learning task under the constraints on privacy loss, local computation, and storage. To validate PrivyNet, we use the convolutional NN (CNN) based image classification task as an example and characterize the dependency of privacy loss and accuracy on the local NN topology in detail. We also demonstrate that PrivyNet is efficient and can help explore and optimize the trade-off between privacy loss and accuracy.