Privacy recently emerges as a severe concern in deep learning, e.g., sensitive data must be prohibited from being shared to the third party during deep neural network development. In this paper, we present Morphed Learning(MoLe), an efficient and secure scheme for delivering deep learning data. MoLe is consisted of data morphing and Augmented Convolutional (Aug-Conv) layer. Data morphing allows the data provider to send morphed data without privacy information, while Aug-Conv layer helps the deep learning developer apply their network on the morphed data without performance penalty. Theoretical analysis show that MoLe can provide strong security with overhead non-related to dataset size or the depth of neural network. Thanks to the low overhead, MoLe can be applied for both training and inference stages. Specifically, using MoLe for VGG-16 network on CIFAR dataset, the computational overhead is 9% and data transmission overhead is 5.12%. Meanwhile the attack success probability for the adversary is less than 7.9 x 10^{-90}.
View on arXiv