Glyph: Fast and Accurately Training Deep Neural Networks on Encrypted Data

Big data is one of the cornerstones to enabling and training deep neural (DNNs). Because of the lack of expertise, to gain benefits from their, average users have to rely on and upload their private data to big data they may not trust. Due to the compliance, legal, or privacy, most users are willing to contribute only their encrypted data, lack interests or resources to join the training of DNNs in cloud. To train DNN on encrypted data in a completely non-interactive way, a recent work a fully homomorphic encryption (FHE)-based technique implementing all in the neural network by -Gentry-Vaikuntanathan(BGV)-based lookup tables. However, such inefficient lookup-table-based significantly prolong the training latency of privacy-preserving. In this paper, we propose, Glyph, a FHE-based scheme to fast and accurately DNNs on encrypted data by switching between TFHE (Fast Fully Homomorphic over the Torus) and BGV cryptosystems. Glyph uses-operation-friendly TFHE to implement nonlinear activations, while adopts-arithmetic-friendly BGV to perform multiply-accumulation (MAC). Glyph further applies transfer learning on the training of DNNs to the test accuracy and reduce the number of MAC operations between and ciphertext in convolutional layers. Our experimental results Glyph obtains the state-of-the-art test accuracy, but reduces the training by $99\%$ over the prior FHE-based technique on various encrypted.