Nappa: NNA-Compatible and Privacy-Preserving DNN Training Framework via Vector Decomposition

Summary

Nappa is a DNN training framework that aims to preserve data privacy while remaining compatible with heterogeneous Neural Network Accelerators (NNAs). It uses a vector decomposition method to allocate operators across NNAs, and a compiler that converts model descriptions into encrypted computation graphs running over trusted and untrusted hardware. The source reports no extra accuracy cost, higher throughput than existing methods, and negligible impact after deployment on a platform with a billion users.