FBS: Accelerating CNN Inference over RNS-CKKS with Fewer Bootstrapping Sparsity

Abstract

FBS accelerates privacy-preserving convolutional neural network inference under RNS-CKKS. It uses weight and channel sparsity to reduce bootstrapping costs and selects pruning strategies based on inference latency. The paper reports up to a 1.91× speedup with negligible accuracy loss.

Publication
Proceedings of the 63rd ACM/IEEE Design Automation Conference
Shuwen Deng
Shuwen Deng
Associate Professor