基于秘密共享的轻量级隐私保护ViT推理框架

Light weighted privacy protection ViT inference framework based on secret sharing

针对广泛应用于图像处理的ViT推理框架存在泄露用户隐私数据的风险,而已有隐私保护推理框架存在计算效率较低、在线通信量较大等问题,提出了一种高效隐私保护推理框架SViT。该框架由2个边缘服务器协作执行基于秘密共享设计的安全计算协议SSoftmax、SLayerNorm、SGeLU,在保持ViT-B/16原始框架结构的情况下,解决了隐私保护框架推理开销大的问题。理论分析与实验表明,相比CrypTen,SViT在计算效率和在线通信开销方面分别提升了2~6倍和4~14倍。

The ViT(vision transformer)inference framework,which was widely used in image processing,was found to have a risk of leaking user privacy data.However,existing privacy protection inference frameworks had problems such as low computational efficiency and high online communication volume.To address this issue,a highly efficient privacy protection inference framework SViT was proposed.Two edge servers collaborated to execute secure computing protocols based on secret sharing design,such as SSoftmax,SLayerNorm,SGeLU,etc.While maintaining the original framework structure of ViT-B/16,the problem of large inference overhead in privacy protection framework was solved.Theoretical analysis and experiments show that compared to Crypton,SViT has improved computational efficiency by 2~6 times and online communication overhead by 4~14 times,respectively.