Data-Oblivious ML Accelerators Using Hardware Security Extensions.

Hossam ElAtali, John Z. Jekel,Lachlan J. Gunn,N. Asokan

Computing Research Repository (CoRR)（2024）

Univ Waterloo

被引用0|浏览26

摘要

We present Dolma, which applies dynamic information flow tracking (DIFT) to the Gemmini matrix multiplication accelerator. With the BliMe CPU extensions, it efficiently guarantees client data confidentiality, even in the presence of malicious/vulnerable software and side channel attacks on servers.

更多

查看译文

关键词

Machine Learning Accelerators,Confidential Information,Matrix Multiplication,Side-channel,Client Data,Stage 2,Graphics Processing Unit,Security Policy,Input Matrix,Resource Usage,Clock Frequency,Instruction Set Architecture,Hardware Accelerators,Parallel Circuit,Blinded Data,Performance Overhead,Session Key