Proj CDeepFuzz Paper Reading: Decompiling x86 Deep Neural Network Executables

Abstract

本文: BTD
github: https://github.com/monkbai/DNN-decompiler/
Task: a decompiler for DNN models to output DNN specifications including: operators, network topology, dimensions and parameters
特点：enable full optimizations on x86(Q)
Method: learning DNN operators, dynamic analysis to reveal network architecture, symbolic execution to facilitate dimensions and parameters
实验：
DNN Compiler：TVM, Glow, NNFusion
DNN Models: Resnet18, VGG16, FastText, Inception, Shufflenet, Mobilenet, Efficientnet
效果：

  1. 能对如Resnet这种等级的DNN反编译。验证方法：重新编译后功能基本一致
  2. can boost 2 representation attacks: 1. adversarial example generation 2. knowledge stealing
  3. cross architecture legacy code reuse
posted @ 2023-09-07 17:32  雪溯  阅读(15)  评论(0)  编辑  收藏  举报