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：
- 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
- can boost 2 representation attacks: 1. adversarial example generation 2. knowledge stealing
- cross architecture legacy code reuse
- Decompiling Executables CDeepFuzz Network Readingdecompiling executables cdeepfuzz network combinatorial cdeepfuzz learning reading differentiation cdeepfuzz automatic reading pre-trained cdeepfuzz natural reading state-of-the-art cdeepfuzz the reading cdeepfuzz networks reading testing deep-neural-network-driven autonomous cdeepfuzz decompiling executables executables configure compiler cannot