BTD: Unleashing the Power of Decompilation for x86 Deep Neural Network Executables
summary
BTD decompiler reconstructs high‑level DNN models from x86 executables by combining trace-based symbolic execution, an LSTM opcode classifier to map assembly functions to operators, and instrumentation that dumps weights at runtime. Tested on eight compiler versions (TVM, Glow, oneDNN/Infusion) across seven ImageNet models, it achieves >99% operator recovery and near‑100% parameter recovery in most cases, rebuilds functionally identical models, and enables white‑box attacks like model inversion.
AI-written summary · about
chapters (16)
- 0:00BTD and DNN decompilation
- 1:11DNN executables
- 5:15Motivation for decompiling DNNs
- 9:29Attacks on deep learning models
- 11:59Extracting DNN executables
- 12:42BTD design
- 14:55Operator recovery in BTD
- 16:01Topology recovery in BTD
- 16:44Parameter recovery
- 19:16BTD demo
- 19:38BTD evaluation
- 20:44Operator and parameter accuracy
- 21:49White-box attacks on DNNs
- 22:32Case study: attacking a DNN
- 22:53Key takeaways
- 23:34Q&A
related talks
32mDnD: Decompiling Deep Neural Network Compiled Binary
23mAn In-Depth Analysis of Disassembly on Full-Scale x86/x64 Binaries
54mLost and Found in Translation
14mNeural Reverse Engineering of Stripped Binaries using Augmented Control Flow Graphs
48mHow to Build A Decompiler
40mDynamic Binary Instrumentation Techniques to Address Native Code Obfuscation
21mDemystify Today's Binary Disassembling and How Modern ABI Makes it Easier
45m"Advanced Malware Threat: handling obfuscated code using..."