Lost and Found in Translation
summary
Overview of ML-assisted decompilation and security implications: demonstrates using an encoder-decoder Transformer trained on automatically generated assembly↔C pairs from competitive programming repos, with no heavy preprocessing; empirical result shows model output still needs large edits (nearly every other token corrected on average) though some functions decompile nearly perfectly, so a human-in-the-loop decompiler assistant is realistic. Also highlights tooling risks from code-generation: studies of Copilot-style models find up to 40% vulnerable completions in crafted scenarios, and assembly can be ~15× longer than source, complicating translation.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 0:58Who am I
- 1:43Agenda
- 3:48Google Whistleblower
- 4:56GitHub CoPilot
- 5:56Security
- 12:56Unsupervised Learning
- 13:15Reinforcements
- 15:14Two Rules of Thumb
- 16:48Machine Learning Models
- 17:13Feature Engineering
- 29:30Pseudocode
- 30:24Natural Language Translation
- 35:43Neural Machine Translation
- 38:24Sequence to Sequence Models
- 39:05Encoder Decoder Architecture
- 40:07Recurrent Neural Networks
- 41:00Encoder
- 42:21Attention
- 43:26Bert
- 44:53Language Models
- 46:57Preprocessing
- 49:07Data Acquisition
- 52:04Summary
- 52:56Questions
related talks
45m"Advanced Malware Threat: handling obfuscated code using..."
48mHow to Build A Decompiler
01h 22mBeyond Decoding: Meta-Generation Algorithms for Large Language Models
24mBTD: Unleashing the Power of Decompilation for x86 Deep Neural Network Executables
26mReverse Engineering Using LLMs
19mDecompiling for Constant-Time Analysis
29mSyntia: Synthesizing the Semantics of Obfuscated Code
34mTerrible tokenizer troubles in large language models