Learning Execution Semantics from Micro-Traces for Binary Similarity

Attended CS Colloquium on 2023-03-28.
[2012.08680] Trex: Learning Execution Semantics from Micro-Traces for Binary Similarity
In Natural Language, semantics usually release info about surroundings
In Binary, not so much.
We can't do Dynamic Analysis due to Scalability problems.
Teach AI approximate learning on binary code instructions and make mental executions.
- Dynamic Execution has coverage problems.
- So here, we set arbitrary registers to run small code. This is for pre-training purposes, not to make accurate extrapolations.
Masked Language Modeling. Mask certain portions of binaries and make the AI guess.
Questions may arise in Microexecutions and Logical reasoning. Employing Higher level programming language mapping is possible often, but not researched through.
Now, fine-tune this to use static analysis.
Can analyze Semantically Similar Binary Functions
Function Dependencies and Signatures are also predictable
Can find vulnerabilities
Limitations: in-code only. Cannot find multimodal programming vulnerabilities. i.e., involving many interacting components.

이 문서를 언급한 문서들