SK²Decompile: LLM-based Two-Phase Binary Decompilation from Skeleton to Skin

SK²Decompile is a novel two-phase framework for binary decompilation using Large Language Models (LLMs). Our approach decomposes the complex decompilation task into two manageable phases:

Phase 1 Structure Recovery (Skeleton): Transform binary/pseudo-code into obfuscated intermediate representations (current model)

Phase 2 Identifier Naming (Skin): Generate human-readable source code with meaningful identifiers 🤗 HF Link

Usage:

0. Install vllm and transformers via pip; install clang-format via apt.
1. Prepare a Linux-x64 executable file (ELF).
2. Use IDA to decompile it (you can also simply use this website: https://dogbolt.org/).
3. Convert the data into the corresponding format (https://huggingface.co/LLM4Binary/sk2decompile-struct-6.7b/blob/main/reverse_sample.json)

python normalize_pseudo.py --input_json reverse_sample.json --output_json reverse_sample.json

4. Run inference:

   python sk2decompile.py --dataset_path reverse_sample.json \
       --model_path LLM4Binary/sk2decompile-struct-6.7b \
       --recover_model_path LLM4Binary/sk2decompile-ident-6.7b
   

Project overview: https://github.com/albertan017/LLM4Decompile/tree/main/sk2decompile

Notes:

• IDA decompilation results should be preprocessed before inference.
• Use vllm to recover function structure (sk2decompile-struct) and variable names (sk2decompile-ident) step by step.
• Training was done on C language Linux-x64 code with IDA pseudocode; performance may degrade for other languages or architectures.