LLM4Decompile

Reverse Engineering: Decompiling Binary Code with Large Language Models

Updates

[2023-03-16]: Add llm4decompile-6.7b-uo model which is trained without prior knowledge of the optimization levels (O0~O3), the average re-executability is around 0.219, performs the best in our models.

About

LLM4Decompile is the pioneering open-source large language model dedicated to decompilation. Its current version supports decompiling Linux x86_64 binaries, ranging from GCC's O0 to O3 optimization levels, into human-readable C source code. Our team is committed to expanding this tool's capabilities, with ongoing efforts to incorporate a broader range of architectures and configurations.
Decompile-Eval is the first decompilation benchmark that focuses on assessing the re-compilability and re-executability aspects of decompiled code. It is the C language adaptation of the HumanEval dataset and provides a suite of C solutions and assertions in evaluating the practical utility of decompiled code.

Evaluation Results

Metrics

Re-compilability assesses if the decompiled code can successfully be recompiled with the original compiler settings and configurations.
Re-executability evaluates whether the decompiled code can execute properly and pass all the predefined test cases.

Re-compilability and re-executability serve as critical indicators in validating the effectiveness of a decompilation process. When decompiled code can be recompiled, it provides strong evidence of syntactic integrity. It ensures that the decompiled code is not just readable, but also adheres to the structural and syntactical standards expected by the compiler. However, syntax alone does not guarantee semantic equivalence to the original pre-compiled program. Re-executability provides this critical measure of semantic correctness. By re-compiling the decompiled output and running the test cases, we assess if the decompilation preserved the program logic and behavior. Together, re-compilability and re-executability indicate syntax recovery and semantic preservation - both essential for usable and robust decompilation.

Figure 1 presents the steps involved in our decompilation evaluation. First, the source code (denoted as src) is compiled by the GCC compiler with specific parameters, such as optimization levels, to produce the executable binary. This binary is then disassembled into assembly language (asm) using the objdump tool. The assembly instructions are subsequently decompiled to reconstruct the source code in a format that's readable to humans (noted as src'). To assess the quality of the decompiled code (src'), it is tested for its ability to be recompiled with the original GCC compiler (re-compilability) and for its functionality through test assertions (re-executability).

Results

Models

Our LLM4Decompile includes models with sizes between 1.3 billion and 33 billion parameters, and we have made these models available on Hugging Face.

Model	Checkpoint	Size	Re-executability	Note
llm4decompile-1.3b	🤗 HF Link	1.3B	10.6%	-
llm4decompile-6.7b	🤗 HF Link	6.7B	21.4%	-
llm4decompile-33b	🤗 HF Link	33B	21.5%	-
llm4decompile-6.7b-nsp	🤗 HF Link	6.7B	20.9%	Note 1
llm4decompile-6.7b-uo	🤗 HF Link	6.7B	21.9%	Note 2

Note 1: The NSP model is trained with assembly code, the average re-executability is around 0.17.

Note 2: The unified optimization (UO) model is trained without prior knowledge of the optimization levels (O0~O3), the average re-executability is around 0.21. The pre-processing of the UO model is slightly different (no prior knowledge of the On), please check the model page.

Quick Start

Here is an example of how to use our model.

Preprocessing: Compile the C code into binary, and disassemble the binary into assembly instructions.

import subprocess
import os
import re

digit_pattern = r'\b0x[a-fA-F0-9]+\b'# binary codes in Hexadecimal
zeros_pattern = r'^0+\s'#0s
OPT = ["O0", "O1", "O2", "O3"]
fileName = 'path/to/file'
with open(fileName+'.c','r') as f:#original file
    c_func = f.read()
for opt_state in OPT:
    output_file = fileName +'_' + opt_state
    input_file = fileName+'.c'
    compile_command = f'gcc -c -o {output_file}.o {input_file} -{opt_state} -lm'#compile the code with GCC on Linux
    subprocess.run(compile_command, shell=True, check=True)
    compile_command = f'objdump -d {output_file}.o > {output_file}.s'#disassemble the binary file into assembly instructions
    subprocess.run(compile_command, shell=True, check=True)
    
    input_asm = ''
    with open(output_file+'.s') as f:#asm file
        asm= f.read()
    asm = asm.split('Disassembly of section .text:')[-1].strip()
    for tmp in asm.split('\n'):
        tmp_asm = tmp.split('\t')[-1]#remove the binary code
        tmp_asm = tmp_asm.split('#')[0].strip()#remove the comments
        input_asm+=tmp_asm+'\n'
    input_asm = re.sub(zeros_pattern, '', input_asm)
    before = f"# This is the assembly code with {opt_state} optimization:\n"#prompt
    after = "\n# What is the source code?\n"#prompt
    input_asm_prompt = before+input_asm.strip()+after
    with open(fileName +'_' + opt_state +'.asm','w',encoding='utf-8') as f:
        f.write(input_asm_prompt)

Decompilation: Use LLM4Decompile to translate the assembly instructions into C:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_path = 'arise-sustech/llm4decompile-1.3b'
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path,torch_dtype=torch.bfloat16).cuda()

with open(fileName +'_' + opt_state +'.asm','r') as f:#original file
    asm_func = f.read()
inputs = tokenizer(asm_func, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=500)
c_func_decompile = tokenizer.decode(outputs[0][len(inputs[0]):-1])

Decompile-Eval

Data are stored in llm4decompile/decompile-eval/decompile-eval.json, using JSON list format. There are 164*4 (O0, O1, O2, O3) samples, each with five keys:

task_id: indicates the ID of the problem.
type: the optimization stage, is one of [O0, O1, O2, O3].
c_func: C solution for HumanEval problem.
c_test: C test assertions.
input_asm_prompt: assembly instructions with prompts, can be derived as in our preprocessing example.

To run the evaluation on a single GPU and single process:

cd LLM4Decompile
python ./evaluation/run_evaluation_llm4decompile_singleGPU.py

To run the evaluation using TGI (10x faster, support multiple GPUs and multi-process): First, please install the text-generation-inference following the official link

git clone https://github.com/albertan017/LLM4Decompile.git
cd LLM4Decompile
pip install -r requirements.txt

# Before running the evaluation script, please update the model_path to your local model path.
bash ./scripts/run_evaluation_llm4decompile.sh

On Going

Larger training dataset with the cleaning process.
Support for popular languages/platforms and settings.
Support for executable binaries.
Integration with decompilation tools (e.g., Ghidra, Rizin)

License

This code repository is licensed under the MIT and DeepSeek License.

Citation

@misc{tan2024llm4decompile,
      title={LLM4Decompile: Decompiling Binary Code with Large Language Models}, 
      author={Hanzhuo Tan and Qi Luo and Jing Li and Yuqun Zhang},
      year={2024},
      eprint={2403.05286},
      archivePrefix={arXiv},
      primaryClass={cs.PL}
}

zhang-ge-hao/LLM4Decompile