CLAP: Learning Transferable Binary Code Representations with Natural Language Supervision

About | News | QuickStart | Citation

About

CLAP (Contrastive Language-Assembly Pre-training) is a framework that learns binary code representations through natural language supervision. By aligning binary code with natural language explanations, it improves analysis performance in few-shot and zero-shot scenarios. Utilizing a dataset engine capable of automatically generating 195 million pairs of code snippets and their descriptions, CLAP offers a method with exceptional transferability in the field of binary code analysis. Our goal is to provide an effective tool for researchers and practitioners in binary code analysis, with our models accessible on the Hugging Face Model Hub.

News

[2024/2/29] CLAP is available on Hugging Face Model Hub (clap-asm and clap-text).
[2024/2/28] CLAP is now on ArXiv.

QuickStart

This document will help you set up and start using the CLAP model for various tasks, including fine-grained classification of sorting algorithms, malware, and cryptographic algorithms without any further training.

Requirements

Python 3.6 or higher
PyTorch
Transformers library
A CUDA-enabled GPU is highly recommended for faster processing.

Ensure you have Python and PyTorch installed on your system. Then, install the Transformers library using pip:

pip install transformers

Preparing Tokenizers and Models

Import necessary libraries and initialize the model and tokenizers:

import torch
from transformers import AutoModel, AutoTokenizer

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

asm_tokenizer = AutoTokenizer.from_pretrained("hustcw/clap-asm", trust_remote_code=True)
text_tokenizer = AutoTokenizer.from_pretrained("hustcw/clap-text", trust_remote_code=True)
asm_encoder = AutoModel.from_pretrained("hustcw/clap-asm", trust_remote_code=True).to(device)
text_encoder = AutoModel.from_pretrained("hustcw/clap-text", trust_remote_code=True).to(device)

Example Use Cases

Fine-Grained Sorting Algorithm Classification (Zero-Shot)

Load your assembly (asm) code dataset. For demonstration, we use a JSON file containing assembly code snippets related to bubble sort:

with open("./CaseStudy/bubblesort.json") as fp:
    asm = json.load(fp)

Define your classification prompts:

prompts = [
"This is a function related to bubble sort",
"This is a function related to selection sort",
...
]

Encode the assembly code and prompts, then perform classification:

# Encode assembly code
asm_input = asm_tokenizer([asm], padding=True, return_tensors="pt").to(device)
asm_embedding = asm_encoder(**asm_input)

# Encode prompts
text_input = text_tokenizer(prompts, return_tensors='pt').to(device)
text_embeddings = text_encoder(**text_input)

# Classification
logits = torch.einsum("nc,ck->nk", [asm_embedding.last_hidden_state, text_embeddings.last_hidden_state.T])
preds = torch.softmax(logits / 0.07, dim=1).squeeze(0).tolist()

# Output predictions
for i, prompt in enumerate(prompts):
    print(f"Probability: {preds[i]*100:.3f}%, Text: {prompt}")

Repeat the process for any other classification tasks you want, such as malware classification and cryptographic algorithm identification, by loading the respective datasets and defining the relevant natural language prompts.

Citation

If this work is helpful for your research, please consider giving a star 🌟 and citing our work.

@misc{wang2024clap,
title={CLAP: Learning Transferable Binary Code Representations with Natural Language Supervision},
author={Hao Wang and Zeyu Gao and Chao Zhang and Zihan Sha and Mingyang Sun and Yuchen Zhou and Wenyu Zhu and Wenju Sun and Han Qiu and Xi Xiao},
year={2024},
eprint={2402.16928},
archivePrefix={arXiv},
primaryClass={cs.SE}
}