trex: A Python repository from Columbia University MLSec

Introduction

Trex is a tool to match semantically similar functions based on transfer learning.

Installation

We recommend conda to setup the environment and install the required packages.

First, create the conda environment,

conda create -n trex python=3.8 numpy scipy scikit-learn requests

and activate the conda environment:

conda activate trex

Then, install the latest PyTorch (assume you have GPU):

conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia

Enter the trex root directory: e.g., path/to/trex, and install trex:

pip install --editable .

For large datasets install PyArrow:

pip install pyarrow

For faster training install NVIDIA's apex library:

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" \
  --global-option="--deprecated_fused_adam" --global-option="--xentropy" \
  --global-option="--fast_multihead_attn" ./

Preparation

Pretrained models:

Create the checkpoints and checkpoints/pretrain subdirectory in path/to/trex

mkdir checkpoints, mkdir checkpoints/pretrain

Download our pretrained weight parameters and put in checkpoints/pretrain

Sample data for finetuning similarity

We provide the sample training/testing files of finetuning in here. Download them ad put in data-src/similarity. If you want to prepare the finetuning data yourself, make sure you follow the format shown in data-src/similarity.

The pipeline of data processing should follow command/pretrain/prepare_json.py (read the raw binary, e.g., elf, and obtain the raw bytes for each function and save them in data-raw/funcbytes), command/finetune/prepare_finetune_trace.py (take data-raw/funcbytes as input and generate function code+dummy traces in data-raw/functraces for finetuning), and command/finetune/prepare_finetune.py (generate the actual finetuning dataset in `data-src/similarity).

We have to binarize the data to make it ready to be trained. To binarize the training data for finetuning, run:

python command/finetune/preprocess.py

The binarized training data ready for finetuning (for detecting similarity) will be stored at data-bin/similarity

Training

To finetune the model, run:

./command/finetune/finetune.sh

The scripts loads the pretrained weight parameters from checkpoints/pretrain/ and finetunes the model.

Sample data for pretraining on micro-traces

We also provide (10K) samples and scripts to demonstrate how to pretrain the model. To binarize the training data for pretraining, run:

python command/pretrain/preprocess_pretrain_10k.py

The binarized training data ready for pretraining will be stored at data-bin/pretrain_10k

To pretrain the model, run:

./command/pretrain/pretrain_10k.sh

The pretrained model will be checkpointed at checkpoints/pretrain_10k

Dataset

We put our dataset here.

CUMLSec/trex