xiaojunxu / dnn-binary-code-similarity Goto Github PK

View Code? Open in Web Editor NEW

135.0 5.0 49.0 16.91 MB

License: Other

Python 100.00%

dnn-binary-code-similarity's Introduction

DNN Binary Code Similarity Detection

This repo provides an implementation of the Gemini network for binary code similarity detection in this paper.

Prepration and Data

Unzip the data by running:

unzip data.zip

The network is written using Tensorflow 1.4 in Python 2.7. You can install the dependencies by running:

pip install -r requirements.txt

Model Implementation

The model is implemented in graphnnSiamese.py.

Run the following code to train the model:

python train.py

or run python train.py -h to check the optional arguments.

After training, run the following code to evaluate the model:

python eval.py

or run python eval.py -h to check the optional arguments.

dnn-binary-code-similarity's People

Stargazers

Watchers

Forkers

gaoyz0625 xiaohuisu yuedeji ikangaroo jiweizhi chubbymaggie wgmueller1 coder-chenzhi wliuxingxiangyu hamedmx prosp3r0 iamsabhoho twelveand0 hustcw yichuanyanyu26 sakuracy adalisan scion01 zeovan jianwei76 wangrun xianglinyang ericharley dingelish themoep beicichen wendy23git michaelpdu danijimmy19 shouguoyang wmkhoo apurva-cuj lyq12345 claymore715 louisccc rhmpzyyx evandowning foresta-yang naiyupang djeeone12 lordoz234 gystar 4b5f5f4b nharrisonau monster-oneday sigmoid-bar

dnn-binary-code-similarity's Issues

Hello, I need the data set mentioned in your paper

The data set I need is"Dataset IV：This dataset contains vulnerable functions obtained from the vulnerability dataset in [18]. In total, it contains of 154vulnerable functions."
[18] Qian Feng, Rundong Zhou, Chengcheng Xu, Yao Cheng, Brian Testa, and Heng Yin. 2016. Scalable Graph-based Bug Search for Firmware Images. In ACM Conference on Computer and Communications Security (CCS’16).
But I didn't find the address of this data set in his paper. Could you help me

May i ask whether your experimental result contain top1,top10 and so on?

Hi,
In the 4.7 section,i find the top-50 result could achieve over 80% accuracy.So i would like to get other result like top1 or top10.

solved

training auc is equal validation auc?

in line 140 of the train.py

        if (i % TEST_FREQ == 0):
            auc, fpr, tpr, thres = get_auc_epoch(gnn, Gs_train, classes_train,
                    BATCH_SIZE, load_data=valid_epoch)     #valid_epoch?
            gnn.say("Testing model: training auc = {0} @ {1}".format(
                auc, datetime.now()))
            auc, fpr, tpr, thres = get_auc_epoch(gnn, Gs_dev, classes_dev,
                    BATCH_SIZE, load_data=valid_epoch)
            gnn.say("Testing model: validation auc = {0} @ {1}".format(
                auc, datetime.now()))

partial results:

Initial training auc = 0.8385795480339078 @ 2021-01-14 10:49:18.571989
Initial validation auc = 0.8385795480339078 @ 2021-01-14 10:49:26.453895
EPOCH 1/100, loss = 0.6696747210700994 @ 2021-01-14 10:52:33.539342
Testing model: training auc = 0.9225966099742847 @ 2021-01-14 10:52:41.815191
Testing model: validation auc = 0.9225966099742847 @ 2021-01-14 10:52:49.352049
Model saved in ./saved_model/graphnn-model_best
EPOCH 2/100, loss = 0.6475605093730598 @ 2021-01-14 10:55:58.523499
Testing model: training auc = 0.9278334338892803 @ 2021-01-14 10:56:06.467269
Testing model: validation auc = 0.9278334338892803 @ 2021-01-14 10:56:14.051002

so should I change it as the following?

        if (i % TEST_FREQ == 0):
            auc, fpr, tpr, thres = get_auc_epoch(gnn, Gs_train, classes_train,
                    BATCH_SIZE, load_data=None) #change valid_epoch into None
            gnn.say("Testing model: training auc = {0} @ {1}".format(
                auc, datetime.now()))
            auc, fpr, tpr, thres = get_auc_epoch(gnn, Gs_dev, classes_dev,
                    BATCH_SIZE, load_data=valid_epoch)
            gnn.say("Testing model: validation auc = {0} @ {1}".format(
                auc, datetime.now()))