Skip to content

VulRepair: A T5-Based Automated Software Vulnerability Repair

License

Notifications You must be signed in to change notification settings

awsm-research/VulRepair

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

28 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DOI

VulRepair Replication Package

VulRepair

A T5-based Automated Software Vulnerability Repair

Auto-Repair Real-World Software Vulnerabilities

VulRepair Performance on Top-25 Most Dangerous CWEs in 2021

Rank CWE Type Name %PP Proportion
1 CWE-787 Out-of-bounds Write 30% 16/53
2 CWE-79 Cross-site Scripting 0 0/1
3 CWE-125 Out-of-bounds Read 32% 54/170
4 CWE-20 Improper Input Validation 45% 68/152
5 CWE-78 OS Command Injection 33% 1/3
6 CWE-89 SQL Injection 20% 1/5
7 CWE-416 Use After Free 53% 29/55
8 CWE-22 Path Traversal 25% 2/8
9 CWE-352 Cross-Site Request Forgery 0 0/2
10 CWE-434 Dangerous File Type - -
11 CWE-306 Missing Authentication for Critical Function - -
12 CWE-190 Integer Overflow or Wraparound 53% 31/59
13 CWE-502 Deserialization of Untrusted Data - -
14 CWE-287 Improper Authentication 50% 3/6
15 CWE-476 NULL Pointer Dereference 66% 46/70
16 CWE-798 Use of Hard-coded Credentials - -
17 CWE-119 Improper Restriction of Operations 37% 141/386
18 CWE-862 Missing Authorization 0 0/2
19 CWE-276 Incorrect Default Permissions - -
20 CWE-200 Exposure of Sensitive Information 61% 39/64
21 CWE-522 Insufficiently Protected Credentials 0 0/4
22 CWE-732 Incorrect Permission Assignment 50% 1/2
23 CWE-611 Improper Restriction of XML Reference 0 0/3
24 CWE-918 Server-Side Request Forgery (SSRF) 0 0/1
25 CWE-77 Command Injection 100% 2/2
TOTAL 41% 434/1048

Top-10 Most Accurately Repaired CWE Types of VulRepair

Rank CWE Type Name %PP Proportion
1 CWE-755 Improper Handling of Exceptional Conditions 100% 1/1
2 CWE-706 Use of Incorrectly-Resolved Name or Reference 100% 1/1
3 CWE-326 Inadequate Encryption Strength 100% 2/2
4 CWE-667 Improper Locking 100% 1/1
5 CWE-369 Divide By Zero 100% 5/5
6 CWE-77 Command Injection 100% 2/2
7 CWE-388 Error Handling 100% 1/1
8 CWE-436 Interpretation Conflict 100% 1/1
9 CWE-191 Integer Underflow 100% 2/2
10 CWE-285 Improper Access Control 75% 6/8
TOTAL 92% 22/24

VulRepair Performance on Top-10 Majority CWE Types in Testing Data

Rank CWE Type Name %PP Proportion
1 CWE-119 Improper Restriction of Operations 37% 141/386
2 CWE-125 Out-of-bounds Read 32% 54/170
3 CWE-20 Improper Input Validation 45% 68/152
4 CWE-264 Permissions, Privileges, and Access Controls 51% 36/71
5 CWE-476 NULL Pointer Dereference 66% 46/70
6 CWE-200 Exposure of Sensitive Information 61% 39/64
7 CWE-399 Resource Management Errors 62% 37/60
8 CWE-190 Integer Overflow or Wraparound 53% 31/59
9 CWE-416 Use After Free 53% 29/55
10 CWE-362 Race Condition 43% 23/54
TOTAL 44% 504/1141

The raw predictions of VulRepair can be accessed here

[FSE 2022 Technical track] [Paper #152] [7 mins talk]
VulRepair: A T5-based Automated Software Vulnerability Repair
To appear in ESEC/FSE 2022 (14-18 November, 2022).

Table of contents

  1. How to replicate
  2. Appendix
  3. Acknowledgements
  4. License
  5. Citation

How to replicate

About the Environment Setup

First of all, clone this repository to your local machine and access the main dir via the following command:

git clone https://github.com/awsm-research/VulRepair.git
cd VulRepair

Then, install the python dependencies via the following command:

pip install transformers
pip install torch
pip install numpy
pip install tqdm
pip install pandas
pip install tokenizers
pip install datasets
pip install gdown
pip install tensorboard
pip install scikit-learn

Alternatively, we provide requirements.txt with version of packages specified to ensure the reproducibility, you may install via the following commands:

pip install -r requirements.txt

If having an issue with the gdown package, try the following commands:

git clone https://github.com/wkentaro/gdown.git
cd gdown
pip install .
cd ..
  • We highly recommend you check out this installation guide for the "torch" library so you can install the appropriate version on your device.

  • To utilize GPU (optional), you also need to install the CUDA library, you may want to check out this installation guide.

  • Python 3.9.7 is recommended, which has been fully tested without issues.

About the Datasets

All of the dataset has the same number of columns (i.e., 7 cols), we focus on the following 2 columns to conduct our experiments:

  1. source (str): The localized vulnerable function written in C (preprocessed by Chen et al.)
  2. target (str): The repair ground-truth (preprocessed by Chen et al.)
source target
... ...

Descriptive statistics of our experimental dataset

1st Qt. Median 3rd Qt. Avg.
Function Length 138 280 593 586
Patch Length 12 24 48 55
Cyclomatic Complexity of Functions 3 8 19 23

Note.

  1. This dataset is originally provided by Bhandari et al. and Fan et al., and it is further preprocessed by Chen et al.

    For more information, please kindly refer to this repository.

  2. We process cyclomatic complexity (CC) using Joern tool

    Dataset with labelled CC. can be found here

About the Models

Model Naming Convention

Model Name Model Specification Related to RQ
M1 (VulRepair) BPE Tokenizer + Pre-training (PL/NL) + T5 RQ1, RQ2, RQ3, RQ4
M2 (CodeBERT) BPE Tokenizer + Pre-training (PL/NL) + BERT RQ1, RQ2, RQ3
M3 BPE Tokenizer + No Pre-training + T5 RQ2, RQ4
M4 BPE Tokenizer + Pre-training (NL) + T5 RQ2
M5 BPE Tokenizer + No Pre-training + BERT RQ2
M6 BPE Tokenizer + Pre-training (NL) + BERT RQ2
M7 Word-level Tokenizer + Pre-training (PL/NL) + T5 RQ3, RQ4
M8 BPE Tokenizer + Vanilla XFMR RQ3
M9 Word-level Tokenizer + Pre-training (PL/NL) + BERT RQ3
M10 Word-level Tokenizer + No Pre-training + T5 RQ4

How to access the models

  • We host our VulRepair on the Model Hub provided by Huggingface Transformers which can be access here.
  • All other models can be downloaded from this public Google Cloud Space.

About VulRepair Replication

To reproduce the results of our VulRepair (M1 model), run the following commands (Inference only):

cd M1_VulRepair_PL-NL
python vulrepair_main.py \
    --output_dir=./saved_models \
    --model_name=model.bin \
    --tokenizer_name=MickyMike/VulRepair \
    --model_name_or_path=MickyMike/VulRepair \
    --do_test \
    --encoder_block_size 512 \
    --decoder_block_size 256 \
    --num_beams=50 \
    --eval_batch_size 1

Note. please adjust the "num_beams" parameters accordingly to obtain the results we present in the discussion section. (i.e., num_beams= 1, 2, 3, 4, 5, 10)

To retrain the VulRepair model from scratch, run the following commands (Training + Inference):

# training
cd M1_VulRepair_PL-NL
python vulrepair_main.py \
    --model_name=model.bin \
    --output_dir=./saved_models \
    --tokenizer_name=Salesforce/codet5-base \
    --model_name_or_path=Salesforce/codet5-base \
    --do_train \
    --epochs 75 \
    --encoder_block_size 512 \
    --decoder_block_size 256 \
    --train_batch_size 4 \
    --eval_batch_size 4 \
    --learning_rate 2e-5 \
    --max_grad_norm 1.0 \
    --evaluate_during_training \
    --seed 123456  2>&1 | tee train.log
# Inference
python vulrepair_main.py \
    --output_dir=./saved_models \
    --model_name=model.bin \
    --tokenizer_name=Salesforce/codet5-base \
    --model_name_or_path=Salesforce/codet5-base \
    --do_test \
    --encoder_block_size 512 \
    --decoder_block_size 256 \
    --num_beams=50 \
    --eval_batch_size 1

About the Experiment Replication

We recommend to use GPU with 8 GB up memory for training since T5 and BERT architecture is very computing intensive.

Note. If the specified batch size is not suitable for your device, please modify --eval_batch_size and --train_batch_size to fit your GPU memory.

How to replicate RQ1

You need to replicate M1(VulRepair) and M2(CodeBERT) to replicate the results of RQ1:

  • Click here for the instruction of replicating M1(VulRepair)
  • Click here for the instruction of replicating M2(CodeBERT)

How to replicate RQ2

You need to replicate M1(VulRepair), M2(CodeBERT), M3, M4, M5, M6 to replicate the results of RQ2:

  • Click here for the instruction of replicating M1(VulRepair)
  • Click here for the instruction of replicating M2(CodeBERT)
  • Click here for the instruction of replicating M3
  • Click here for the instruction of replicating M4
  • Click here for the instruction of replicating M5
  • Click here for the instruction of replicating M6

How to replicate RQ3

You need to replicate M1(VulRepair), M2(CodeBERT), M7, M8, M9 to replicate the results of RQ2:

  • Click here for the instruction of replicating M1(VulRepair)
  • Click here for the instruction of replicating M2(CodeBERT)
  • Click here for the instruction of replicating M7
  • Click here for the instruction of replicating M8
  • Click here for the instruction of replicating M9

How to replicate RQ4

You need to replicate M1(VulRepair), M3, M7, M10 to replicate the results of RQ2:

  • Click here for the instruction of replicating M1(VulRepair)
  • Click here for the instruction of replicating M3
  • Click here for the instruction of replicating M7
  • Click here for the instruction of replicating M10

Appendix

Results of RQ1

Methods % Perfect Prediction
VulRepair 44%
CodeBERT 31%
VRepair 21%

Results of RQ2

T5 % Perfect Prediction
PL/NL (VulRepair) 44%
No Pre-training 30%
NL 6%
BERT % Perfect Prediction
PL/NL (CodeBERT) 31%
No Pre-training 29%
NL 1%

Results of RQ3

VulRepair % Perfect Prediction
Subword Tokenizer 44%
Word-level Tokenizer 35%
VRepair % Perfect Prediction
Subword Tokenizer 34%
Word-level Tokenizer 23%
CodeBERT % Perfect Prediction
Subword Tokenizer 31%
Word-level Tokenizer 17%

Results of RQ4

VulRepair % Perfect Prediction
Pre-train + BPE + T5 44%
Pre-train + Word-level + T5 35%
No Pre-train + BPE + T5 30%
No Pre-train + Word-level + T5 1%

Acknowledgements

License

MIT License

Citation

@inproceedings{fu2022vulrepair,
  title={VulRepair: A T5-based Automated Software Vulnerability Repair},
  author={Fu, Michael and Tantithamthavorn, Chakkrit and Le, Trung and Nguyen, Van and Dinh, Phung},
  journal={To appear in the ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE)},
  year={2022}
}