PP-Vul: Privacy-Preserving Vulnerability Detection Using Homomorphic Encryption

This repository contains the implementation of PP-Vul.

Requirements

We have confirmed that PP-Vul can be executed on a 64-bit Ubuntu 18.04 system with python

• Python 3.10
• SEAL-Python 4.0 (https://github.com/Huelse/SEAL-Python)
• PyTorch

How to use PP-Vul

1. Dataset pre-processing

• The raw dataset of VulCNN can be found at: https://github.com/CGCL-codes/VulCNN/tree/main/dataset
• The raw dataset of BigVul can be found at: https://github.com/ZeoVan/MSR_20_Code_vulnerability_CSV_Dataset
• The raw dataset of PrimeVul can be found at: https://github.com/DLVulDet/PrimeVul
• Normalization code can be found at: https://github.com/CGCL-codes/VulCNN/blob/main/normalization.py
• For convenience, the preprocessed dataset of VulCNN can be found at: https://drive.usercontent.google.com/download?id=1NGslBtKLA1gUex4tjNWbF01O3sFLSHPD&export=download&authuser=1
• If you are using the preprocessed dataset, Dataset pre-precessing can be skipped.

1.1. Normalization

• Download the raw dataset from the link (https://github.com/CGCL-codes/VulCNN/tree/main/dataset) into dataset folder.
• Unzip the dataset and rename the folder

mv Dataset-sard rawdata

• Normalize the dataset

python normalization.py -i ../dataset/rawdata

1.2. Embedding

• Move to preprocessing folder.

python codet5embedding.py -i ../dataset/normalized/Vul -o ../dataset/embedding/1_line/Vul -n 1
python codet5embedding.py -i ../dataset/normalized/No-Vul -o ../dataset/embedding/1_line/No-Vul -n 1

1.3. Split dataset (train:valid:test=7:2:1)

python split_data.py -i ../dataset/embedding -o ../dataset/1_line 

2. PP-Vul

2.1. Model training (HE-friendly)

• Default model: Kernel height = 4, Number of filters = 16, Activation function = $x^3$, Depthwise convolution
• Move to pp-vul folder.

python classifier.py -i ../dataset/1_line 

2.2. Privacy-preserving inference using homomorphic encrytion

python main.py -i ../dataset/1_line 

3. Plaintext model

• Default model: Number of kernel = 10, Number of filters = 32, Activation function = ReLU, Max pooling
• Move to plaintext_model folder.

python classifier.py -i ../dataset/1_line