TransURL

This is an implementation of the paper: TransURL:Improving Malicious URL Detection with Multi-layer Transformer Encoding and Multi-scale Pyramid Features

Overview

The proposed framework, which is based on a character-aware pretrained language model and includes a novel feature concatenation method, a multi-scale learning module with pyramid spatial attention mechanism, demonstrates good performance in the detection of malicious URLs.

Model Designs

Directory Guide

Folder/File Name	Purpose
/Data	Where the dataset files are stored under their unique dataset names. Contains all the dataset we used in our paper.
/character_bert_wiki	Where the pre-trained model CharBERT based on BERT is stored. You can also download using this link.
Multiple_attention.py	Contains the proposed multi-level attention module.
data_processing.py	Contains functions for preprocessing the dataset, including converting URL strings into BERT's input format and splitting it into training and validation sets.
Model_MMA.py	Where the proposed Malicious URL Detection model PyraTrans architecture is stored.
Model_CharBERT.py	This file contains the original CharBERT model part.
Train.py	Trains the model according to given parameters, You need to modify the "pre-trained model path" and "dataset path".
Test_binary.py	Tests the pre-trained model in the binary classification experiments according to given parameters.
Test_Multiple.py	Tests the pre-trained model in the multiple classification experiment according to given parameters.
bert_utils.py	Some basic source codes of original bert which are used in the CharBERT structure.
adverserial_generator.py	Where the functions to generate adversarial examples by using specified samples are stored.

Requirements

Python 3.8
Pytorch 2.0.0

Usage

In all datasets for training or testing, each line includes the label and the URL text string following the template:

<URL String><label>

Example:

The URL label is either +1 (Malicious) or 0 (Benign).

http://www.exampleforbenign.com/urlpath/ 0
http://www.exampleformalicious.com/urlpath/ 1

You can adjust the hyper parameters for your specific computing device. Also can further tuning the batch_size, learning_rate and train_epoch. To train the model, use the 'Train.py' file, and you can find descriptions of the training parameters that influence the training phase below.