URLNet

Introduction

This is an implementation of URLNet - Learning a URL Representation with Deep Learning for Malicious URL Detection https://arxiv.org/abs/1802.03162

URLNet is a convolutional neural network (CNN) based model used to detect malicious URLs. The model exploits features of URL text string at both character and word levels.

Resources

URLNet requires Python 3.6 and the following packages:

• tensorflow 1.8

• tflearn 0.3

• numpy 1.14

Model Designs

Sample Commands

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

<URL label><tab><URL string>

Example:

+1  http://www.exampledomain.com/urlpath/...

-1  http://www.exampledomain.com/urlpath/...

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

The model can be trained by running the following command:

python train.py --data.data_dir <data_file_directory> --data.dev_pct 0.2 --data.delimit_mode <url_delimit_mode> --data.min_word_freq <url_min_word_freq> \
--model.emb_mode <embedding_mode> --model.emb_dim <size_of_embedding_dimensions> --model.filter_sizes <convolutional_filter_sizes_separated_by_comma> \
--train.nb_epochs <nb_of_training_epochs> --train.batch_size <nb_of_urls_per_batch> \
--log.print_every <print_acc_after_this_nb_steps> --log.eval_every <evaluate_on_dev_set_after_this_nb_steps> --log.checkpoint_every <checkpoint_model_after_this_nb_steps> --log.output_dir <model_output_folder_directory>

The training will save all the related word and character dictionaries into an output folder, and the model checkpoints are saved into checkpoints/ folder. By default, maximum 5 checkpoints are stored in the folder.

The model can be tested by running the following command:

python test.py --data.data_dir <data_file_directory> --data.delimit_mode <url_delimit_mode> --data.word_dict_dir <word_dictionary_directory> --data.subword_dict_dir <character_in_word_dictionary_directory> --data.char_dict_dir <character_dictionary_directory> \
--log.checkpoint_dir <model_checkpoint_directory> --log.output_dir <test_result_directory> \
--model.emb_mode <embedding_mode> --model.emb_dim <nb_of_embedding_dimensions> \
--test.batch_size <nb_of_urls_per_batch>

The test will save the test results for each URL, including 3 columns: label (original label), predict (prediction label), and score (softmax score). The orders of test results from top to bottom are the same as their orders in the test dataset. If the score is more than 0.5, prediction label is +1 (Malicious). Else, the prediction is -1 (Benign).

Example: ~~~~~~~~~~~~~~~~~~~~~~ label predict score

1 1 0.884

-1 -1 0.359 ~~~~~~~~~~~~~~~~~~~~~~

To obtain test metrics such as True Positive, False Positive, True Negative, False Negative, and the AUC curves, run the following command:

python auc.py --input_path <test_result_directory> --input_file
<test_result_file>

The command will output the test metrics into a file with the same name as the input_file but with an .auc extension.

Parameters

Training parameters include:

Parameter	Description	Default
data.max_len_words	The maximum number of words in a URL. The URL is either truncated or padded with a <PADDING> token to reach this length.	200
data.max_len_chars	The maximum number of characters in a URL. The URL is either truncted or padded with a <PADDING> character to reach this length.	200
data.max_len_subwords	The maximum number of characters in each word in a URL. Each word is either truncated or padded with a <PADDING> character to reach this length.	20
data.min_word_freq	Words that have frequency less than or equal to this parameter are considered as rare words, and represented as a single <UNKNOWN> token.	1
data.dev_pct	Percentage of training data used for validation	0.001
data.data_dir	Directory of the training dataset	train_10000.txt
data.delimit_mode	0: delimit URL by special characters (i.e. non-alphanumerical characters), 1: delimit URL by special characters (i.e. non-alphanumerical characters) and treat each special characters as words	1
model.emb_dim	Dimension size of word and character embedding.	32
model.emb_mode	1: only character-based CNN, 2: only word-based CNN, 3: character and word CNN, 4: character-level word CNN, 5: character and character-level word CNN	1
model.filter_sizes	Sizes of convolutional filters. If more than one branches of CNN, all will have the same set of filter sizes. Separate the filter sizes by comma.	3,4,5,6
train.batch_size	Number of URLs in each training batch	128
train.nb_epochs	Number of training epochs	5
train.lr	Learning rate	0.001
train.l2_reg_lambda	regularization parameter of loss function	0
log.output_dir	Output folder to save dictionaries and model checkpoints	runs/10000/
log.print_every	Output the training loss and accuracy after this number of batches	50
log.eval_every	Evaluate the model on the validation set after this number of batches	500
log.checkpoint_every	Checkpoint the model after this number of batches. Only save the model checkpoint if the validation loss is improved.	500

 

Test parameters include:

Parameter	Description	Default
data.max_len_words	The maximum number of words in a URL. The URL is either truncated or padded with a <PADDING> token to reach this length.	200
data.max_len_chars	The maximum number of characters in a URL. The URL is either truncted or padded with a <PADDING> character to reach this length.	200
data.max_len_subwords	The maximum number of characters in each word in a URL. Each word is either truncated or padded with a <PADDING> character to reach this length.	20
data.delimit_mode	0: delimit URL by special characters (i.e. non-alphanumerical characters), 1: delimit URL by special characters (i.e. non-alphanumerical characters) and treat each special characters as words	1
data.data_dir	Directory of the test dataset	test_1000.txt
data.word_dict_dir	Directory of the word dictionary file. Dictionary file is in pickle extension .p	runs/10000/words_dict.p
data.char_dict_dir	Directory of the character dictionary file. Dictionary file is in pickle extension .p	runs/10000/chars_dict.p
data.subword_dict_dir	Directory of the character-in-word dictionary file. Dictionary file is in pickle extension .p	runs/10000/ngrams_dict.p
model.emb_dim	Dimension size of word and character embedding.	32
model.emb_mode	1: only character-based CNN, 2: only word-based CNN, 3: character and word CNN, 4: character-level word CNN, 5: character and character-level word CNN	1
test.batch_size	Number of URLs in each test batch	128
log.checkpoint_dir	Directory of the model checkpoints	runs/10000/checkpoints/
log.output_dir	Directory of the test results	runs/10000/train_10000_test_1000.txt

The test parameters such as model.emb_mode and data.delimit_mode have to be consistent with the trained model to get accurate test results.

Refer to the file run.sh for example commands.

Acknowledgement

We are very grateful to our collaborators VirusTotal, for providing us access to their data, without which this research would not have been possible.