Skip to content

Code of the AISec23 paper "Drift Forensics of Malware Classifiers"

License

Notifications You must be signed in to change notification settings

isneslab/DriftAnalysis

Repository files navigation

Drift Forensics of Malware Classifiers

Repository containing code for our AISec23 paper:

Theo Chow, Zeliang Kan, Lorenz Linhardt, Lorenzo Cavallaro, Daniel Arp, and Fabio Pierazzi , Drift Forensics of Malware Classifiers , In Proc. of the ACM Workshop on Artificial Intelligence and Security (AISec), 2023

If you use this repository in your own research, please cite our AISec23 paper as follows:

@inproceedings{chow2023driftforensics,
  title = {Drift Forensics of Malware Classifiers},
  author = {Chow, Theo and Kan, Zeliang and Linhardt, Lorenz and Cavallaro, Lorenzo and Arp, Daniel and Pierazzi, Fabio},
  booktitle = {Proc. of the {ACM} Workshop on Artificial Intelligence and Security ({AISec})},
  year = {2023},
}

Link to dataset can be found Here

Getting Started

Installation

This project requires Python 3 as well as the statistical learning stack of NumPy, SciPy and Scikit-learn, secml.

First, install package dependencies using the listing in requirements.txt.

pip install -r requirements.txt

Run experiments

To reproduce the paper results, run

python paper_results.py

Load dataset

First load in the desired dataset and obtain the X predictors X, y predictors y, timestamps t, family labels f, feature names feature_names and md5 md5.

PATH = "../Datasets/extended-features/"
X, y, t, f, feature_names, md5 = load_transcend(f"{PATH}extended-features-X-updated.json",
                                                f"{PATH}extended-features-y-updated.json",
                                                f"{PATH}extended-features-meta-updated.json",
                                                f"{PATH}meta_info_file.tsv")

Reduce feature space

Reduce the feature space to a manageable amount and save the feature indexes as a pkl file

X, feature_names = util.feature_reduction(X, y, feature_names, "pkl_files/feature_index_1000.pkl", feature_size=1000)

Dataset class

Put the data in to a dataset class, this gives us flexibiliy when selecting samples. Currently there are 2 main functions in the dataset class, splitting the dataset in to time aware splits for analysis and finding occurences of features in the dataset.

dataset = Dataset(X, y, t, f, feature_names, md5)

Search up feature name IDs

ids = dataset.get_feature_id_from_name("android")

Find IDs in family

dataset.sample_select_from_feature_id(families=['Dowgin','Dnotua','Kuguo','Airpush','Revmob'],ids=ids,contains=True, year=2015, month=1)

Split dataset and return time aware indexes for training and test

train, test = dataset.time_aware_split_index('month', 6, 1)

Analysis

The analysis class runs the experiment outlined in the paper. Currently, there are 3 main experiments, base, half and snoop. The results of this will be logged in a MySQL database and the results in a pkl file. By default, a file name pkl_files needs to be created.

analyse = Analysis(X, y, t, f, feature_names, train, test)

training_family = ['Dowgin','Dnotua','Kuguo','Airpush','Revmob']
testing_family = ['Dowgin','Dnotua','Kuguo','Airpush','Revmob']

analyse.run(training_family=training_family, testing_family=testing_family,experiment='snoop', dataset='Transcend')

Visualizing data

To visualise the results, we first load in the corresponding data in question. The ResultsLoader() class gives an easy way in accessing saved experiments.

training_familes = ['Dowgin','Dnotua','Kuguo','Airpush','Revmob']
testing_families = ['Dowgin','Dnotua','Kuguo','Airpush','Revmob']
ResultsLoader().query_database_for_ID('half',training_familes,testing_families,'Transcend')

Load in the desired data using the ID returned by ResultsLoader()

result1 = ResultsLoader().load_file_from_id(5)
result2 = ResultsLoader().load_file_from_id(6)

For performance, distribution and difference plots

Viz(result1,result2).plot_performance_distribution()
Viz(result1,result2).plot_single('difference')

About

Code of the AISec23 paper "Drift Forensics of Malware Classifiers"

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •  

Languages