Welcome to the Robust Malware Detection Challenge. This challenge is organized in conjunction with the 1st Workshop on Adversarial Learning Methods for Machine Learning and Data Mining @ KDD 2019.
The bulk of adversarial machine learning research has been focused on crafting attacks and defenses for image classification. In this challenge, we consider adversarial machine learning in the context of robust malware detection. In the era of modern cyber warfare, cyber adversaries craft adversarial perturbations to malicious code to evade malware detectors. Crafting adversarial examples in the malware classification setup is more challenging than image classifications: malware adversarial examples must not only fool the classifier, their adversarial perturbations must not alter the malicious payload. The gist of this challenge is to defend against adversarial attacks by building robust detectors and/or attack robust malware detectors based on binary indicators of imported functions used by the malware. The challenge has two tracks:
- Defense Track: Build high-accuracy deep models that are robust to adversarial attacks.
- Attack Track: Craft adversarial malicious PEs that evades detection on adversarially trained models. We do not consider adversarial examples for benign PEs assuming that their authors do not intend to have them misclassified as malware.
and is based on the following paper
Datasets for the two tracks can be shared upon request, please email advmalwarechallenge@gmail.com
and we will send you a link to the dataset. We note that the dataset contains feature vectors of the PE files, not the actual binaries. For a formal description of the challenge, please refer to this document.
All the required packages are specified in the yml files under helper_files
. If you have conda
installed, you can just cd
to the main directory and execute the following with osx_environment.yml
or linux_environment.yml
on OSx or Linux, respectively.
conda env create —f ./helper_files/(osx|linux)_environment.yml
This will create an environment called nn_mal
.
To activate this environment, execute:
source activate nn_mal
With adding the current directory to the python path:
export PYTHONPATH=`pwd`
the code is ready.
Note: If you’re running the code on Mac OS with Cuda, then according to Pytorch.org “macOS Binaries dont support CUDA, install from source if CUDA is needed”
The code main's script is framework.py
, which implements an adversarial training framework based on the paper above.
Over multiple epochs, it fits the model using a minmax formulation (specified by the training_method
parameter in the parameters.ini
file) and reports the performance on a validation set.
It also reports the performance on a test set and evaluates its robustness against an attack/inner maximizer (specified by the evasion_method
parameter in the parameters.ini
file).
We modified the code to support both challenge tracks. Towards the end of the main()
function, there are 4 conditional code blocks. They can be activated based on the [challenge]
parameters in parameters.ini
as follows.
- For training a model (this is intended for the defend track, but can also be used for the attack track to build a subtitute model for generating transferable adversarial examples)
[challenge]
eval = False
attack = False
defend = True
, ensure that the benign_filepath
and malicious_filepath
point to the defend dataset
[dataset]
benign_filepath = /home/defend/benign/
malicious_filepath = /home/defend/malicious/
Note: the model can be trained from scratch or load its weights from a file as specified in the current version of the parameters.ini
.
load_model_weights = True
model_weights_path = ./helper_files/[training:natural|evasion:dfgsm_k]_demo-model.pt
To train it from scratch, set load_model_weights
to False
. The trained model will be stored in ./helper_files
, participants of the defend track need to send us the generated file.
- For crafting and storing adversarial examples (this is intended for the attack track)
[challenge]
eval = False
attack = True
defend = False
adv_examples_path = "PATH TO WHERE ADVERSARIAL EXAMPLES SHOULD BE STORED"
, ensure that the benign_filepath
and malicious_filepath
point to the attack dataset; also, set the malicious_files_list
as follows.
[dataset]
benign_filepath = /home/attack/benign/
malicious_filepath = /home/attack/malicious/
malicious_files_list = attack_file_list.p
The malicious_files_list
parameter is set to ensure the malicious PEs are looped over in the same order across the participants (and hence the entries of the '.npy' file will correspond to the PEs in the same order).
The above setup will run the evasion_method
specified in parameters.ini
against the model defined by the load_model_weights
and model_weights_path
parameters and generate a numpy '.npy' file at adv_examples_path
, participants of the attack track need to send us this file.
The last two code blocks will be used by the challenge organizers to evaluate the submissions by setting eval = True
as shown next and handled by the script eval_subm_script.py
. Nevertheless, participants may also use this mode to evaluate their techniques.
- When evaluating submissions for the defend track, we will use the following setup.
[challenge]
eval = True
attack = False
defend = True
, point to the holdout dataset.
[dataset]
benign_filepath = "PATH TO BENIGN HOLDOUT DATASET/"
malicious_filepath = "PATH TO MALICIOUS HOLDOUT DATASET/"
num_files_to_use = 3800
, point to the participant's submitted model
[general]
load_model_weights = True
model_weights_path = "PATH TO THE SUBMITTED MODEL"
, and we will attack the model with a set of adversaries specified by the evasion_method
and compute the F1 score based on the performance (metrics file) against the strongest adversary.
- When evaluating submissions for the attack track, we will use the following setup.
[challenge]
eval = True
attack = True
defend = False
adv_examples_path = "PATH TO THE SUBMITTED ADVERSARIAL EXAMPLES"
, point to the attack dataset.
[dataset]
benign_filepath = "PATH TO BENIGN ATTACK DATASET/"
malicious_filepath = "PATH TO MALICIOUS ATTACK DATASET/"
num_files_to_use = 3800
, and point to the secret model
[general]
load_model_weights = True
model_weights_path = "PATH TO THE SECRET MODEL"
, and we will report the evasion rate on the secret model.
The rest of the repo is organized as follows.
- inner_maximizers: a module for implementing inner maximizer algorithms that satisfy the malware perturbation constraints of their binary features (e.g., 'rfgsm_k', 'bca_k' from [1]). These algorithms can be used for training and/or attacking the model based on
parameters.ini
. For instance, the current version ofparameters.ini
setstraining_method = natural
andevasion_method = dfgsm_k
. This means ifframework.py
is made to train a model (this can be set by other parameters inparameters.ini
), then it will train anatural
model (no adversarial training) and on the test test, it will evaluate the model's evasion rate based ondfgsm_k
attack. Participants can also add their own inner maximizer and modify thetraining_method
/evasion_method
inparameters.ini
accordingly. - nets: a module for defining the malware detector's model architecture. Currently, there is only one model architecture defined in
ff_classifier.py
. Participants may define their own model architecture in a similar format to that offf_classifier.py
. If you plan to change the model architecture, ensure the following:- The model's input has the same dimensionality of the PE feature vector.
- The output layer should be an
nn.LogSoftmax(dim=1))
of 2 neurons. - Change Line 76 of
framework.py
to construct the new model - You may change
parameters.ini
parameters:ff_h1
,ff_h2
, and any model-specific parameter.
- datasets: a module for loading the dataset (malicious/benign). Participants need not to modify this.
- run_experiments.py: a script that participants may find helpful. It shows how the
parameters.ini
file can be changed programetically and runningframework.py
afterward. For instance, participants in the attack challenge may first train a robust model on their own, then use it to craft the transferable black-box adversarial examples. For this,parameters.ini
must be changed from training a model to crafting adversarial examples (we will show this shortly). - eval_subm_script.py: the evaluation script to be used by the organizers to evaluate the submissions.
- helper_files: contains supporting files for setting up the Python environment, PE feature mapping, and a baseline natural model
[training:natural|evasion:dfgsm_k]_demo-model.pt
. In training mode (for the defend challenge), trained models in*.pt
format will be saved in this directory and participants of the defend challenge need to share these files with us for evalution. The directory also has a couple of sample.ini
files that will be used by theeval_subm_script.py
to evaluate the submissions. You can have a look at them to understand further howparameters.ini
can be set. - utils: a module that implements multiple functions for computing performance metric and texifying them.
- blindspot_covarge: This can be ignored. It implements an adversarial training metric from [1]
More specific instructions for each track will be provided in the datasets' READMEs.
Results (accuracy metrics, bscn measures, and evasion rates) will be populated under (to-be-generated) result_files
directory every time framework.py
is run. These json files will be used to evaluate the submissions and their names are in the form of
[training:{training_method}|evasion:{evasion_method}]_{experiment_suffix}.json
The results can be compiled into LaTeX tables saved under result_files
by runnig the function create_tex_tables()
with the valid filepath to the result files under utils/script_functions.py
. By default, you can do the following
cd utils/
python script_functions.py
As mentioned earlier, the trained models will be saved under helper_files
.
If you make use of this code and you'd like to cite us, please consider the following:
@inproceedings{al2018adversarial,
title={Adversarial deep learning for robust detection of binary encoded malware},
author={Al-Dujaili, Abdullah and Huang, Alex and Hemberg, Erik and O'Reilly, Una-May},
booktitle={2018 IEEE Security and Privacy Workshops (SPW)},
pages={76--82},
year={2018},
organization={IEEE}
}
We thank Edward Raff for the valuable feedback.