Create your own machine learning model from scratch

This repo accompanies a BSidesLV 2017 talk, "Your model isn't that special". It serves two purposes:

1. It's a playground/tutorial for machine learning malware models, including end-to-end deep learning for malicious file detection.
2. It's a demonstration that one cannot simply throw deep learning architectures inspired in other domains and expect good performance for static malware classification. (The models may not be optimized for malware, and certainly our training data sizes are too small to train large capacity models.)

Notebooks for BSidesLV 2017 "Your Model Isn't That Special"

For a gentle walk through the BSidesLV 2017 talk, it's recommended that you peruse through the notebooks in the following order

1. BSidesLV -- your model isn't that special -- (1) MLP
2. BSidesLV -- your model isn't that special -- (2) End-to-End
3. BSidesLV -- your model isn't that special -- (3) MalwaResNet

Getting started

Requirements

Python dependencies

This code was developed using Python 3.6. Necessary packages can be installed by typing

pip install -r requirements.txt

in your shell.

Bring your own samples

Create subdirectories malicious/ and benign/ off of the main branch, and populate them with malicious and benign samples, respectively. Hint: this may be the most important step in creating your machine learning model. For the first multilayer perceptron model, you should make sure to have at least 100K samples between the two subdirectories. For the end-to-end models that have very large capacity, one should generally try for millions of samples between the two directories. For this demo, we're going to stick with 100K samples, and note that the expressive end-to-end models are overfitting.

Training models

You may then try any of the malware models contained in the classifier/ directory. For example python classifier/modeltest_multilayer.py will extract features for all the the samples in malicious/ and benign/ and cache them into sample_index.json and X.dat numpy array, then build a multilayer percpetron on top of those features.

The end-to-end deep learning models do not require any feature extraction, however, can take a very long time (and a lot of data!) to train. On a single Titan X GPU, we trained the simple end-to-end model in about 24 hours on 100K malicious and benign samples. For the end-to-end deep learning models, it is recommended that you significantly increase the number of training epochs beyond what is contained in the model test scripts.