Malware Env for OpenAI Gym

---

Citing

If you use this code in a publication please cite the following paper:

Hyrum S. Anderson, Anant Kharkar, Bobby Filar, David Evans, Phil Roth, "Learning to Evade Static PE Machine Learning Malware Models via Reinforcement Learning", in ArXiv e-prints. Jan. 2018.

@ARTICLE{anderson2018learning,
  author={Anderson, Hyrum S and Kharkar, Anant and Filar, Bobby and Evans, David and Roth, Phil},
  title={Learning to Evade Static PE Machine Learning Malware Models via Reinforcement Learning},
  journal={arXiv preprint arXiv:1801.08917},
  archivePrefix = "arXiv",
  eprint = {1801.08917},
  primaryClass = "cs.CR",
  keywords = {Computer Science - Cryptography and Security},
  year = 2018,
  month = jan,
  adsurl = {http://adsabs.harvard.edu/abs/2018arXiv180108917A},
}

This is a malware manipulation environment for OpenAI's gym. OpenAI Gym is a toolkit for developing and comparing reinforcement learning algorithms. This makes it possible to write agents that learn to manipulate PE files (e.g., malware) to achieve some objective (e.g., bypass AV) based on a reward provided by taking specific manipulation actions.

Objective

Create an AI that learns through reinforcement learning which functionality-preserving transformations to make on a malware sample to break through / bypass machine learning static-analysis malware detection.

Basics

There are two basic concepts in reinforcement learning: the environment (in our case, the malware sample) and the agent (namely, the algorithm used to change the environment). The agent sends actions to the environment, and the environment replies with observations and rewards (that is, a score).

This repo provides an environment for manipulating PE files and providing rewards that are based around bypassing AV. An agent can be deployed that have already been written for the rich gym framework. For example

• https://github.com/pfnet/chainerrl [recommended]
• https://github.com/matthiasplappert/keras-rl

Setup

The EvadeRL framework is built on Python3.6 we recommend first creating a virtualenv (details can be found here) with Python3.6 then performing the following actions ensure you have the correct python libraries:

pip install -r requirements.txt

EvadeRL also leverages a Library to Instrument Executable Formats aptly named LIEF. It allows our agent to modify the binary on-the-fly. To add it to your virtualenv just pip install one of their pre-built packages. Examples below:

Linux

pip install https://github.com/lief-project/LIEF/releases/download/0.7.0/linux_lief-0.7.0_py3.6.tar.gz

OSX

pip install https://github.com/lief-project/LIEF/releases/download/0.7.0/osx_lief-0.7.0_py3.6.tar.gz

Once completed ensure you've moved malware samples into the

gym_malware/gym_malware/envs/utils/samples/

If you are unsure where to acquire malware samples see the Data Acquisition section below. If you have samples in the correct directory you can check to see if your environment is correctly setup by running :

python test_agent_chainer.py

Note that if you are using Anaconda, you may need to

conda install libgcc

in order for LIEF to operate properly.

Data Acquisition

If you have a VirusTotal API key, you may download samples to the gym_malware/gym_malware/envs/utils/samples/ using the Python script download_samples.py.

Gym-Malware Environment

EvadeRL pits a reinforcement agent against the malware environment consisting of the following components:

• Action Space
• Independent Malware Classifier
• OpenAI framework malware environment (aka gym-malware)

Action Space

The moves or actions that can be performed on a malware sample in our environment consist of the following binary manipulations:

• append_zero
• append_random_ascii
• append_random_bytes
• remove_signature
• upx_pack
• upx_unpack
• change_section_names_from_list
• change_section_names_to random
• modify_export
• remove_debug
• break_optional_header_checksum

The agent will randomly select these actions in an attempt to bypass the classifier (info on default classifier below). Over time, the agent learns which combinations lead to the highest rewards, or learns a policy (like an optimal plan of attack for any given observation).

Independent Classifier

Included as a default model is a gradient boosted decision trees model trained on 50k malicious and 50k benign samples with the following features extracted:

• Byte-level data (e.g. histogram and entropy)
• Header
• Section
• Import/Exports