gym-idsgame is a reinforcement learning environment for simulating attack and defense operations in an abstract network intrusion
game. The environment extends the abstract model described in (Elderman et al. 2017). The model constitutes a
two-player Markov game between an attacker agent and a defender agent that face each other in a simulated computer
network. The reinforcement learning environment exposes an interface to a partially observed Markov decision process
(POMDP) model of the Markov game. The interface can be used to train, simulate, and evaluate attack- and defend policies against each other.
Moreover, the repository contains code to reproduce baseline results for various reinforcement learning algorithms, including:
- Tabular Q-learning
- Neural-fitted Q-learning using the DQN algorithm.
- REINFORCE with baseline
- Actor-Critic REINFORCE
- PPO
Please use this bibtex if you make use of this code in your publications (paper: https://arxiv.org/abs/2009.08120):
@INPROCEEDINGS{Hamm2011:Finding,
AUTHOR="Kim Hammar and Rolf Stadler",
TITLE="Finding Effective Security Strategies through Reinforcement Learning and
{Self-Play}",
BOOKTITLE="International Conference on Network and Service Management (CNSM 2020)
(CNSM 2020)",
ADDRESS="Izmir, Turkey",
DAYS=1,
MONTH=nov,
YEAR=2020,
KEYWORDS="Network Security; Reinforcement Learning; Markov Security Games",
ABSTRACT="We present a method to automatically find security strategies for the use
case of intrusion prevention. Following this method, we model the
interaction between an attacker and a defender as a Markov game and let
attack and defense strategies evolve through reinforcement learning and
self-play without human intervention. Using a simple infrastructure
configuration, we demonstrate that effective security strategies can emerge
from self-play. This shows that self-play, which has been applied in other
domains with great success, can be effective in the context of network
security. Inspection of the converged policies show that the emerged
policies reflect common-sense knowledge and are similar to strategies of
humans. Moreover, we address known challenges of reinforcement learning in
this domain and present an approach that uses function approximation, an
opponent pool, and an autoregressive policy representation. Through
evaluations we show that our method is superior to two baseline methods but
that policy convergence in self-play remains a challenge."
}
- Design
- Included Environments
- Requirements
- Installation
- Usage
- Manual Play
- Baseline Experiments
- Future Work
- Author & Maintainer
- Copyright and license
A rich set of configurations of the Markov game are registered as openAI gym environments.
The environments are specified and implemented in gym_idsgame/envs/idsgame_env.py see also gym_idsgame/__init__.py.
This is an environment where the agent is supposed to play the attacker in the Markov game and the defender is following the defend_minimal baseline defense policy.
The defend_minimal policy entails that the defender will always defend the attribute with the minimal value out of all of its neighbors.
Registered configurations:
idsgame-minimal_defense-v0idsgame-minimal_defense-v1idsgame-minimal_defense-v2idsgame-minimal_defense-v3idsgame-minimal_defense-v4idsgame-minimal_defense-v5idsgame-minimal_defense-v6idsgame-minimal_defense-v7idsgame-minimal_defense-v8idsgame-minimal_defense-v9idsgame-minimal_defense-v10idsgame-minimal_defense-v11idsgame-minimal_defense-v12idsgame-minimal_defense-v13idsgame-minimal_defense-v14idsgame-minimal_defense-v15idsgame-minimal_defense-v16idsgame-minimal_defense-v17idsgame-minimal_defense-v18idsgame-minimal_defense-v19idsgame-minimal_defense-v20
This is an environment where the agent is supposed to play the defender and the attacker is following the attack_maximal baseline attack policy.
The attack_maximal policy entails that the attacker will always attack the attribute with the maximum value out of all of its neighbors.
Registered configurations:
idsgame-maximal_attack-v0idsgame-maximal_attack-v1idsgame-maximal_attack-v2idsgame-maximal_attack-v3idsgame-maximal_attack-v4idsgame-maximal_attack-v5idsgame-maximal_attack-v6idsgame-maximal_attack-v7idsgame-maximal_attack-v8idsgame-maximal_attack-v9idsgame-maximal_attack-v10idsgame-maximal_attack-v11idsgame-maximal_attack-v12idsgame-maximal_attack-v13idsgame-maximal_attack-v14idsgame-maximal_attack-v15idsgame-maximal_attack-v16idsgame-maximal_attack-v17idsgame-maximal_attack-v18idsgame-maximal_attack-v19idsgame-maximal_attack-v20
This is an environment where the agent is supposed to play as the defender and the attacker is following a random baseline attack policy.
Registered configurations:
idsgame-random_attack-v0idsgame-random_attack-v1idsgame-random_attack-v2idsgame-random_attack-v3idsgame-random_attack-v4idsgame-random_attack-v5idsgame-random_attack-v6idsgame-random_attack-v7idsgame-random_attack-v8idsgame-random_attack-v9idsgame-random_attack-v10idsgame-random_attack-v11idsgame-random_attack-v12idsgame-random_attack-v13idsgame-random_attack-v14idsgame-random_attack-v15idsgame-random_attack-v16idsgame-random_attack-v17idsgame-random_attack-v18idsgame-random_attack-v19idsgame-random_attack-v20
An environment where the agent is supposed to play as the attacker and the defender is following a random baseline defense policy.
Registered configurations:
idsgame-random_defense-v0idsgame-random_defense-v1idsgame-random_defense-v2idsgame-random_defense-v3idsgame-random_defense-v4idsgame-random_defense-v5idsgame-random_defense-v6idsgame-random_defense-v7idsgame-random_defense-v8idsgame-random_defense-v9idsgame-random_defense-v10idsgame-random_defense-v11idsgame-random_defense-v12idsgame-random_defense-v13idsgame-random_defense-v14idsgame-random_defense-v15idsgame-random_defense-v16idsgame-random_defense-v17idsgame-random_defense-v18idsgame-random_defense-v19idsgame-random_defense-v20
This is an environment where neither the attacker nor defender is part of the environment, i.e. it is intended for 2-agent simulations or RL training. In the experiments folder you can see examples of using this environment for training PPO-attacker vs PPO-defender, DQN-attacker vs REINFORCE-defender, etc..
Registered configurations:
idsgame-v0idsgame-v1idsgame-v2idsgame-v3idsgame-v4idsgame-v5idsgame-v6idsgame-v7idsgame-v8idsgame-v9idsgame-v10idsgame-v11idsgame-v12idsgame-v13idsgame-v14idsgame-v15idsgame-v16idsgame-v17idsgame-v18idsgame-v19idsgame-v20
- Python 3.5
- OpenAI Gym
- NumPy
- Pyglet (OpenGL 3D graphics)
- GPU for 3D graphics acceleration (optional)
- jsonpickle (for configuration files)
- torch (for baseline algorithms)
# install from pip
pip install gym-idsgame==1.0.12
# git clone and install from source
git clone https://github.com/Limmen/gym-idsgame
cd gym-idsgame
pip3 install -e .
# local install from source
$ pip install -e gym-idsgame
# force upgrade deps
$ pip install -e gym-idsgame --upgrade
# run unit tests
pytest
# run it tests
cd experiments
make testsThe environment can be accessed like any other OpenAI environment with gym.make.
Once the environment has been created, the API functions
step(), reset(), render(), and close() can be used to train any RL algorithm of
your preference.
import gymnasium as gym
from gym_idsgame.envs import IdsGameEnv
env_name = "idsgame-maximal_attack-v3"
env = gym.make(env_name)The environment ships with implementation of several baseline algorithms, e.g. the tabular Q(0) algorithm, see the example code below.
import gymnasium as gym
from gym_idsgame.agents.training_agents.q_learning.q_agent_config import QAgentConfig
from gym_idsgame.agents.training_agents.q_learning.tabular_q_learning.tabular_q_agent import TabularQAgent
random_seed = 0
util.create_artefact_dirs(default_output_dir(), random_seed)
q_agent_config = QAgentConfig(gamma=0.999, alpha=0.0005, epsilon=1, render=False, eval_sleep=0.9,
min_epsilon=0.01, eval_episodes=100, train_log_frequency=100,
epsilon_decay=0.9999, video=True, eval_log_frequency=1,
video_fps=5, video_dir=default_output_dir() "/results/videos/" str(random_seed), num_episodes=20001,
eval_render=False, gifs=True, gif_dir=default_output_dir() "/results/gifs/" str(random_seed),
eval_frequency=1000, attacker=True, defender=False, video_frequency=101,
save_dir=default_output_dir() "/results/data/" str(random_seed))
env_name = "idsgame-minimal_defense-v2"
env = gym.make(env_name, save_dir=default_output_dir() "/results/data/" str(random_seed))
attacker_agent = TabularQAgent(env, q_agent_config)
attacker_agent.train()
train_result = attacker_agent.train_result
eval_result = attacker_agent.eval_resultYou can run the environment in a mode of "manual control" as well:
from gym_idsgame.agents.manual_agents.manual_defense_agent import ManualDefenseAgent
random_seed = 0
env_name = "idsgame-random_attack-v2"
env = gym.make(env_name)
ManualDefenseAgent(env.idsgame_config)The experiments folder contains results, hyperparameters and code to reproduce reported results using this environment.
For more information about each individual experiment, see this README.
cd experiments # cd into experiments folder
make cleancd experiments # cd into experiments folder
make allcd experiments # cd into experiments folder
make v0cd experiments/training/v0/random_defense/tabular_q_learning/ # cd into the experiment folder
make runcd experiments/training/v0/random_defense/tabular_q_learning/ # cd into the experiment folder
make cleancd experiments/training/v0/random_defense/tabular_q_learning/ # cd into the experiment folder
make tensorboardBy default when cloning the repo the experiment results are not included, to fetch the experiment results,
install and setup git-lfs then run:
git lfs fetch --all
git lfs pullKim Hammar [email protected]
MIT
(C) 2020, Kim Hammar