oxbrbtx/PrimAITE
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

901 - merged with dev

Browse Source
This commit is contained in:
SunilSamra
2023-07-05 11:34:15 +01:00
parent 3f440c0a28 e199dc52c0
commit f121b0e21c
73 changed files with 4047 additions and 1376 deletions
Download Patch File Download Diff File Expand all files Collapse all files

197
src/primaite/environment/observations.py
View File

@@ -1,18 +1,13 @@
"""Module for handling configurable observation spaces in PrimAITE."""
import logging
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING, Dict, Final, List, Tuple, Union
from typing import Dict, Final, List, Tuple, TYPE_CHECKING, Union
import numpy as np
from gym import spaces
from primaite.acl.acl_rule import ACLRule
from primaite.common.enums import (
FileSystemState,
HardwareState,
RulePermissionType,
SoftwareState,
)
from primaite.common.enums import FileSystemState, HardwareState, RulePermissionType, SoftwareState
from primaite.nodes.active_node import ActiveNode
from primaite.nodes.service_node import ServiceNode
@@ -34,6 +29,7 @@ class AbstractObservationComponent(ABC):
self.env: "Primaite" = env
self.space: spaces.Space
self.current_observation: np.ndarray # type might be too restrictive?
self.structure: List[str]
return NotImplemented
@abstractmethod
@@ -41,6 +37,11 @@ class AbstractObservationComponent(ABC):
"""Update the observation based on the current state of the environment."""
self.current_observation = NotImplemented
@abstractmethod
def generate_structure(self) -> List[str]:
"""Return a list of labels for the components of the flattened observation space."""
return NotImplemented
class NodeLinkTable(AbstractObservationComponent):
"""Table with nodes and links as rows and hardware/software status as cols.
@@ -62,7 +63,7 @@ class NodeLinkTable(AbstractObservationComponent):
"""
_FIXED_PARAMETERS: int = 4
_MAX_VAL: int = 1_000_000
_MAX_VAL: int = 1_000_000_000
_DATA_TYPE: type = np.int64
def __init__(self, env: "Primaite"):
@@ -84,6 +85,8 @@ class NodeLinkTable(AbstractObservationComponent):
# 3. Initialise Observation with zeroes
self.current_observation = np.zeros(observation_shape, dtype=self._DATA_TYPE)
self.structure = self.generate_structure()
def update(self):
"""Update the observation based on current environment state.
@@ -98,9 +101,7 @@ class NodeLinkTable(AbstractObservationComponent):
self.current_observation[item_index][1] = node.hardware_state.value
if isinstance(node, ActiveNode) or isinstance(node, ServiceNode):
self.current_observation[item_index][2] = node.software_state.value
self.current_observation[item_index][
3
] = node.file_system_state_observed.value
self.current_observation[item_index][3] = node.file_system_state_observed.value
else:
self.current_observation[item_index][2] = 0
self.current_observation[item_index][3] = 0
@@ -108,9 +109,7 @@ class NodeLinkTable(AbstractObservationComponent):
if isinstance(node, ServiceNode):
for service in self.env.services_list:
if node.has_service(service):
self.current_observation[item_index][
service_index
] = node.get_service_state(service).value
self.current_observation[item_index][service_index] = node.get_service_state(service).value
else:
self.current_observation[item_index][service_index] = 0
service_index += 1
@@ -130,12 +129,44 @@ class NodeLinkTable(AbstractObservationComponent):
protocol_list = link.get_protocol_list()
protocol_index = 0
for protocol in protocol_list:
self.current_observation[item_index][
protocol_index + 4
] = protocol.get_load()
self.current_observation[item_index][protocol_index + 4] = protocol.get_load()
protocol_index += 1
item_index += 1
def generate_structure(self):
"""Return a list of labels for the components of the flattened observation space."""
nodes = self.env.nodes.values()
links = self.env.links.values()
structure = []
for i, node in enumerate(nodes):
node_id = node.node_id
node_labels = [
f"node_{node_id}_id",
f"node_{node_id}_hardware_status",
f"node_{node_id}_os_status",
f"node_{node_id}_fs_status",
]
for j, serv in enumerate(self.env.services_list):
node_labels.append(f"node_{node_id}_service_{serv}_status")
structure.extend(node_labels)
for i, link in enumerate(links):
link_id = link.id
link_labels = [
f"link_{link_id}_id",
f"link_{link_id}_n/a",
f"link_{link_id}_n/a",
f"link_{link_id}_n/a",
]
for j, serv in enumerate(self.env.services_list):
link_labels.append(f"link_{link_id}_service_{serv}_load")
structure.extend(link_labels)
return structure
class NodeStatuses(AbstractObservationComponent):
"""Flat list of nodes' hardware, OS, file system, and service states.
@@ -184,6 +215,7 @@ class NodeStatuses(AbstractObservationComponent):
# 3. Initialise observation with zeroes
self.current_observation = np.zeros(len(shape), dtype=self._DATA_TYPE)
self.structure = self.generate_structure()
def update(self):
"""Update the observation based on current environment state.
@@ -206,10 +238,37 @@ class NodeStatuses(AbstractObservationComponent):
if node.has_service(service):
service_states[i] = node.get_service_state(service).value
obs.extend(
[hardware_state, software_state, file_system_state, *service_states]
[
hardware_state,
software_state,
file_system_state,
*service_states,
]
)
self.current_observation[:] = obs
def generate_structure(self):
"""Return a list of labels for the components of the flattened observation space."""
services = self.env.services_list
structure = []
for _, node in self.env.nodes.items():
node_id = node.node_id
structure.append(f"node_{node_id}_hardware_state_NONE")
for state in HardwareState:
structure.append(f"node_{node_id}_hardware_state_{state.name}")
structure.append(f"node_{node_id}_software_state_NONE")
for state in SoftwareState:
structure.append(f"node_{node_id}_software_state_{state.name}")
structure.append(f"node_{node_id}_file_system_state_NONE")
for state in FileSystemState:
structure.append(f"node_{node_id}_file_system_state_{state.name}")
for service in services:
structure.append(f"node_{node_id}_service_{service}_state_NONE")
for state in SoftwareState:
structure.append(f"node_{node_id}_service_{service}_state_{state.name}")
return structure
class LinkTrafficLevels(AbstractObservationComponent):
"""Flat list of traffic levels encoded into banded categories.
@@ -263,9 +322,7 @@ class LinkTrafficLevels(AbstractObservationComponent):
self._entries_per_link = self.env.num_services
# 1. Define the shape of your observation space component
shape = (
[self._quantisation_levels] * self.env.num_links * self._entries_per_link
)
shape = [self._quantisation_levels] * self.env.num_links * self._entries_per_link
# 2. Create Observation space
self.space = spaces.MultiDiscrete(shape)
@@ -273,6 +330,8 @@ class LinkTrafficLevels(AbstractObservationComponent):
# 3. Initialise observation with zeroes
self.current_observation = np.zeros(len(shape), dtype=self._DATA_TYPE)
self.structure = self.generate_structure()
def update(self):
"""Update the observation based on current environment state.
@@ -292,14 +351,27 @@ class LinkTrafficLevels(AbstractObservationComponent):
elif load >= bandwidth:
traffic_level = self._quantisation_levels - 1
else:
traffic_level = (load / bandwidth) // (
1 / (self._quantisation_levels - 2)
) + 1
traffic_level = (load / bandwidth) // (1 / (self._quantisation_levels - 2)) + 1
obs.append(int(traffic_level))
self.current_observation[:] = obs
def generate_structure(self):
"""Return a list of labels for the components of the flattened observation space."""
structure = []
for _, link in self.env.links.items():
link_id = link.id
if self._combine_service_traffic:
protocols = ["overall"]
else:
protocols = [protocol.name for protocol in link.protocol_list]
for p in protocols:
for i in range(self._quantisation_levels):
structure.append(f"link_{link_id}_{p}_traffic_level_{i}")
return structure
class AccessControlList(AbstractObservationComponent):
"""Flat list of all the Access Control Rules in the Access Control List.
@@ -329,7 +401,6 @@ class AccessControlList(AbstractObservationComponent):
]
"""
0,
# Terms (for ACL observation space):
# [0, 1, 2] - Permission (0 = NA, 1 = DENY, 2 = ALLOW)
# [0, num nodes] - Source IP (0 = NA, 1 = any, then 2 -> x resolving to IP addresses)
@@ -389,8 +460,7 @@ class AccessControlList(AbstractObservationComponent):
nodes = list(self.env.nodes.values())
for node in nodes:
if (
isinstance(node, ServiceNode)
or isinstance(node, ActiveNode)
isinstance(node, ServiceNode) or isinstance(node, ActiveNode)
) and node.ip_address == source_ip:
source_ip_int = int(node.node_id) + 1
break
@@ -400,8 +470,7 @@ class AccessControlList(AbstractObservationComponent):
nodes = list(self.env.nodes.values())
for node in nodes:
if (
isinstance(node, ServiceNode)
or isinstance(node, ActiveNode)
isinstance(node, ServiceNode) or isinstance(node, ActiveNode)
) and node.ip_address == dest_ip:
dest_ip_int = int(node.node_id) + 1
if protocol == "ANY":
@@ -444,6 +513,12 @@ class AccessControlList(AbstractObservationComponent):
# print("current obs", obs, "\n" ,len(obs))
self.current_observation[:] = obs
def generate_structure(self):
"""Return a list of labels for the components of the flattened observation space."""
structure = []
return structure
class ObservationsHandler:
"""Component-based observation space handler.
@@ -461,8 +536,17 @@ class ObservationsHandler:
def __init__(self):
self.registered_obs_components: List[AbstractObservationComponent] = []
self.space: spaces.Space
self.current_observation: Union[Tuple[np.ndarray], np.ndarray]
# internal the observation space (unflattened version of space if flatten=True)
self._space: spaces.Space
# flattened version of the observation space
self._flat_space: spaces.Space
self._observation: Union[Tuple[np.ndarray], np.ndarray]
# used for transactions and when flatten=true
self._flat_observation: np.ndarray
self.flatten: bool = False
def update_obs(self):
"""Fetch fresh information about the environment."""
@@ -471,12 +555,11 @@ class ObservationsHandler:
obs.update()
current_obs.append(obs.current_observation)
# If there is only one component, don't use a tuple, just pass through that component's obs.
if len(current_obs) == 1:
self.current_observation = current_obs[0]
self._observation = current_obs[0]
else:
self.current_observation = tuple(current_obs)
# TODO: We may need to add ability to flatten the space as not all agents support tuple spaces.
self._observation = tuple(current_obs)
self._flat_observation = spaces.flatten(self._space, self._observation)
def register(self, obs_component: AbstractObservationComponent):
"""Add a component for this handler to track.
@@ -503,12 +586,31 @@ class ObservationsHandler:
for obs_comp in self.registered_obs_components:
component_spaces.append(obs_comp.space)
# If there is only one component, don't use a tuple space, just pass through that component's space.
# if there are multiple components, build a composite tuple space
if len(component_spaces) == 1:
self.space = component_spaces[0]
self._space = component_spaces[0]
else:
self.space = spaces.Tuple(component_spaces)
# TODO: We may need to add ability to flatten the space as not all agents support tuple spaces.
self._space = spaces.Tuple(component_spaces)
if len(component_spaces) > 0:
self._flat_space = spaces.flatten_space(self._space)
else:
self._flat_space = spaces.Box(0, 1, (0,))
@property
def space(self):
"""Observation space, return the flattened version if flatten is True."""
if self.flatten:
return self._flat_space
else:
return self._space
@property
def current_observation(self):
"""Current observation, return the flattened version if flatten is True."""
if self.flatten:
return self._flat_observation
else:
return self._observation
@classmethod
def from_config(cls, env: "Primaite", obs_space_config: dict):
@@ -538,6 +640,9 @@ class ObservationsHandler:
# Instantiate the handler
handler = cls()
if obs_space_config.get("flatten"):
handler.flatten = True
for component_cfg in obs_space_config["components"]:
# Figure out which class can instantiate the desired component
comp_type = component_cfg["name"]
@@ -551,3 +656,17 @@ class ObservationsHandler:
handler.update_obs()
return handler
def describe_structure(self):
"""Create a list of names for the features of the obs space.
The order of labels follows the flattened version of the space.
"""
# as it turns out it's not possible to take the gym flattening function and apply it to our labels so we have
# to fake it. each component has to just hard-code the expected label order after flattening...
labels = []
for obs_comp in self.registered_obs_components:
labels.extend(obs_comp.structure)
return labels

292
src/primaite/environment/primaite_env.py
View File

@@ -1,11 +1,11 @@
# Crown Copyright (C) Dstl 2022. DEFCON 703. Shared in confidence.
"""Main environment module containing the PRIMmary AI Training Evironment (Primaite) class."""
import copy
import csv
import logging
from datetime import datetime
import uuid as uuid
from pathlib import Path
from typing import Dict, Tuple, Union
from random import choice, randint, sample, uniform
from typing import Dict, Final, Tuple, Union
import networkx as nx
import numpy as np
@@ -13,11 +13,13 @@ import yaml
from gym import Env, spaces
from matplotlib import pyplot as plt
from primaite import getLogger
from primaite.acl.access_control_list import AccessControlList
from primaite.agents.utils import is_valid_acl_action_extra, is_valid_node_action
from primaite.common.custom_typing import NodeUnion
from primaite.common.enums import (
ActionType,
AgentFramework,
FileSystemState,
HardwareState,
NodePOLInitiator,
@@ -25,6 +27,7 @@ from primaite.common.enums import (
NodeType,
ObservationType,
Priority,
SessionType,
SoftwareState,
)
from primaite.common.service import Service
@@ -43,9 +46,9 @@ from primaite.pol.green_pol import apply_iers, apply_node_pol
from primaite.pol.ier import IER
from primaite.pol.red_agent_pol import apply_red_agent_iers, apply_red_agent_node_pol
from primaite.transactions.transaction import Transaction
from primaite.utils.session_output_writer import SessionOutputWriter
_LOGGER = logging.getLogger(__name__)
_LOGGER.setLevel(logging.INFO)
_LOGGER = getLogger(__name__)
class Primaite(Env):
@@ -61,7 +64,6 @@ class Primaite(Env):
self,
training_config_path: Union[str, Path],
lay_down_config_path: Union[str, Path],
transaction_list,
session_path: Path,
timestamp_str: str,
):
@@ -70,26 +72,23 @@ class Primaite(Env):
:param training_config_path: The training config filepath.
:param lay_down_config_path: The lay down config filepath.
:param transaction_list: The list of transactions to populate.
:param session_path: The directory path the session is writing to.
:param timestamp_str: The session timestamp in the format:
<yyyy-mm-dd>_<hh-mm-ss>.
"""
self.session_path: Final[Path] = session_path
self.timestamp_str: Final[str] = timestamp_str
self._training_config_path = training_config_path
self._lay_down_config_path = lay_down_config_path
self.training_config: TrainingConfig = training_config.load(
training_config_path
)
self.training_config: TrainingConfig = training_config.load(training_config_path)
_LOGGER.info(f"Using: {str(self.training_config)}")
# Number of steps in an episode
self.episode_steps = self.training_config.num_steps
super(Primaite, self).__init__()
# Transaction list
self.transaction_list = transaction_list
# The agent in use
self.agent_identifier = self.training_config.agent_identifier
@@ -182,6 +181,9 @@ class Primaite(Env):
# It will be initialised later.
self.obs_handler: ObservationsHandler
self._obs_space_description = None
"The env observation space description for transactions writing"
# Open the config file and build the environment laydown
with open(self._lay_down_config_path, "r") as file:
# Open the config file and build the environment laydown
@@ -208,16 +210,14 @@ class Primaite(Env):
plt.savefig(file_path, format="PNG")
plt.clf()
except Exception:
_LOGGER.error("Could not save network diagram")
_LOGGER.error("Exception occured", exc_info=True)
print("Could not save network diagram")
_LOGGER.error("Could not save network diagram", exc_info=True)
# Initiate observation space
self.observation_space, self.env_obs = self.init_observations()
# Define Action Space - depends on action space type (Node or ACL)
if self.training_config.action_type == ActionType.NODE:
_LOGGER.info("Action space type NODE selected")
_LOGGER.debug("Action space type NODE selected")
# Terms (for node action space):
# [0, num nodes] - node ID (0 = nothing, node ID)
# [0, 4] - what property it's acting on (0 = nothing, state, SoftwareState, service state, file system state) # noqa
@@ -226,7 +226,7 @@ class Primaite(Env):
self.action_dict = self.create_node_action_dict()
self.action_space = spaces.Discrete(len(self.action_dict))
elif self.training_config.action_type == ActionType.ACL:
_LOGGER.info("Action space type ACL selected")
_LOGGER.debug("Action space type ACL selected")
# Terms (for ACL action space):
# [0, 2] - Action (0 = do nothing, 1 = create rule, 2 = delete rule)
# [0, 1] - Permission (0 = DENY, 1 = ALLOW)
@@ -237,27 +237,29 @@ class Primaite(Env):
self.action_dict = self.create_acl_action_dict()
self.action_space = spaces.Discrete(len(self.action_dict))
elif self.training_config.action_type == ActionType.ANY:
_LOGGER.info("Action space type ANY selected - Node + ACL")
_LOGGER.debug("Action space type ANY selected - Node + ACL")
self.action_dict = self.create_node_and_acl_action_dict()
self.action_space = spaces.Discrete(len(self.action_dict))
else:
_LOGGER.info(
f"Invalid action type selected: {self.training_config.action_type}"
)
# Set up a csv to store the results of the training
try:
header = ["Episode", "Average Reward"]
_LOGGER.error(f"Invalid action type selected: {self.training_config.action_type}")
file_name = f"average_reward_per_episode_{timestamp_str}.csv"
file_path = session_path / file_name
self.csv_file = open(file_path, "w", encoding="UTF8", newline="")
self.csv_writer = csv.writer(self.csv_file)
self.csv_writer.writerow(header)
except Exception:
_LOGGER.error(
"Could not create csv file to hold average reward per episode"
)
_LOGGER.error("Exception occured", exc_info=True)
self.episode_av_reward_writer = SessionOutputWriter(self, transaction_writer=False, learning_session=True)
self.transaction_writer = SessionOutputWriter(self, transaction_writer=True, learning_session=True)
@property
def actual_episode_count(self) -> int:
"""Shifts the episode_count by -1 for RLlib."""
if self.training_config.agent_framework is AgentFramework.RLLIB:
return self.episode_count - 1
return self.episode_count
def set_as_eval(self):
"""Set the writers to write to eval directories."""
self.episode_av_reward_writer = SessionOutputWriter(self, transaction_writer=False, learning_session=False)
self.transaction_writer = SessionOutputWriter(self, transaction_writer=True, learning_session=False)
self.episode_count = 0
self.step_count = 0
self.total_step_count = 0
def reset(self):
"""
@@ -266,12 +268,14 @@ class Primaite(Env):
Returns:
Environment observation space (reset)
"""
csv_data = self.episode_count, self.average_reward
self.csv_writer.writerow(csv_data)
if self.actual_episode_count > 0:
csv_data = self.actual_episode_count, self.average_reward
self.episode_av_reward_writer.write(csv_data)
self.episode_count += 1
# Don't need to reset links, as they are cleared and recalculated every step
# Don't need to reset links, as they are cleared and recalculated every
# step
# Clear the ACL
self.init_acl()
@@ -280,6 +284,10 @@ class Primaite(Env):
# Does this for both live and reference nodes
self.reset_environment()
# Create a random red agent to use for this episode
if self.training_config.random_red_agent:
self._create_random_red_agent()
# Reset counters and totals
self.total_reward = 0
self.step_count = 0
@@ -287,6 +295,7 @@ class Primaite(Env):
# Update observations space and return
self.update_environent_obs()
return self.env_obs
def step(self, action):
@@ -302,15 +311,10 @@ class Primaite(Env):
done: Indicates episode is complete if True
step_info: Additional information relating to this step
"""
if self.step_count == 0:
print(f"Episode: {str(self.episode_count)}")
# TEMP
done = False
self.step_count += 1
self.total_step_count += 1
# print("Episode step: " + str(self.step_count))
# Need to clear traffic on all links first
for link_key, link_value in self.links.items():
@@ -320,13 +324,15 @@ class Primaite(Env):
link.clear_traffic()
# Create a Transaction (metric) object for this step
transaction = Transaction(
datetime.now(), self.agent_identifier, self.episode_count, self.step_count
)
transaction = Transaction(self.agent_identifier, self.actual_episode_count, self.step_count)
# Load the initial observation space into the transaction
transaction.set_obs_space_pre(copy.deepcopy(self.env_obs))
transaction.obs_space = self.obs_handler._flat_observation
# Set the transaction obs space description
transaction.obs_space_description = self._obs_space_description
# Load the action space into the transaction
transaction.set_action_space(copy.deepcopy(action))
transaction.action_space = copy.deepcopy(action)
# 1. Implement Blue Action
self.interpret_action_and_apply(action)
@@ -370,9 +376,7 @@ class Primaite(Env):
self.acl,
self.step_count,
)
apply_red_agent_node_pol(
self.nodes, self.red_iers, self.red_node_pol, self.step_count
)
apply_red_agent_node_pol(self.nodes, self.red_iers, self.red_node_pol, self.step_count)
# Take snapshots of nodes and links
self.nodes_post_red = copy.deepcopy(self.nodes)
self.links_post_red = copy.deepcopy(self.links)
@@ -388,35 +392,39 @@ class Primaite(Env):
self.step_count,
self.training_config,
)
# print(f" Step {self.step_count} Reward: {str(reward)}")
_LOGGER.debug(f"Episode: {self.actual_episode_count}, " f"Step {self.step_count}, " f"Reward: {reward}")
self.total_reward += reward
if self.step_count == self.episode_steps:
self.average_reward = self.total_reward / self.step_count
if self.training_config.session_type == "EVALUATION":
if self.training_config.session_type is SessionType.EVAL:
# For evaluation, need to trigger the done value = True when
# step count is reached in order to prevent neverending episode
done = True
print(f" Average Reward: {str(self.average_reward)}")
_LOGGER.info(f"Episode: {self.actual_episode_count}, " f"Average Reward: {self.average_reward}")
# Load the reward into the transaction
transaction.set_reward(reward)
transaction.reward = reward
# 6. Output Verbose
# self.output_link_status()
# 7. Update env_obs
self.update_environent_obs()
# Load the new observation space into the transaction
transaction.set_obs_space_post(copy.deepcopy(self.env_obs))
# 8. Add the transaction to the list of transactions
self.transaction_list.append(copy.deepcopy(transaction))
# Write transaction to file
if self.actual_episode_count > 0:
self.transaction_writer.write(transaction)
# Return
return self.env_obs, reward, done, self.step_info
def __close__(self):
"""Override close function."""
self.csv_file.close()
def close(self):
"""Override parent close and close writers."""
# Close files if last episode/step
# if self.can_finish:
super().close()
self.transaction_writer.close()
self.episode_av_reward_writer.close()
def init_acl(self):
"""Initialise the Access Control List."""
@@ -425,14 +433,9 @@ class Primaite(Env):
def output_link_status(self):
"""Output the link status of all links to the console."""
for link_key, link_value in self.links.items():
print("Link ID: " + link_value.get_id())
_LOGGER.debug("Link ID: " + link_value.get_id())
for protocol in link_value.protocol_list:
print(
" Protocol: "
+ protocol.get_name().name
+ ", Load: "
+ str(protocol.get_load())
)
print(" Protocol: " + protocol.get_name().name + ", Load: " + str(protocol.get_load()))
def interpret_action_and_apply(self, _action):
"""
@@ -447,13 +450,9 @@ class Primaite(Env):
self.apply_actions_to_nodes(_action)
elif self.training_config.action_type == ActionType.ACL:
self.apply_actions_to_acl(_action)
elif (
len(self.action_dict[_action]) == 7
): # ACL actions in multidiscrete form have len 7
elif len(self.action_dict[_action]) == 7: # ACL actions in multidiscrete form have len 6
self.apply_actions_to_acl(_action)
elif (
len(self.action_dict[_action]) == 4
): # Node actions in multdiscrete (array) from have len 4
elif len(self.action_dict[_action]) == 4: # Node actions in multdiscrete (array) from have len 4
self.apply_actions_to_nodes(_action)
else:
logging.error("Invalid action type found")
@@ -519,9 +518,7 @@ class Primaite(Env):
return
elif property_action == 1:
# Patch (valid action if it's good or compromised)
node.set_service_state(
self.services_list[service_index], SoftwareState.PATCHING
)
node.set_service_state(self.services_list[service_index], SoftwareState.PATCHING)
else:
# Node is not of Service Type
return
@@ -678,6 +675,9 @@ class Primaite(Env):
"""
self.obs_handler = ObservationsHandler.from_config(self, self.obs_config)
if not self._obs_space_description:
self._obs_space_description = self.obs_handler.describe_structure()
return self.obs_handler.space, self.obs_handler.current_observation
def update_environent_obs(self):
@@ -1204,8 +1204,7 @@ class Primaite(Env):
# [0, max acl rules - 1] - Position (0 = first index, then 1 -> x index resolving to acl rule in acl list)
# reserve 0 action to be a nothing action
actions = {0: [0, 0, 0, 0, 0, 0, 0]}
# [1, 1, 2, 1, 1, 1, 2] CREATE RULE ALLOW NODE 2 TO NODE 1 ON SERVICE 1 PORT 1 AT INDEX 2
# 1, 2, 1, 6, 0, 0, 1 ALLOW NODE 2 TO NODE 1 ON SERVICE 1 SERVICE ANY PORT ANY AT INDEX 1
action_key = 1
# 3 possible action decisions, 0=NOTHING, 1=CREATE, 2=DELETE
for action_decision in range(3):
@@ -1246,13 +1245,136 @@ class Primaite(Env):
# Change node keys to not overlap with acl keys
# Only 1 nothing action (key 0) is required, remove the other
new_node_action_dict = {
k + len(acl_action_dict) - 1: v
for k, v in node_action_dict.items()
if k != 0
}
new_node_action_dict = {k + len(acl_action_dict) - 1: v for k, v in node_action_dict.items() if k != 0}
# Combine the Node dict and ACL dict
combined_action_dict = {**acl_action_dict, **new_node_action_dict}
print("combined dict", combined_action_dict.items())
return combined_action_dict
def _create_random_red_agent(self):
"""Decide on random red agent for the episode to be called in env.reset()."""
# Reset the current red iers and red node pol
self.red_iers = {}
self.red_node_pol = {}
# Decide how many nodes become compromised
node_list = list(self.nodes.values())
computers = [node for node in node_list if node.node_type == NodeType.COMPUTER]
max_num_nodes_compromised = len(computers) # only computers can become compromised
# random select between 1 and max_num_nodes_compromised
num_nodes_to_compromise = randint(1, max_num_nodes_compromised)
# Decide which of the nodes to compromise
nodes_to_be_compromised = sample(computers, num_nodes_to_compromise)
# choose a random compromise node to be source of attacks
source_node = choice(nodes_to_be_compromised)
# For each of the nodes to be compromised decide which step they become compromised
max_step_compromised = self.episode_steps // 2 # always compromise in first half of episode
# Bandwidth for all links
bandwidths = [i.get_bandwidth() for i in list(self.links.values())]
if len(bandwidths) < 1:
msg = "Random red agent cannot be used on a network without any links"
_LOGGER.error(msg)
raise Exception(msg)
servers = [node for node in node_list if node.node_type == NodeType.SERVER]
for n, node in enumerate(nodes_to_be_compromised):
# 1: Use Node PoL to set node to compromised
_id = str(uuid.uuid4())
_start_step = randint(2, max_step_compromised + 1) # step compromised
pol_service_name = choice(list(node.services.keys()))
source_node_service = choice(list(source_node.services.values()))
red_pol = NodeStateInstructionRed(
_id=_id,
_start_step=_start_step,
_end_step=_start_step, # only run for 1 step
_target_node_id=node.node_id,
_pol_initiator="DIRECT",
_pol_type=NodePOLType["SERVICE"],
pol_protocol=pol_service_name,
_pol_state=SoftwareState.COMPROMISED,
_pol_source_node_id=source_node.node_id,
_pol_source_node_service=source_node_service.name,
_pol_source_node_service_state=source_node_service.software_state,
)
self.red_node_pol[_id] = red_pol
# 2: Launch the attack from compromised node - set the IER
ier_id = str(uuid.uuid4())
# Launch the attack after node is compromised, and not right at the end of the episode
ier_start_step = randint(_start_step + 2, int(self.episode_steps * 0.8))
ier_end_step = self.episode_steps
# Randomise the load, as a percentage of a random link bandwith
ier_load = uniform(0.4, 0.8) * choice(bandwidths)
ier_protocol = pol_service_name # Same protocol as compromised node
ier_service = node.services[pol_service_name]
ier_port = ier_service.port
ier_mission_criticality = 0 # Red IER will never be important to green agent success
# We choose a node to attack based on the first that applies:
# a. Green IERs, select dest node of the red ier based on dest node of green IER
# b. Attack a random server that doesn't have a DENY acl rule in default config
# c. Attack a random server
possible_ier_destinations = [
ier.get_dest_node_id()
for ier in list(self.green_iers.values())
if ier.get_source_node_id() == node.node_id
]
if len(possible_ier_destinations) < 1:
for server in servers:
if not self.acl.is_blocked(
node.ip_address,
server.ip_address,
ier_service,
ier_port,
):
possible_ier_destinations.append(server.node_id)
if len(possible_ier_destinations) < 1:
# If still none found choose from all servers
possible_ier_destinations = [server.node_id for server in servers]
ier_dest = choice(possible_ier_destinations)
self.red_iers[ier_id] = IER(
ier_id,
ier_start_step,
ier_end_step,
ier_load,
ier_protocol,
ier_port,
node.node_id,
ier_dest,
ier_mission_criticality,
)
overwhelm_pol = red_pol
overwhelm_pol.id = str(uuid.uuid4())
overwhelm_pol.end_step = self.episode_steps
# 3: Make sure the targetted node can be set to overwhelmed - with node pol
# # TODO remove duplicate red pol for same targetted service - must take into account start step
o_pol_id = str(uuid.uuid4())
o_red_pol = NodeStateInstructionRed(
_id=o_pol_id,
_start_step=ier_start_step,
_end_step=self.episode_steps,
_target_node_id=ier_dest,
_pol_initiator="DIRECT",
_pol_type=NodePOLType["SERVICE"],
pol_protocol=ier_protocol,
_pol_state=SoftwareState.OVERWHELMED,
_pol_source_node_id=source_node.node_id,
_pol_source_node_service=source_node_service.name,
_pol_source_node_service_state=source_node_service.software_state,
)
self.red_node_pol[o_pol_id] = o_red_pol

23
src/primaite/environment/reward.py
View File

@@ -41,27 +41,19 @@ def calculate_reward_function(
reference_node = reference_nodes[node_key]
# Hardware State
reward_value += score_node_operating_state(
final_node, initial_node, reference_node, config_values
)
reward_value += score_node_operating_state(final_node, initial_node, reference_node, config_values)
# Software State
if isinstance(final_node, ActiveNode) or isinstance(final_node, ServiceNode):
reward_value += score_node_os_state(
final_node, initial_node, reference_node, config_values
)
reward_value += score_node_os_state(final_node, initial_node, reference_node, config_values)
# Service State
if isinstance(final_node, ServiceNode):
reward_value += score_node_service_state(
final_node, initial_node, reference_node, config_values
)
reward_value += score_node_service_state(final_node, initial_node, reference_node, config_values)
# File System State
if isinstance(final_node, ActiveNode):
reward_value += score_node_file_system(
final_node, initial_node, reference_node, config_values
)
reward_value += score_node_file_system(final_node, initial_node, reference_node, config_values)
# Go through each red IER - penalise if it is running
for ier_key, ier_value in red_iers.items():
@@ -80,14 +72,9 @@ def calculate_reward_function(
if step_count >= start_step and step_count <= stop_step:
reference_blocked = not reference_ier.get_is_running()
live_blocked = not ier_value.get_is_running()
ier_reward = (
config_values.green_ier_blocked * ier_value.get_mission_criticality()
)
ier_reward = config_values.green_ier_blocked * ier_value.get_mission_criticality()
if live_blocked and not reference_blocked:
_LOGGER.debug(
f"Applying reward of {ier_reward} because IER {ier_key} is blocked"
)
reward_value += ier_reward
elif live_blocked and reference_blocked:
_LOGGER.debug(