def __init__(self,
ep: EnvironmentBounds,
qm: Optional[np.ndarray] = None) -> None:
self.state_space = w.HashEncoding(
ep,
[
w.Feature_active_node_properties(ep),
w.Feature_active_node_age(ep)
# w.Feature_actions_tried_at_node(ep)
],
7000)
# NOTE: For debugging purpose it's convenient instead to use
# Ravel encoding for node properties
self.state_space_debugging = w.RavelEncoding(
ep,
[
w.HashEncoding(
ep,
[
# Feature_discovered_node_count(),
# Feature_discovered_credential_count(),
w.Feature_discovered_ports_sliding(ep),
w.Feature_discovered_nodeproperties_sliding(ep),
w.Feature_discovered_notowned_node_count(ep, 3),
],
100),
w.Feature_active_node_properties(ep)
])
self.action_space = w.AbstractAction(ep)
super().__init__("attack_at_source", self.state_space,
self.action_space, qm)
def __init__(self, ep: EnvironmentBounds):
self.ep = ep
self.global_features = w.ConcatFeatures(
ep,
[
# w.Feature_discovered_node_count(ep),
# w.Feature_owned_node_count(ep),
w.Feature_discovered_notowned_node_count(ep, None)
# w.Feature_discovered_ports(ep),
# w.Feature_discovered_ports_counts(ep),
# w.Feature_discovered_ports_sliding(ep),
# w.Feature_discovered_credential_count(ep),
# w.Feature_discovered_nodeproperties_sliding(ep),
])
self.node_specific_features = w.ConcatFeatures(
ep,
[
# w.Feature_actions_tried_at_node(ep),
w.Feature_success_actions_at_node(ep),
w.Feature_failed_actions_at_node(ep),
w.Feature_active_node_properties(ep),
w.Feature_active_node_age(ep)
# w.Feature_active_node_id(ep)
])
self.state_space = w.ConcatFeatures(
ep, self.global_features.feature_selection +
self.node_specific_features.feature_selection)
self.action_space = w.AbstractAction(ep)
def __init__(self, ep: EnvironmentBounds, qm: Optional[np.ndarray] = None):
self.ep = ep
self.state_space = w.HashEncoding(
ep,
[
# Feature_discovered_node_count(),
# Feature_discovered_credential_count(),
w.Feature_discovered_ports_sliding(ep),
w.Feature_discovered_nodeproperties_sliding(ep),
w.Feature_discovered_notowned_node_count(ep, 3)
],
5000
) # should not be too small, pick something big to avoid collision
self.action_space = w.RavelEncoding(
ep, [w.Feature_active_node_properties(ep)])
super().__init__("attack_source", self.state_space, self.action_space,
qm)
if debugging:
print(f"port_count = {ep.port_count}, property_count = {ep.property_count}")
gym_env.environment
# training_env.environment.plot_environment_graph()
gym_env.environment.network.nodes
gym_env.action_space
gym_env.action_space.sample()
gym_env.observation_space.sample()
o0 = gym_env.reset()
o_test, r, d, i = gym_env.step(gym_env.sample_valid_action())
o0 = gym_env.reset()
o0.keys()
fe_example = w.RavelEncoding(ep, [w.Feature_active_node_properties(ep), w.Feature_discovered_node_count(ep)])
a = w.StateAugmentation(o0)
w.Feature_discovered_ports(ep).get(a, None)
fe_example.encode_at(a, 0)
# %%
# Evaluate a random agent that opportunistically exploits
# credentials gathere in its local cache
credlookup_run = learner.epsilon_greedy_search(
gym_env,
ep,
learner=rca.CredentialCacheExploiter(),
episode_count=10,
iteration_count=iteration_count,
epsilon=0.90,
render=False,
o_test, r, d, i = cyberbattlechain_10.step(
cyberbattlechain_10.sample_valid_action())
o0 = cyberbattlechain_10.reset()
o0.keys()
# %%
ep = w.EnvironmentBounds.of_identifiers(
maximum_node_count=22,
maximum_total_credentials=22,
identifiers=cyberbattlechain_10.identifiers)
print(f"port_count = {ep.port_count}, property_count = {ep.property_count}")
fe_example = w.RavelEncoding(ep, [
w.Feature_active_node_properties(ep),
w.Feature_discovered_node_count(ep)
])
a = w.StateAugmentation(o0)
w.Feature_discovered_ports(ep).get(a, None)
fe_example.encode_at(a, 0)
iteration_count = 9000
training_episode_count = 50
eval_episode_count = 5
# %%
random_run = learner.epsilon_greedy_search(
cyberbattlechain_10,
ep,
learner=learner.RandomPolicy(),