DYNAS adds defensive noise to reinforcement-learning penetration tests
Original title: DYNAS: A Dynamic Network Penetration Testing Framework Using Deep Hierarchical Reinforcement Learning
Pen-DHRL posts positive mean returns with defenses on from the start, while flat baselines stay negative.
DYNAS is a dynamic network attack simulator with large-scale scenarios, stateful intrusion-detection responses, scan false positives and negatives, temporary service outages, and action timeouts. Pen-DHRL is a recurrent manager-worker policy that picks persistent subgoals from IDS-conditioned context. Against GNN, Invariant, and Attention baselines, it reports higher reward per step and more captured hosts in selected transfer settings. With all defensive dynamics active from training, only Pen-DHRL shows positive mean episode returns in all four scenarios.
Unconfirmed: Reported gains are approximate simulation results, not measurements on live networks.
- After IDS-off training, reward per step is about 64% and 24% higher on Bridge and Dynamic.
- Captured hosts are about 26% and 30% higher on VarA and VarB than the next baseline.
- Flat baselines are negative in all 12 architecture-scenario pairs when defenses start on.
- Staged curriculum training yields positive mean returns in all 16 combinations.
From the original article
1 min read at arxiv.orgAs network scale and uncertainty complicate automated penetration testing, reinforcement learning (RL) environments must support controlled evaluation under defensive and non-stationary conditions.
By Md. Rakibul Hassan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.