Reproduce “NASimEmu: Network Attack Simulator & Emulator for Training Agents Generalizing to Novel Scenarios” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jaromír Janisch, Tomáš Pevný, Viliam Lisý

NASimEmu, a framework from the Czech Technical University in Prague, unifies a scalable network attack simulator with a realistic emulator to train deep reinforcement learning agents for penetration testing. It enables efficient agent training in simulation while verifying their real-world applicability through deployment in a live emulated environment, demonstrating that invariant architectures achieve better generalization to novel network scenarios compared to fixed-input models.

Key insights

  • A combined simulator-emulator framework can effectively bridge the reality gap in automated penetration testing, facilitating both efficient agent training and validation of their real-world applicability.
  • The choice of agent architecture is crucial for generalization; permutation-invariant models with shared weights outperform fixed-input MLP architectures when transferring to novel network topologies and sizes.
  • Designing training environments that vary structurally and evaluating agents on entirely unseen scenarios are essential for developing practical, robust penetration testing agents.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.