Jihoon Shin

Jihoon Shin

he/his

Project Specialist at Syracuse University with 7 years of experience in IT systems management and 2 years of experience in machine learning research.

[CTU/UNCUYO/UTEP] Evaluating Generalization Mechanisms in Autonomous Cyber Attack Agents

[CTU/UNCUYO/UTEP] Evaluating Generalization Mechanisms in Autonomous Cyber Attack Agents

Organizations: Czech Technical University (CTU), Universidad Nacional de Cuyo (UNCUYO), and University of Texas at El Paso (UTEP)
Role: Research Assistant – Meta-Reinforcement Learning Implementation
Project Period: June 2025–Present

Overview
This project investigates the generalization problem of autonomous attacker agents in cybersecurity environments. It examines how reinforcement learning agents adapt their learned policies when facing new, unseen network topologies in realistic cybersecurity simulations.

The research focuses on:

  • Data exfiltration attack scenarios
  • Bridging the gap between theoretical RL assumptions and realistic network dynamics
  • Comparing different adaptation and generalization strategies across learning paradigms

Preprint:
Lukáš, O., Shin, J., Rivas, E., Forni, D., Rigaki, M., Catania, C., Piplai, A., Kiekintveld, C., & Garcia, S. (2026).
Evaluating Generalization Mechanisms in Autonomous Cyber Attack Agents.
arXiv:2603.10041.
https://doi.org/10.48550/arXiv.2603.10041

Official Repository:
https://github.com/stratosphereips/RedMorph

This work is an international collaboration with CTU & UNCUYO, and is currently in manuscript preparation.

Research Contributions

  • Redesigned the NetSecGame environment to support dynamically changing network topologies.
  • Developed a systematic training and evaluation methodology for attacker adaptation.
  • Evaluated and compared multiple learning strategies, including:
    • Deep Q-Networks (DQN)
    • Model-Agnostic Meta Learning (MAML)
    • Conceptual Q-learning and SARSA with state abstraction
    • Large Language Model (LLM)-based agents
  • Contributed to the development of an Explainable AI (XAI) framework to analyze action distributions and interpret generalization behavior.

My Contributions

  • Implemented MAML and Reptile meta-learning agents within the NetSecGame framework.
  • Conducted empirical evaluation comparing meta-learning agents.
  • Contributed to experimental analysis and manuscript development (in progress).