Scholarlink

We present a deep reinforcement learning model that enables agents to learn optimal counter-strategies in partially observable games without prior knowledge of opponent behavior. Using FlipIt, a two-player security game where players compete for resource ownership with limited state information, we combine deep neural networks with Q-learning to maximize the defender’s resource ownership time. Our approach learns solely through observing opponent moves and adapting to their patterns. We derive the theoretical optimal strategy for defenders facing attackers with unknown probability distributions and demonstrate that our model learns cost-effective counter-strategies aligned with theoretical predictions. This work advances reinforcement learning applications in partially observable security games with implications for cybersecurity.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Reinforcement Learning for Short-Duration FlipIt Security Game

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sie Hendrata Dharmawan,
  • Peter Chin

摘要

We present a deep reinforcement learning model that enables agents to learn optimal counter-strategies in partially observable games without prior knowledge of opponent behavior. Using FlipIt, a two-player security game where players compete for resource ownership with limited state information, we combine deep neural networks with Q-learning to maximize the defender’s resource ownership time. Our approach learns solely through observing opponent moves and adapting to their patterns. We derive the theoretical optimal strategy for defenders facing attackers with unknown probability distributions and demonstrate that our model learns cost-effective counter-strategies aligned with theoretical predictions. This work advances reinforcement learning applications in partially observable security games with implications for cybersecurity.