Scholarlink

<p>Penetration testing is an effective method for assessing cybersecurity posture. However, as cyber attacks become increasingly sophisticated and network environments grow in complexity, traditional intelligent penetration testing models encounter significant challenges in efficiently identifying optimal strategies and achieving rapid convergence. To address these issues, this paper proposes an automatic penetration testing model based on reinforcement learning, which abstracts the testing process as a Markov Decision Process (MDP) and incorporates a curiosity mechanism to enhance exploration efficiency. The core of this approach lies in training an agent using a reinforcement learning algorithm equipped with a curiosity mechanism, enabling it to automatically explore and select optimal penetration testing strategies based on reward feedback from the environment, and to persist until sensitive privileges are obtained. Experiments conducted in a simulated network environment with 50 hosts and millions of states demonstrate that the proposed model reduces convergence time by approximately 50% compared to the standard Proximal Policy Optimization (PPO) algorithm. By integrating GPU acceleration within a high-performance computing (HPC) framework and employing distributed parallel processing, the model effectively satisfies the real-time and scalability requirements of penetration testing in complex network environments.</p>

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

Automatic penetration testing model based on reinforcement learning for complex network environments

  • 期刊论文

  • 发表日期：

    2025年7月18日

查看全文
  • Yang Chen,
  • Junjiang He,
  • Wenbo Fang,
  • Shenwen Yang,
  • Jiangchuan Chen,
  • Tao Li,
  • Xiaolong Lan

摘要

Penetration testing is an effective method for assessing cybersecurity posture. However, as cyber attacks become increasingly sophisticated and network environments grow in complexity, traditional intelligent penetration testing models encounter significant challenges in efficiently identifying optimal strategies and achieving rapid convergence. To address these issues, this paper proposes an automatic penetration testing model based on reinforcement learning, which abstracts the testing process as a Markov Decision Process (MDP) and incorporates a curiosity mechanism to enhance exploration efficiency. The core of this approach lies in training an agent using a reinforcement learning algorithm equipped with a curiosity mechanism, enabling it to automatically explore and select optimal penetration testing strategies based on reward feedback from the environment, and to persist until sensitive privileges are obtained. Experiments conducted in a simulated network environment with 50 hosts and millions of states demonstrate that the proposed model reduces convergence time by approximately 50% compared to the standard Proximal Policy Optimization (PPO) algorithm. By integrating GPU acceleration within a high-performance computing (HPC) framework and employing distributed parallel processing, the model effectively satisfies the real-time and scalability requirements of penetration testing in complex network environments.