Network_Defense_Symmetric_Competitive
收藏资源简介:
该数据集是一个用于多智能体强化学习研究的模拟环境数据集,包含1.024亿个时间步。数据集模拟了网络安全攻防对抗场景,包含两个对抗性智能体:红方智能体(攻击者)和蓝方智能体(防御者)。红方智能体的目标是发现系统漏洞、提升权限、破坏关键资产并维持持久性访问,其动作空间基于MITRE ATT&CK攻击框架的各个阶段进行建模。蓝方智能体作为防御者,目标包括维护系统可用性、减少攻击面、检测恶意行为、使用蜜罐技术欺骗攻击者以及最终驱逐入侵者,其动作空间由各种防御配置和缓解措施组成。该环境建模为纯零和博弈,即红方智能体的收益直接等于蓝方智能体的绝对损失,反之亦然。数据集适用于网络安全领域的多智能体强化学习算法研究、攻防策略分析以及对抗性AI系统的训练与评估。
This dataset is a simulation environment dataset for multi-agent reinforcement learning research, containing 102.4 million time steps. It simulates a cybersecurity attack and defense adversarial scenario, featuring two adversarial agents: the Red Agent (attacker) and the Blue Agent (defender). The Red Agents objectives are to discover system vulnerabilities, escalate privileges, compromise critical assets, and maintain persistent access, with its action space modeled based on various stages of the MITRE ATT&CK attack framework. The Blue Agent, as the defender, aims to maintain system availability, reduce the attack surface, detect malicious behavior, deceive attackers using honeypot techniques, and ultimately expel intruders, with its action space composed of various defense configurations and mitigation measures. The environment is modeled as a pure zero-sum game, meaning the Red Agents gain directly equals the absolute loss of the Blue Agent, and vice versa. The dataset is suitable for research in multi-agent reinforcement learning algorithms in cybersecurity, analysis of attack and defense strategies, and training and evaluation of adversarial AI systems.
数据集概述
数据集名称: Network_Defense_Symmetric_Competitive
许可证: MIT
语言: 英语
数据规模: 1亿至10亿时间步(100M < n < 1B)
内容描述: 该数据集包含 102,400,000 个时间步,用于多智能体强化学习场景,是一个纯零和博弈数据集。其中包含两个对立的智能体角色:
- 红方智能体(攻击方): 目标是发现漏洞、提升权限、危害资产并维持持久性。其动作空间可参考 MITRE ATT&CK 框架的各阶段建模。
- 蓝方智能体(防御方): 目标是维护系统可用性、减少攻击面、检测恶意行为、使用蜜罐欺骗攻击者并驱逐入侵者。其动作空间包括防御配置和缓解措施。
博弈性质: 纯零和博弈,红方的收益即为蓝方的绝对损失,反之亦然。
相关资源:
- 防御方智能体演示视频:https://www.youtube.com/watch?v=3g3plOjZkJw




