基于加权孤立森林的容器逃逸实时检测算法OA
Weighted isolation forest based real-time container escape detection method
容器逃逸是云原生环境面临的最严重安全威胁之一,攻击者可利用内核漏洞或配置错误从容器突破到宿主机,进而控制整个集群.现有检测方法多基于静态规则或单点系统调用监控,难以识别复杂的序列化攻击行为和未知逃逸手法.为此,提出一种基于加权孤立森林的容器逃逸实时检测方法.首先利用 eBPF 技术实现容器运行时系统调用的低开销实时采集,其次设计基于滑动窗口的多维特征提取方法,融合统计特征、序列特征与时序特征,然后根据容器逃逸的攻击机理,将 mount、clone 等高危系统调用定义高危特征集,引入权重放大因子 γ 构建加权孤立森林模型,最后采用指数加权移动平均实现自适应阈值优化.实验结果表明,该方法能够有效检测已知和未知的容器逃逸攻击,从而为云原生环境下的容器安全防护提供了新的技术手段.
Container escape is one of the most critical security threats in cloud-native environments.Attackers can exploit kernel vulnerabili-ties or configuration errors to break out from containers to the host operation system,thereby gaining control over the entire cluster.Unfortu-nately,the existing detection methods are mostly based on static rules or single-point system call monitoring,making it difficult to identify complex sequential attack behaviors and unknown escape techniques.To address these issues,this paper proposes a real-time detection meth-od built around weighted isolation forest.First,we use eBPF to collect system call traces from running containers—lightweight,low overhead,no impact on actual workloads.Second,we apply a sliding window to extract multi-dimensional features,including statistical,sequential,and temporal information.Third,drawing on known attack mechanisms,we identify high-risk system calls(mount,clone,and so on)and intro-duce a weight factor γ to make the isolation forest pay more attention to them.Fourth,we add an adaptive threshold using exponential weighted moving average to keep the detector responsive.Experimental results demonstrate that the proposed method can effectively detect both known and unknown container escape attacks,providing a new technical approach for container security in cloud-native environments.
马建阳;王强
中铁建数智科技(北京)有限公司 网络安全部,北京 100039中国电子科技集团第四十一研究所 装备事业部,山东 青岛 266500
信息技术与安全科学
容器逃逸加权孤立森林系统调用序列自适应阈值
container escapeweighted isolation forestsystem call sequenceadaptive threshold
《网络安全与数据治理》 2026 (9)
53-59,7
工信部高质量发展专项项目(2023ZY01033)
评论