电力攻击已成为数据中心面临的关键威胁之一。电力攻击通过恶意提高服务器功耗并使之超过配电系统功率极限，引起服务器故障或断路器动作，甚至造成整个数据中心供电中断。目前尚未有成熟可靠的电力攻击检测技术。随着多租户、虚拟化、容器化等技术在数据中心大量部署，进一步加大了用户行为识别的复杂性和电力攻击检测的难度。本课题拟从多租户环境下虚拟机功耗识别、电力攻击特征识别与提取、电力攻击防护与分级电力管理、电力平衡与电力安全等方面展开研究。本课题基于分布式机器学习方法构建典型负载下数据中心电力使用模式，并对电力攻击的行为模式进行建模分析和挖掘，提出数据中心电力攻击检测方法，根据数据中心服务级协议构建数据中心电力安全防护与电力平衡策略。本课题将依托国网电力信息通信公司的数据中心运行日志与电力使用数据，构建数据中心电力攻击检测原型系统。本课题的研究将对数据中心运营、信息系统安全等提供参考和决策依据。

Power attack is emerging as one of the key threats in data centers. Power attacks try to increase servers’ power consumption maliciously enough to exceed the upper limit of electricity provisioning system such that the breakers will trip to discard electric workload or servers will fail. In some cases, power attacks can even result in power outage in the whole data center. Currently there are no maturely reliable solution for power attacks in data centers. Moreover, deployment of multi-tenant provisioning, virtualization, and containerization in data centers increases the complexity to identify the guest users’ behavior and the difficulties to detect power attacks. This project tries to investigate characterization on power attacks in data centers, virtual machine level power usage accounting, and classified multilevel power provisioning policies. Firstly, we try to characterize power usage patterns of typical workloads in data centers using distributed machine learning, and then we will conduct data mining and analysis of power attacks and derive power attacks models. Secondly, we will propose virtual machine level detection approaches for power attack detection. Finally, we will propose classified multilevel power provisioning policies to secure power provisioning in data centers. We will cooperate with experts from the State Grid and build prototype based on real world data center and its power usage patterns. The expected research outputs of this project may help both data center operators and system security stakeholders.