Securing Data Center Against Power Attacks

Modern data centers employ complex and specialized power management architectures in the pursuit of energy and thermal efficiency. Interestingly, this rising complexity has exposed a new attack surface in an already vulnerable environment. In this work, we uncover a potent threat stemming from a compromised power management module in the hypervisor to motivate the need to safeguard the data centers from power attacks. HyperAttack — an internal power attack —maliciously increases the data center power consumption by more than 70 % , while minimally affecting the service level agreement. We propose a machine learning-based secure architecture, SCALE , to detect anomalous power consumption behavior and prevent against power outages due to HyperAttack escalations. SCALE delivers 99 % classification accuracy, with a maximum false positive rate of 3.8 % .