Energy Theft Detection in Multi-Tenant Data Centers with Digital Protective Relay Deployment

Document Type


Publication Date



© 2016 IEEE. High performance data centers serve as the backbone of the prevailing cloud computing paradigm. Among data centers with different operational structures, multi-tenant data centers (MTDCs) are increasingly popular among various internet service providers for the ease of deployment. Despite the offered benefits, MTDCs are vulnerable to various cyberattacks. An important cyberattack is energy theft which can be launched by malicious tenants to reduce monetary cost of the electricity consumption. It can be achieved through attacking smart meters in the data center to undercount the energy usage of the attacker. Since the attackers could consume an excessive amount of energy without incurring elevated utility cost, energy theft discourages frugality in terms of energy consumption, which is highly undesirable in the era of sustainable computing. Despite fruitful research results on MTDCs, none of existing works address energy theft. When energy theft occurs, it might be necessary for the data center operator to examine smart meter of all tenants to find the compromised ones which could induce excessive labor cost. Localization of energy theft detection is an effective way to limit the labor cost in detecting energy thefts in MTDCs. It can be facilitated through deploying Digital Protective Relays (DPR) in the data center where a DPR is a device for fault detection and event logging in the power system. In this paper, an anomaly rate range based dynamic programming algorithm is proposed for inserting minimal DPRs into the data center, where the anomaly rate range is computed using Minimum Covariance Determinant (MCD) algorithm. To the best of our knowledge, this is the first work addressing the energy theft issue in multi-tenant data centers. The simulation results demonstrate that our algorithm inserts 19.2 percent less DPRs into the data center compared to a natural baseline algorithm. Meanwhile, in an attempt to identify all energy theft cases, our DPR insertion solution requires 12.8 percent less tenants to be checked compared with the baseline algorithm. More importantly, we demonstrate that using MCD alone cannot achieve accurate detection while using DPR alone cannot handle collusive energy theft. In contrast, integrating DPR with MCD can achieve a high detection accuracy (of 97.6 percent) for collusive energy theft.

Publication Title

IEEE Transactions on Sustainable Computing