A comparative evaluation of unsupervised anomaly detection techniques in smart water metering networks
Loading...
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Namibia
Abstract
Advances in electronics and wireless communication technologies have enabled the development of smart water metering networks (SWMNs). These networks enable water utilities to measure the water consumption of the connected households better than the traditional metering. Every household's traditional water meter is replaced by a smart meter. Smart meter communicates the household's water consumption to the utility. Data collected by the smart meters can easily become unreliable due to different data attacks such as data tampering, integrity attacks or flooding. Data
integrity has become a major concern in SWMNs hence this work is motivated by the demand for data integrity in this networks. Thus, different security algorithms and techniques needs to be put in place in order to detect which smart meter is compromised. Anomaly detection techniques are some of the security techniques that can be applied in SWMNs in order to protect all SWMN stakeholders. Although a lot of privacy and security solutions for attacks are proposed in smart grids, there is a dearth in security solutions for SWMNs in literature. This work is a comparative
evaluation of three unsupervised anomaly detection techniques in Smart Water Metering Networks (SMWNs). These techniques are k-Nearest Neighbor (kNN), cluster-based local outlier factor (CBLOF), and the histogram-based outlier score (HBOS). The comparative study aims at providing a better unsupervised anomaly detection technique that can be adopted in SWMNs. This work aimed to find a better anomaly detection technique that can be used in a SWMNs, to track
down the smart meters that are producing anomalies, the time the anomaly occurred, and lastly, to plot the consumer's profiles. Simulations are conducted in OMNeT ++ v5 .0 platform running on the Ubuntu 14.04OS. INET-3.4.0 framework is used on top of OMNeT ++ v5 .0 in order to simulate and model a realistic network. Data collected from the Tsumeb East area is used to configure the normal meter reading. Several scenarios depicting deviations from the normal conditions are
simulated. The performance of each technique is compared with one another. False positive rates (FPR), detection rate (DR) and accuracy rate are used as performance metrics. One-way Anova and Turkey HSD are used for statistical analysis. Simulations runs for 60s; the first 20s aim to build the intelligence for the algorithm and the remaining 40s as a test phase. Simulation results show the kNN achieves almost zero FPR throughout while giving almost 5% for DR in all simulation runs. On other hand, CBLOF achieves the highest DR between 95% - 100%, while
giving the worst performance in terms of FPR. This shows that CBLOF's anomaly score is too stringent as it penalizes even normal data points. The HBOS exhibits intermediate performance in either metrics. HBOS gives out the intermediate results. Statistical results for accuracy rate show that although HBOS gives intermediate results, it tends to run very close to CBLOF. kNN gives a high accuracy rate while CBLOF gives a lower accuracy rate since it struggles to identify normal
data. This works concludes that kNN is a better anomaly detection technique to be used in SWMNs because it has a better DR, low FPR and a high accuracy rate as long as the magnitudes for the anomalous readings are not too close to what is being considered normal. Future work will explore a hybrid anomaly detection technique. Future work will also look at how to distinguish anomalous data from normal changes in customer water usage habits.
Description
A thesis submitted in partial fulfilment of the requirements for the degree of Master of Science (Electronics and Computer Engineering)
Keywords
Anomaly detection, Smart water, Metering networks