Evaluating Energy and Thermal Efficiency of Anomaly Detection Algorithms in Edge Devices

Abstract

Although IoT delivers several benefits, it also raises concerns regarding privacy and security, from revenue disruption in industrial facilities to life-threatening situations caused by smart houses hacking. As a consequence, anomaly detection algorithms stand out to improve data reliability. However, little has been said about the implications of running these computationally expensive programs in hardware-constrained edge devices. Therefore, in this paper, we present an evaluation of six anomaly detection algorithms running in an edge device regarding performance, accuracy, temperature, and power consumption. The results showed that time complexity, resources demand, and detection approach directly impact on the feasibility of running anomaly detection algorithms in edge devices. Based on these results, we present a recommendation on which algorithms would best satisfy the requirements of several IoT environments.

Comparative analysis of anomaly recognition methods in real time

“…Optimizing existing systems used to detect anomalies can greatly improve their performance [18]. On the other hand, on the basis of the analysis performed, the methods themselves can be optimized, as well as combined methods can be created that make it possible to level the existing costs according to the principle of complementarity.…”

Detecting abnormal sensors via machine learning: An IoT farming WSN-based architecture case study

“…We did perform an empirical evaluation of the proposed architecture using six unsupervised anomaly detection algorithms with support to the analysis of multivariate data sets [16]. An overview of these algorithms is presented next and a summary categorizing each algorithm is provided in Table 1.…”

TiWS-iForest: Isolation Forest in Weakly Supervised and Tiny ML scenarios

“…This task is usually unsupervised, meaning the labels describing if a sample is anomalous, are few or totally absent. In this context the Isolation Forest (IF) is a very appealing algorithm due to its good detecting performances compared to its algorithmic complexity [8]. However, in the case of ultra-constrained devices, even small improvements in the algorithm can make a difference: the detection performance is only one of the factors that might be considered in the choice of an algorithm to be run on an edge device; other factor might be memory, latency, computational power and energy cost to run on battery power [8].…”