Anomaly detection in resource constrained environments and its applications

Abstract

Anomaly detection is a well-known problem that has been researched in numerous applications across domains such as in Fraud detection, System health monitoring, event detection. This Thesis focuses on an anomaly detection approach that is effective in resource-constrained environments. The approach adapts the usage of a well known anomaly detection approach called Isolation Forest in a manner that it is e↵ective in resource constrained environments with streaming data. Isolation Forest utilises the concept of sub-sampling in creating ordered BSTs that ultimately leads to anomaly detection. It is, however, a bulky approach that assumes the luxury of large storage space and is also ine↵ective with dynamic streaming data so common nowadays in varied application domains. In this work, we first present the Preprocessed Isolation Forest (P iF orest) approach for anomaly detection that works well in resource constrained environments and is also e↵ective on streaming data. P iF orest, as stated earlier, is based on the Isolation Forest algorithm but has substantially low storage and prediction complexity. There are several important application domains for the proposed P iF orest technique. One such area that we dwell upon in this thesis is Wireless Body Area Networks. Wireless Body Area Network (WBAN) is a network of computing devices that are either implanted in or surround the body surface of a patient and provide continuous health monitoring. We utilise the proposed P iF orest technique as part of a two-level lightweight adaptive approach to discard unusual faulty measurements by the body sensors and to generate an alarm only when the patient enters an emergency situation. The two levels of the application comprise: fault detection using an adaptive mean-variance approach on streaming data with a constant amortized time and space complexity of O(1); the second level anomaly detection using the proposed P iF orest approach. The efficacy of the proposed P iF orest approach for anomaly detection in a constrained environment on streaming data and its application in a WBAN environment is evaluated with standard datasets. The results indicate performance at par with existing state-of-the-art techniques even when the latter is working in an unconstrained environment with static data. Subsequently, we implement the algorithm on a real world resource constrained environment comprising a Wireless Sensor Network node and demonstrate its practicability. We also test and validate the working of the approach in the mentioned WBAN application with real physiological datasets. Keywords: Anomaly Detection, Concept Drift, Isolation Forest, Sliding window, Streaming data, PiForest, Fault Detection, Body Area Networks.