Energy-Efficient Anomaly Detection With Primary and Secondary Attributes in Edge-Cloud Collaboration Networks

An energy-efficient anomaly detection is fundamental to maintain a healthy status of domain applications in edge-cloud collaboration networks. Generally, various kinds of multimodal sensory data capture heterogeneous attributes, where a certain attribute, called the primary one, may be more significant in detecting certain anomaly. This observation drives us to propose a novel energy-efficient anomaly detection mechanism, where attributes sensed by multimodal smart things (msts) are categorized as primary and secondary ones according to their relevance with the characteristic of this anomaly. This technique includes two steps: 1) an initial anomaly detection in single edge networks. Edge nodes associated with the primary attribute adopt a lightweight object detection model to initially detect the potential occurrence of this anomaly. Certain edge networks are determined where an anomaly is suspected and 2) an anomaly refinement with multimodal and multiattribute smart things in marginal edge networks. The cloud identifies and issues a specific query request to gather anomaly-aware sensory data from smart things with secondary attributes, for refining the detection accuracy of this anomaly, where an adaptive weighted fusion model is developed to analyze sensory data coupling of msts. The experimental results show that this technique performs better than the state of the art on the reduction of energy consumption and query time.