Anomalous behavior detection is an important component in many applications. Anomalies can represent problematic situations where early detection is critical to make situational assessments in the event of unexpected conditions. For many problems, the state of the art in machine learning is batch learning. However, online anomaly detection algorithms have emerged as an alternative which offer rapid access to useful insights with fewer computing capacity requirements. Ubiquitous and accelerated data streams have led to the development of several machine learning algorithms that favor adaptive learning. Real-time digital environments necessitate tackling specific data analysis challenges such as copious volumes of accelerated, infinite data streams and the phenomenon of concept drift. These Big Data challenges also represent real opportunities for improving a multitude of processes. In this paper, a novel thresholding approach for online anomaly detection based on the long short-term memory recurrent neural network encoder-decoder architecture is proposed. Comparative results in applying various advanced algorithms on 5 streaming datasets from the UCI repository and 3 synthetic datasets are discussed. All datasets contain anomaly ground truth information to evaluate the effectiveness of algorithms in detecting anomalous instances in streaming data.

2022 IEEE International Conference on Big Data (Big Data), December 17-20, 2022, Osaka, Japan