Online Deep anomaly Detection – ODD

Anomaly detection is a challenge per se. It is unsupervised by nature, as abnormal events are rare, varied, and cumbersome to collect. By exploring deep neural networks for representation learning, the main categories of anomaly detection methods are deep one-class classifiers, autoencoders, generative adversarial networks, and self-supervised learning methods. The challenge is even more ambitious in anomaly detection in streaming data. Indeed, time series are prone to specificities, such as seasonality, non-stationarity, or conceptual drift. Moreover, anomalies strongly depend on the temporal contextualization and early detection must be done in real time.

The objective of this project is to address the main challenges encountered in deep anomaly detection in streaming data. Two online detection problems may arise depending on whether the anomaly is short-termed or sustainable: point anomaly detection or change point detection. In this perspective, three research directions are investigated in this project:
- We address recent advances in anomaly detection according to the online detection paradigm, by proposing two original strategies: Firstly, we revisit deep representations for anomaly detection in the light of the specificities of time series; Secondly, we explore an adaptive ensemble deep learning for anomaly detection in evolving data streams.
- We explore optimal transport theory to provide online processing methods for anomaly detection and for the characterization of the change point and the distribution shift in the case of a concept drift.
- We aim to leverage the theoretical underpinnings of the change point detection literature, in order to implement them in a deep representation space by using advances in generative models in deep learning.

While the major contributions will essentially be fundamental in Machine Learning, the project will address several application areas. In addition to the time series from sensors, we are interested in complex multivariate time series for condition monitoring of rotating machines, in detecting environmental pollution in bio-sensors signals, as well as in the integration of the spatial contextualization for anomaly detection in images.