Anomaly Detection

Our most recent work includes the development of an unsupervised anomaly detector for multivariate time series. The method has been designed to meet constraints that guarantee deployment into practice. Concretely, the method needs to be highly scalable and stable.

To meet this goal, we have proposed USAD an encoder-decoder architecture within an adversarial training framework that allows to combine the advantages of autoencoders and adversarial training, while compensating for the limitations of each technique, making the resulting method fast to train, scalable and robust, while achieving a high performance.

All the code is available in Github

Students

Julien Audibert - CIFRE PhD with Orange

Publications

J. Audibert, P. Michiardi, F. Guyard, S. Marti, MA. Zuluaga. USAD : UnSupervised Anomaly Detection on Multivariate Time Series. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, August 23-27, 2020