TELCO

Abstract—Anomaly detection in Multivariate Time-Series(MTS) data plays an important role in multiple domains,especially in cybersecurity, for the detection of unknownattacks. DC-VAE is a recent approach we have proposedfor anomaly detection in network measurement multivariatedata, which uses Variational Auto Encoders (VAEs) andDilated Convolutional Neural Networks (DCNNs) to modelcomplex and high-dimensional MTS data. However, detect-ing anomalies using VAEs can result in performance degra-dation and even catastrophic forgetting when trained on dy-namic and evolving network measurements, particularly inthe event of concept drifts. We extend DC-VAE to a continuallearning setup, leveraging the generative AI properties of theunderlying models to deal with continually evolving data.We introduce GenDeX, an approach to Generative AI-basedanomaly detection which compresses the patterns extractedfrom past measurements into a generative model that cansynthesize MTS data out of input Gaussian noise, mimick-ing the characteristics of the MTS data used for training.GenDeX relies on a Deep Generative Replay paradigm torealize continual learning, combining synthesized past MTSmeasurements with new observations to update the detectionmodel. Using a large-scale, multi-dimensional network mon-itoring dataset collected from an operational mobile InternetService Provider (ISP), we showcase the functionality of DC-VAE in the event of concept drifts, and study in-depth itsgenerative characteristics, assessing GenDeX syntheticallygenerated MTS examples. GenDeX enables DC-VAE adapt-ing to continually evolving data, overcoming the limitationsof catastrophic forgetting.Index Terms—Anomaly Detection, Generative AI, VAE, Mul-tivariate Time-Series, GenDeX