Continual Supervised Anomaly Detection
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: continual learning, anomaly detection, generative replay
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose a new continual learning method for supervised anomaly detection using generative replay with a VAE-based model.
Abstract: This paper proposes a continual-learning method for anomaly detection when a few labeled anomalies are available for training in addition to normal instances. Although several continual-learning methods have been proposed for anomaly detection, they have been dedicated to unsupervised anomaly detection, in which we can use only normal instances for training. However, few anomalies, which are valuable for constructing anomaly detectors, are often available in practice. In our continual-learning method, we use a hybrid model of a Variational AutoEncoder (VAE) and a binary classifier, and compute the anomaly score from the outputs of both models. The VAE is trained by minimizing the reconstruction errors of training data to detect unseen anomalous instances, and the binary classifier is trained to identify whether the input is a seen anomaly. Combining these two models enables us to efficiently detect both seen and unseen anomalies. Furthermore, the proposed method generates anomalous instances in addition to normal instances for generative replay to reduce the negative effects of catastrophic forgetting. In generative replay, anomalous instances are more difficult to generate than normal instances because few anomalous instances are available for training in anomaly detection. To overcome this problem, we formulate the generation of anomalous instances as an optimization problem, in which we find a latent vector of the VAE corresponding to anomalous instances, and generate anomalies by solving it using gradient descent. Our experimental results show that the proposed method is superior to anomaly detection methods using conventional continual learning.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5199
Loading