Anomaly Detection using Contrastive Normalizing Flows

Detecting test data deviating from training data is a central problem for safe and robust machine learning. Likelihoods learned by a generative model, e.g., a normalizing flow via standard log-likelihood training, perform poorly as an anomaly score. We propose to use an unlabelled auxiliary dataset and a probabilistic outlier score for anomaly detection. We use a self-supervised feature extractor trained on the auxiliary dataset and train a normalizing flow on the extracted features by maximizing the likelihood on in-distribution data and minimizing the likelihood on the auxiliary dataset. We show that this is equivalent to learning the normalized positive difference between the in-distribution and the auxiliary feature density. We conduct experiments on benchmark datasets and show a robust improvement compared to likelihood, likelihood ratio methods and state-of-the-art anomaly detection methods.