UC Irvine

Anomaly detection is important in various applications, from cyber-security, transportation, industry, and finance to healthcare. The anomaly detection problem is to identify anomalies originating from a different data-generating process from normal data. The rare occurrence of anomalies and their unknown causes makes it hard to collect and model them. Thus, anomaly detection methods utilize normal data to build anomaly detectors. In this dissertation, we apply deep anomaly detection methods--methods that apply deep learning techniques--to solve anomaly detection problems. We contribute multiple generic frameworks for various anomaly detection setups.

First, we challenge the common clean training data assumption (free of anomalies) and stress that practical training data is often contaminated with unnoticed anomalies. We propose a novel unsupervised training strategy for training an anomaly detector in the presence of unlabeled anomalies that is compatible with a broad class of models.

Second, selecting informative data points for expert feedback can significantly improve anomaly detection performance. The critical challenges are selecting the most informative samples for expert review and effectively incorporating their feedback to bolster anomaly detection capabilities. To address these challenges, we propose a new data labeling strategy and a new learning framework for active and semi-supervised anomaly detection.

Third, real-world applications may face distribution shifts. We consider the online learning problem where the shifts occur at unknown positions and with unknown intensities. We derive a new Bayesian online inference approach to automatically infer these distribution shifts and adapt the model to the detected changes. This approach applies to both supervised and unsupervised learning settings. We also consider the problem of adapting an anomaly detector to drift in the normal data distribution, especially when no training data is available for the “new normal.” This setting is called zero-shot anomaly detection. We propose a simple yet effective method that combines batch normalization and meta-training for zero-shot anomaly detection.

The learning frameworks introduced in this dissertation are model-agnostic and apply to various data types. Extensive experiments demonstrate the efficacy of our proposed approaches.