Advancing anomaly detection with one-class classification and proxy tasks : a focus on video irregularities and face forgeries

Anomaly detection has been a fast-evolving research topic since the 19th century, focusing on identifying rare or unexpected patterns that deviate significantly from the majority of observations. In modern society, it has seen growing relevance in domains like healthcare, cybersecurity, public safety, and environmental monitoring. However, detecting anomalies remains challenging due to their defining attributes— uncertainty, diversity, and rarity. These attributes make it infeasible to enumerate all possible anomalous patterns or to assemble balanced datasets for model training, ultimately limiting the generalization capability of detection methods. In this thesis, we seek to address these challenges using the concepts of one-class classification and proxy tasks, targeting two representative domains in anomaly detection: video irregularities and face forgeries. For both, we propose detection frameworks that rely solely on normal data during training and reformulate the detection task through well-crafted proxy objectives. These frameworks advance state-of-the-art performance in both domains and contribute methodological innovations in recurrent networks, vision transformers, and pseudo-fake generation. This thesis consists of three parts: (1) Chapters 1 and 2 introduce the research topic and relevant background; (2) Chapters 3, 4, and 5 present three anomaly detection frameworks developed in this work, each implementing one-class classification and proxy tasks. Two of them detect video irregularities through frame prediction, with one centered on Convolutional Long Short-Term Memory and the other on vision transformers. The third detects face forgeries via patch identification, which leverages patch-based pseudo-fakes generated by an inpainting algorithm. Each framework is examined in terms of theory, architecture, and experimental outcomes; (3) Chapter 6 summarizes our contributions and outlines future research directions.