Titre de la dissertation: Deep Visual Anomaly Detection under Data Contamination and Anomaly Heterogeneity.

Promoteurs: Monsieur Stéphane Dupont, Monsieur Souhaib Ben Taieb

Résumé de la dissertation : Anomaly detection (AD) is the task of identifying rare and unusual events that deviate from expected behaviour. A standard approach involves learning a « compact  » representation of the normal samples. Then, anomalies are identified as deviations from this learned normality. Its importance spans a wide range of safety-critical domains, including medical diagnosis, fraud detection, industrial inspection, and predictive maintenance of complex systems. Despite recent progress in deep learning, the practical use of deep AD models is often hindered by fundamental challenges that limit their reliability and adoption. This thesis tackles four of the most pressing challenges: (i) the co-occurrence of different anomaly types, (ii) the presence of contamination in training and test datasets, (iii) the non-stationarity of normal data distributions, and (iv) the lack of performance guarantees on risk functions such as false positive rate.The first challenge arises because real-world anomalies are rarely of a single type. In industrial inspection tasks, different types often appear together, which makes detection especially difficult. To address this issue, we develop a unified framework that detects both structural and logical anomalies by combining deep feature reconstruction with novel constraints that capture logical consistency. Experiments on multiple benchmark datasets confirm its effectiveness in detecting complex, co-occurring anomalies.The second challenge is contamination in the data used to train and evaluate AD models. It violates the common assumption that we have access to « clean » datasets consisting only of normal samples during training. In practice, unnoticed anomalies often contaminate the data. This can result in biased models that struggle to distinguish between normal and anomalous instances reliably. To address this, the thesis contributes two complementary solutions. First, we reframe AD as a semi-supervised classification task and introduce risk-based methods that use small sets of labelled samples, together with unlabelled data, to correct for contamination. These methods include a shallow model with a regularised unbiased risk estimator and a deep learning method with a nonnegative risk estimator, both of which are supported by theoretical guarantees. Second, we introduce a test-time adaptation framework that dynamically adjusts outputs of a pre-trained AD model trained on a contaminated dataset without requiring any new labels. By combining prior knowledge captured by a pre-trained model with evidence from incoming data at test-time, this method allows AD systems to remain reliable even under adverse conditions.The third challenge is non-stationarity, where the statistical properties of normal data evolve over time. We explore this challenge through the application of AD in solar power plants using thermal images. The data exhibits strong seasonal patterns, irregular sampling, and temporal dependencies. To address this challenge, we propose a forecasting-based approach that uses deep sequence models to predict future thermal images from past observations. Anomalies are then detected when observed images deviate significantly from predictions. This approach directly captures spatio-temporal patterns while accounting for irregular sampling and temporal dependencies, making it well-suited for complex industrial monitoring tasks.The fourth challenge concerns the reliability of predictions made by deep AD models. In critical applications, stakeholders require not only accurate predictions but also performance guarantees on risk functions such as the false positive rate. While recent methods for uncertainty estimation provide partial insight into the model performance, they often fail to guide practical decision-making. To overcome this, we develop a risk-controlling thresholding framework that provides distribution-free, finite-sample guarantees on any chosen risk measure. This strategy is based on a novel density forecasting model leveraging conditional normalising flows to estimate the likelihood of an observation being normal given past data. We further enhance this framework by introducing adaptive thresholds that better separate normal and anomalous cases, and incorporating an abstention mechanism to defer highly uncertain predictions to domain experts.In summary, this thesis addresses four core barriers to the reliable use of deep AD. Through new frameworks, theoretical analysis, and extensive validation in real-world applications such as industrial inspection and solar power plant monitoring, this work aims to make AD not only more accurate and robust but also more transparent and trustworthy.