When Unsupervised Domain Adaptation meets One-class Anomaly Detection: Addressing the Two-fold Unsupervised Curse by Leveraging Anomaly Scarcity

TL;DR

This paper introduces the first fully unsupervised domain adaptation (UDA) framework for unsupervised anomaly detection (UAD) and proposes a pioneering solution to this curse, considered intractable so far, by assuming that anomalies are rare.

Abstract

This paper introduces the first fully unsupervised domain adaptation (UDA) framework for unsupervised anomaly detection (UAD). The performance of UAD techniques degrades significantly in the presence of a domain shift, difficult to avoid in a real-world setting. While UDA has contributed to solving this issue in binary and multi-class classification, such a strategy is ill-posed in UAD. This might be explained by the unsupervised nature of the two tasks, namely, domain adaptation and anomaly detection. Herein, we first formulate this problem that we call the two-fold unsupervised curse. Then, we propose a pioneering solution to this curse, considered intractable so far, by assuming that anomalies are rare. Specifically, we leverage clustering techniques to identify a dominant cluster in the target feature space. Posed as the normal cluster, the latter is aligned with the source normal features. Concretely, given a one-class source set and an unlabeled target set composed mostly of normal data and some anomalies, we fit the source features within a hypersphere while jointly aligning them with the features of the dominant cluster from the target set. The paper provides extensive experiments and analysis on common adaptation benchmarks for anomaly detection, demonstrating the relevance of both the newly introduced paradigm and the proposed approach. The code will be made publicly available.

Figures (5)

• Figure 1: Illustration of the two-fold unsupervised curse: (a) The decision boundary learned from the source set without any adaptation does not allow generalization to the target domain. (b) Direct alignment of the unlabeled target with the one-class source features leads to the confusion of normal and abnormal samples.
• Figure 2: Comparison of our setting with previous works: (a) supervised source anomaly detection with supervised domain adaptation ildr-2019, (b) unsupervised one-class source anomaly detection with few-shot domain adaptation msra-2021irad-2023iris-attack-detection-2022, (c)our considered setting: unsupervised one-class source anomaly detection with unsupervised domain adaptation.
• Figure 3: Overview of the proposed method: The top branch uses a trainable feature extractor with a DSVDD objective for one-class source data. The bottom branch clusters the features using a frozen CLIP visual encoder to identify the dominant feature cluster and align it with normal source representations. • and $\bigstar$ denote normal and anomalous samples respectively.
• Figure 4: Assessing the validity of anomaly scarcity assumption.
• Figure 5: K-Meanskmeans-1979 components variation on VisDA.