Removing Geometric Bias in One-Class Anomaly Detection with Adaptive Feature Perturbation

TL;DR

PLUME tackles one-class anomaly detection without relying on dataset-specific geometric biases by operating in frozen pretrained feature spaces and generating adaptive pseudo-anomalies through a linear perturbation map. A dedicated perturbator, built with a VAE, produces per-sample perturbations that softly modify feature vectors via $A_i = \bf{I} + \boldsymbol{\alpha}_i \boldsymbol{\beta}_i^\top$, while a compact classifier is guided by a contrastive loss to cluster normal embeddings and repel anomalies. Empirical results on CIFAR-10/100 and SPARK show state-of-the-art AUC performance, with ablations confirming the critical role of the linear perturbation and contrastive learning components. The approach is computationally efficient, interoperable with various backbones, and reduces reliance on geometry-biased data, making it suitable for real-world anomaly detection tasks across domains.

Abstract

One-class anomaly detection aims to detect objects that do not belong to a predefined normal class. In practice training data lack those anomalous samples; hence state-of-the-art methods are trained to discriminate between normal and synthetically-generated pseudo-anomalous data. Most methods use data augmentation techniques on normal images to simulate anomalies. However the best-performing ones implicitly leverage a geometric bias present in the benchmarking datasets. This limits their usability in more general conditions. Others are relying on basic noising schemes that may be suboptimal in capturing the underlying structure of normal data. In addition most still favour the image domain to generate pseudo-anomalies training models end-to-end from only the normal class and overlooking richer representations of the information. To overcome these limitations we consider frozen yet rich feature spaces given by pretrained models and create pseudo-anomalous features with a novel adaptive linear feature perturbation technique. It adapts the noise distribution to each sample applies decaying linear perturbations to feature vectors and further guides the classification process using a contrastive learning objective. Experimental evaluation conducted on both standard and geometric bias-free datasets demonstrates the superiority of our approach with respect to comparable baselines. The codebase is accessible via our public repository.

Figures (7)

• Figure 4: Performance of SotA methods on CIFAR-10 dataset. Methods on the left (blue background) achieve among the best results by exploiting a geometric bias in the dataset, i.e. typical object orientations. On the contrary, methods on the right (grey background - green dots) are less dataset-specific and are therefore considered as our baselines. Detailed results of all methods can be found in the supplementary.
• Figure 5: Study done on CIFAR-10 Bird class, with features extracted with ResNet50. Illustration of normal data (green) and generated pseudo-anomalies (red). Pseudo-anomalies were produced by the PLUME with the AddMult perturbation method.
• Figure 6: Study done on CIFAR-10 Bird class, with features extracted with ResNet50. Illustration of normal data (green) and generated pseudo-anomalies (red). Pseudo-anomalies were produced by a trained PLUME configuration.
• Figure 7: Study done on CIFAR-10 Truck class, with features extracted with ResNet50. Illustration of embedded vectors from normal data (blue), generated pseudo-anomalies (orange) and real anomalies (pink). Pseudo-anomalies were produced by training PLUME with the AddMult perturbation and without contrastive loss.
• Figure 8: Study done on CIFAR-10 Truck class, with features extracted with ResNet50. Illustration of embedded vectors from normal data (blue), generated pseudo-anomalies (orange) and real anomalies (pink). Pseudo-anomalies were produced by a trained PLUME configuration.