Anomaly Detection Requires Better Representations

Abstract

Anomaly detection seeks to identify unusual phenomena, a central task in science and industry. The task is inherently unsupervised as anomalies are unexpected and unknown during training. Recent advances in self-supervised representation learning have directly driven improvements in anomaly detection. In this position paper, we first explain how self-supervised representations can be easily used to achieve state-of-the-art performance in commonly reported anomaly detection benchmarks. We then argue that tackling the next generation of anomaly detection tasks requires new technical and conceptual improvements in representation learning.

A Appendix

In this paper we report anomaly detection results using the standard uni-modal protocol, which is widely used in the anomaly detection community. In the uni-modal protocol, multi-class datasets are converted to anomaly detection by setting a class as normal and all other classes as anomalies. The process is repeated for all classes, converting a dataset with C classes into C datasets. Finally, we report the mean ROC-AUC % over all C datasets as the anomaly detection results.

1.1 A.1 Anomaly detection comparison of MAE and DINO

We compare between DINO [5] and MAE [11] as a representation for a kNN based anomaly detection algorithm. For MAE, we experimented both with kNN and reconstruction error for anomaly scoring and found that the latter works badly, therefore we report just the kNN results. We evaluate using a variety of datasets, in the uni-modal setting described above. We used the following datasets:

INet-S [29]: The dataset is subset of 10 animal classes taken from ImageNet21k (e.g. “petrel”, “tyrannosaur”, “rat snake”, “duck”, “bee fly”, “sheep”, “beer cub”, “red deer”, “silverback”, “opossum rat”) that do not appear in ImageNet1K dataset. The dataset is coarse-grained and contains images relatively close to ImageNet1K dataset. It intended to convey that even for easy tasks the MAE doesn’t achieve as good results as DINO.

CIFAR-10 [18]: Consists of low-resolution $32\times 32$ images from 10 different classes.

CUB-200 [37]: Bird species image dataset which contains 11,788 images of 200 subcategories. In the experiment we calculated mean ROC-AUC% over the 20 first categories.

1.2 A.2 Multi-modal datasets

In these experiment we specify a single class as anomalous, and treat all images which does not contain it as normal.

MS-COCO-I [21]: We build a multi-modal anomaly detection dataset comprised of scenes benchmarks, where each image is evaluated against other images featuring similar scenes. We choose 10 object categories (“bicycle”,“traffic light”, “bird” , “backpack”, “frisbee”, “bottle”, “banana”, “chair”, “tv”, “microwave”, “book”) from different MS-COCO super-categories. To construct a multi-modal anomaly detection benchmark, we designate an object category from the list as the anomalous class, and training images of a similar super-category that do not contain it as our normal train set. Our test set contains all the test images from that super-category, where images containing the anomalous object are labelled as anomalies. This process is repeated for the 10 object categories resulting in 10 different evaluations. We report their average ROC-AUC %.

MS-COCO-O: We introduce a similar benchmark to MS-COCO-I, focusing on single objects rather than scenes. We crop all objects from our 10 super-categories (described above) according to the MS-COCO supplied bounding boxes. We repeat a similar process, using a similar object category as normal and the rest as anomalies.

CUB-200 [37]: We create a multi-modal anomaly detection benchmark based on the CUB-200 dataset. We focus on the 20 first categories, designating only one as an anomaly each time.

1.3 A.3 Tabular domain

Various datasets used for tabular data anomaly detection were used for the experiments. A total of 31 datasets from Outlier Detection DataSets (ODDS)¹ are employed. For the evaluation of GOAD and ICL we used the official repositories and made an effort to select the best configuration available. For all density estimation evaluations we used kNN with $k=5$ nearest neighbors. To convert GOAD and ICL into the standard paradigm of representation learning followed by density estimation: i) we use the original approaches to train a feature encoder (followed by a classifier which we discard) ii) we use the feature encoder to represent each sample iii) density estimation is performed on the representations using kNN exactly as in Sect. 3.