Env-Aware Anomaly Detection Ignore Style Changes Stay True to Content Stefan Smeu

2025-04-29 0 0 4.26MB 11 页 10玖币

侵权投诉

Env-Aware Anomaly Detection: Ignore Style Changes,

Stay True to Content!

Stefan Smeu∗

Bitdefender, Romania

University of Bucharest

ssmeu@bitdefender.com

Elena Burceanu∗

Bitdefender, Romania

eburceanu@bitdefender.com

Andrei Liviu Nicolicioiu

MPI for Intelligent Systems, Tübingen

andrei.nicolicioiu@tuebingen.mpg.de

Emanuela Haller

Bitdefender, Romania

ehaller@bitdefender.com

Abstract

We introduce a formalization and benchmark for the unsupervised anomaly detec-

tion task in the distribution-shift scenario. Our work builds upon the iWildCam

dataset, and, to the best of our knowledge, we are the ﬁrst to propose such an

approach for visual data. We empirically validate that environment-aware methods

perform better in such cases when compared with the basic Empirical Risk Mini-

mization (ERM). We next propose an extension for generating positive samples for

contrastive methods that considers the environment labels when training, improving

the ERM baseline score by 8.7%.

1 Introduction and related work

Identifying and following novelty [

] is an intriguing human ability that could trigger scientiﬁc

discoveries [

]. Machine learning models that can mimic this behavior and detect novelty when

facing unfamiliar situations are vital for ﬁelds like video surveillance [

], intrusion detection in

cybersecurity [

], manufacturing inspection [

], and many others [

]. Anomaly Detection (

)

is an umbrella term [

] for methods whose goal is to identify samples that deviate from an assumed

notion of normality. Normals and anomalies are supposed to come from different distributions. But

how and up to what limit do they differ? Deﬁning what changes constitute anomalies and what

changes should be ignored is essential.

Deep learning methods proved their representation power in multiple ﬁelds [

] and were assumed to become invariant to irrelevant aspects under the big data regime. Yet,

recent works proved that these representations are susceptible to unwanted biases [

] and prone to

ﬁnding shortcuts [

], relying on spurious features while failing to capture relevant aspects of the

data. Consequently, those models exhibit poor performance when dealing with slightly different,

out-of-distribution (OOD) settings, where spurious features are no longer informative. Avoiding

spurious correlations is a challenging problem, impossible to solve in the in-distribution (ID) training

setup [

]. Recent works [

] tackle this problem by using an informative

process that splits the dataset into multiple environments, extracting additional information. Except

for AnoShift [

], which focuses on network trafﬁc data, all those benchmarks and approaches

address supervised scenarios.

∗Equal contribution

Workshop on Distribution Shifts, 36th Conference on Neural Information Processing Systems (NeurIPS 2022).

arXiv:2210.03103v2 [cs.CV] 23 Nov 2022

Figure 1: OOD Content and Style Setup. a) Dataset: The samples’ input distribution varies on the

Content and Style axis. In training, we have only normal data, while in the testset we also have a

3rd (anomalous) class. b) Step 1: Pretraining algos learn its parameters using the training data, with

labels for the two normal classes. c) Step 2: AD methods use the embeddings learned in Step 1 to

transform their input.

We follow the same line to enable robust AD and exploit the multi-environments approach in

unsupervised anomaly detection from visual data. Moreover, we formalize the notion of anomaly

under the distribution shift paradigm.

To summarize, our main contributions are the following:

•

We propose a benchmark for unsupervised anomaly detection in images, focusing on real-

world cases, where the input distribution is different in sub-groups of data. We formally

emphasize the differences between anomaly detection and classical (supervised) distribution

shift analysis.

•

We validate that shallow AD methods can beneﬁt from working on top of embeddings

pretrained using environment-aware methods (like Fish, IRM, LISA). We prove consistent

improvements over ERM pretraining over a wide range of AD methods.

•

We introduce a way of adjusting the contrastive methods to be aware of multiple environ-

ments, making them more robust in out-of-distribution setups. Empirical validation over

MoCo v3, shows an 8.7% increase in ROC-AUC w.r.t. ERM, on iWildsCam dataset, in the

anomaly detection setup.

2 Generalization facets for Anomaly Detection

Latent factorization of the data

It is useful to formalize the data samples

as being determined by

two latent factors: Content and Style [

]. The

Content

should determine the task at hand, i.e.,

it should be the cause of the desired target. At the same time, the

Style

could represent unrelated

Figure 2: Positive samples in contrastive learning. In EA-MoCo, our env-aware baseline over Mo-

Cov3, for the positive samples we use: the basic augmented anchor (left) and the closest sample from

another, randomly chosen, environment (right). We compute distances over representations obtained

with an diffusion based autoencoder, learned under ERM, over samples from all environments.

features spuriously correlated with the target. Inferring these latent factors is an extremely difﬁcult

problem, seen as a goal of representation learning [

]. It is impossible in the unsupervised setting

without additional inductive biases, or other information [

] and it is outside our scope. Instead,

we start from a weaker assumption, that we have data in which only the

Style

is changed. We aim

to use this factorization in AD to highlight directions toward building methods that are robust to

irrelevant changes (involving Style) while capable of detecting relevant changes (involving Content).

Environments

We call domains or environments [

] sub-groups of the data, each with a different

distribution, but all respecting some common basic rules. Namely, the Content is shared, but Style or

relations involving Style change. Examples of domains include pictures taken indoor vs. outdoor

[

], or in different locations [

], real photos vs sketches [

], or images of animals with changing

associations in each environments [

]. Our goal is to be robust to the Style differences between

different environments while identifying the Content changes as anomalies.

2.1 Out-of-distribution regimes

When dealing with real-world data, the test distributions usually differ from the training ones,

encountering changes in Style or/and Content. We provide next an in-depth characterization of

possible scenarios for AD in those regimes, linking them to common methods that work in each

category for supervised tasks. For explicit examples and details, see Appendix A.2.

A. ID setting

The default paradigm in Machine Learning, both in supervised and unsupervised

learning. Although this is the default paradigm, the usual assumption that train and test data come from

the same distribution is very strong and almost never true for real-world datasets [9, 45, 12, 27, 18].

B. Style OOD

Most works that develop methods robust to some (i.e. Style) distribution changes

reside in this category [

]. Environments have differences based on Style, but have the

same Content and the goal is to learn representations that are invariant across environments.

C. Content OOD

The assumption here is that environments contain changes in distribution that are

always relevant (i.e. changes in Content) for the task and should be noticed. Methods in this category

must detect such changes while optionally performing another basic task. Anomaly, novelty, or OOD

detection methods work in this regime [48].

D. Style and Content OOD

Here, environments bring changes in both Content and Style. We argue

that this is the most realistic setting and it is mainly unaddressed in the anomaly detection literature.

An ideal anomaly detection method will only detect Content anomalies, while being robust to Style

changes. Our main analyses and experiments are performed in this setting, showing the blind spots of

current approaches and possible ways forward.

We formalize and detail the distribution shifting scenarios in Appendix A.2. To the best of our

knowledge, we are the ﬁrst to cover this topic for anomaly detection in particular and for unsupervised

learning in general.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Env-AwareAnomalyDetection:IgnoreStyleChanges,StayTruetoContent!StefanSmeuBitdefender,RomaniaUniversityofBucharestssmeu@bitdefender.comElenaBurceanuBitdefender,Romaniaeburceanu@bitdefender.comAndreiLiviuNicolicioiuMPIforIntelligentSystems,Tübingenandrei.nicolicioiu@tuebingen.mpg.deEmanuelaHallerBit...

展开>> 收起<<

Env-Aware Anomaly Detection Ignore Style Changes Stay True to Content Stefan Smeu.pdf

共11页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Env-Aware Anomaly Detection Ignore Style Changes Stay True to Content Stefan Smeu

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: