[论文解读] Zero-Shot Anomaly Detection via Batch Normalization

Aodong Li, Chen Qiu+4 more|arXiv (Cornell University)|Feb 15, 2023

Anomaly Detection Techniques and Applications被引用 9

一句话总结

Introduces Adaptive Centered Representations (ACR), a zero-shot batch-level anomaly detection method that uses batch normalization and meta-training to generalize to unseen AD tasks across tabular and image data without retraining.

ABSTRACT

Anomaly detection (AD) plays a crucial role in many safety-critical application domains. The challenge of adapting an anomaly detector to drift in the normal data distribution, especially when no training data is available for the "new normal," has led to the development of zero-shot AD techniques. In this paper, we propose a simple yet effective method called Adaptive Centered Representations (ACR) for zero-shot batch-level AD. Our approach trains off-the-shelf deep anomaly detectors (such as deep SVDD) to adapt to a set of inter-related training data distributions in combination with batch normalization, enabling automatic zero-shot generalization for unseen AD tasks. This simple recipe, batch normalization plus meta-training, is a highly effective and versatile tool. Our theoretical results guarantee the zero-shot generalization for unseen AD tasks; our empirical results demonstrate the first zero-shot AD results for tabular data and outperform existing methods in zero-shot anomaly detection and segmentation on image data from specialized domains. Code is at https://github.com/aodongli/zero-shot-ad-via-batch-norm

研究动机与目标

• Address the challenge of adapting anomaly detectors to distribution drift without labeled training data for the new normal.
• Propose a lightweight zero-shot AD method that leverages batch normalization to align different distributions.
• Develop a meta-training scheme that enables automatic zero-shot generalization for unseen AD tasks.
• Provide theoretical justification via a generalization bound and demonstrate effectiveness across tabular and image domains.

提出的方法

• Use batch normalization as an adaptive mechanism to recenter and rescale batches so normal samples cluster near the origin while anomalies diverge.
• Train off-the-shelf deep anomaly detectors (e.g., DSVDD, neural transformation learning) on a meta-training set of interrelated distributions to learn batch-norm based adaptation.
• Introduce Meta Outlier Exposure by mixing training distributions with a fraction of data from other distributions to tighten the normal boundary.
• Formulate a loss incorporating both standard anomaly scores and inverse scores that act as supervised signals for normal vs. anomalous samples within meta-training batches.
• Provide a theoretical bound showing the generalization error to a new distribution is controlled by the total variation between transformed test and averaged training distributions.
• Demonstrate zero-shot AD on tabular data and achieve state-of-the-art anomaly segmentation on image datasets (MVTec AD) without task-specific retraining.

Figure 1: a) Demonstrations of concrete examples of a meta-training set and a testing distribution. It is not necessary for the meta-training set to include the exact types of samples encountered during testing. For instance, when detecting lions within geese, the training data does not need to incl

实验结果

研究问题

• RQ1 Can batch normalization, coupled with meta-training, enable zero-shot generalization of anomaly detectors to unseen normal distributions?
• RQ2 How does the proposed ACR framework perform across diverse data types (tabular, image) and anomaly rates without additional supervision at test time?
• RQ3 What theoretical guarantees exist for zero-shot generalization in this setting?
• RQ4 How does Meta Outlier Exposure influence the learned anomaly boundary during meta-training?

主要发现

• ACR enables automatic zero-shot generalization for AD by training with batch normalization across multiple distributions.
• ACR achieves zero-shot AD on tabular data and outperforms baselines on image data, including non-natural and medical images.
• ACR establishes new state-of-the-art in anomaly segmentation on the MVTec AD benchmark.
• A theoretical generalization bound is derived showing the gap to unseen distributions is bounded by the total variation between the test and training transformed distributions.
• Meta Outlier Exposure improves boundary tightness by treating samples from other distributions as anomalies during training.
• ACR demonstrates robustness to varying anomaly ratios and requires no task-specific retraining at test time.

Figure 2: Illustration of batch normalization for AD with two tasks $P^{\pi}_{1}$ and $P^{\pi}_{2}$ . The method (batch-)normalizes the data in $P^{\pi}_{j}$ separately. If each $P^{\pi}_{j}$ consists mainly of normal samples, most samples will be shifted close to the origin (by subtracting the resp