Unsupervised Surrogate Anomaly Detection

TL;DR

This work introduces surrogate anomaly detection, a framework for learning low-dimensional representations that capture normal data patterns without modeling full data densities. Central to the approach is DEAN, a Deep Ensemble ANomaly detector that uses a minimal surrogate $g_{DEAN}({x})=1$ and an ensemble of simple neural submodels to estimate anomaly scores via $score_F(x)=\frac{1}{|F|}\sum_{f_i\in F} \|f_i(x)-q_i\|^{power}$ with $q_i \approx 1$, enabling robust detection while mitigating trivial minima. The authors formalize five axioms to guide surrogate design and demonstrate that DEAN satisfies these criteria, achieving competitive results on 121 datasets against 19 baselines, and offering scalability through parallelizable submodels and feature bagging. Beyond benchmarks, DEAN shows adaptability for explainability, fairness, and adversarial robustness, indicating broad practical impact for unsupervised anomaly detection in high-dimensional settings.

Abstract

In this paper, we study unsupervised anomaly detection algorithms that learn a neural network representation, i.e. regular patterns of normal data, which anomalies are deviating from. Inspired by a similar concept in engineering, we refer to our methodology as surrogate anomaly detection. We formalize the concept of surrogate anomaly detection into a set of axioms required for optimal surrogate models and propose a new algorithm, named DEAN (Deep Ensemble ANomaly detection), designed to fulfill these criteria. We evaluate DEAN on 121 benchmark datasets, demonstrating its competitive performance against 19 existing methods, as well as the scalability and reliability of our method.

Figures (7)

• Figure 1: Example of surrogate anomaly detection: Anomalies are detected by learning a representation that encodes the regular patterns of normal data and measuring deviations from the expected behavior.
• Figure 2: Critical difference diagrams comparing the AUC-ROC performance. A lower rank indicates better performance, while algorithms with no statistically significant differences are connected by a horizontal line. DEAN is depicted in green, other deep learning algorithms in blue.
• Figure 3: (a) Overall runtime overview across all datasets; DEAN is depicted in green, other deep learning algorithms in blue, and shallow algorithms in yellow. (b) and (c) average AUC-ROC performance changes with varying ensemble size.
• Figure 4: Left: Repetition uncertainty for various surrogate algorithms, Right: DEAN performance with varied hyperparameters and the resulting uncertainty.
• Figure 5: Given complicated alinear data, the functions learned by three neural networks with relu activations are shown. The network without learnable shifts cannot capture the structure of the underlying data, while both a network with learnable shifts in each layer and a network with learnable shifts in all layers except the last can describe the alinearity.