Deep Learning for Anomaly Detection: A Review

Comprehensive literature review

• The review distinguishes its in-depth formulation of methods from earlier reviews of conventional techniques and the high-level outlines in [26].It describes methods’ assumptions, objective functions, intuitions, and capabilities in addressing challenges.Source: §1 · p. 2

Problem nature and challenges

• The review frames anomaly detection’s problem complexities and largely unsolved challenges as the basis for assessing deep methods.It systematically reviews how current methods address these challenges.Source: §1 · p. 2

Categorization and formulation

• Current methods are organized into three frameworks and a hierarchical taxonomy spanning 11 modeling perspectives.The frameworks cover generic feature extraction, learning representations of normality, and end-to-end anomaly score learning.Source: §1 · p. 2

Future opportunities

• The review discusses future opportunities and their implications for addressing relevant challenges.It presents possible opportunities as a direction for further work.Source: §1 · p. 2

Source codes and datasets

• The work collects publicly accessible source code for nearly all method categories and real-world datasets with anomalies for empirical comparison benchmarks.Source: §1 · p. 2

2.1 Major Problem Complexities

• Anomalies can be unknown, heterogeneous, rare, and imbalanced, and may occur as point, contextual, or collective anomalies.Collective anomalies are anomalous as a group even when their individual members may appear normal.Source: §2.1 · pp. 2–3

2.2 Main Challenges Tackled by Deep Anomaly Detection

• CH1 is to reduce false positives and improve recall, since methods can miss sophisticated anomalies while incorrectly flagging normal instances.The paper emphasizes the cost of failing to spot anomalies.Source: §2.2 · p. 3
• CH2 is to detect anomalies hidden in high-dimensional or dependent data while capturing intricate feature interactions and preserving useful information in reduced representations.Dependencies may be temporal, spatial, or graph-based.Source: §2.2 · p. 3
• CH3 and CH4 are to learn generalizable normality and abnormality representations from scarce or incomplete labels, while remaining resilient to mislabeled data and unlabeled anomalies.The amount of noise can vary across datasets, and noisy instances may be irregularly distributed.Source: §2.2 · p. 4
• CH5 is to model conditional and group anomalies and detect cases whose abnormality emerges only by combining heterogeneous data sources.Existing methods mainly target point anomalies and single data sources.Source: §2.2 · p. 4
• CH6 is to explain anomaly decisions, especially in safety-critical settings where black-box detection can risk bias against under-represented groups.Deep methods can unify detection and explanation, learn task-tailored representations end-to-end, and capture complex structures across data types; balancing interpretability and effectiveness remains challenging.Source: §2.2 · pp. 4–5

3.1 Preliminaries

• Deep neural networks compose linear and nonlinear functions; common activation functions and layer types support MLP, CNN, and RNN architectures.Examples of recurrent architectures include vanilla RNN, GRU, and LSTM.Source: §3.1 · p. 5
• Deep anomaly detection learns a neural mapping to a representation space or directly learns anomaly scores, with direct scores increasing as instances become more anomalous.Representation learning requires a separate step to calculate anomaly scores; direct score learning takes raw inputs.Source: §3.1 · pp. 5–6

3.2 Categorization of Deep Anomaly Detection

• The taxonomy classifies deep anomaly detection methods into three main paradigms and 11 fine-grained categories.The proposed taxonomy also associates categories with the detection challenges they can address.Source: §3.2 · p. 6, Figure 1 · p. 6
• The paradigms separate feature extraction from detection, jointly learn representations of normality, or unify detection and neural anomaly-score learning end to end.Representation learning is divided by whether shallow anomaly measures guide learning, yielding seven fine-grained categories; end-to-end score learning has four subcategories.Source: §3.2 · p. 6, Figure 2 · p. 7

4 DEEP LEARNING FOR FEATURE EXTRACTION

• Deep feature-extraction methods map high-dimensional data into a lower-dimensional space, then apply an anomaly scorer that is independent of the learned mapping.The mapping typically reduces dimensionality from D to K, where D≫K.Source: §4 · p. 7
• Deep features can capture semantic-rich representations and nonlinear relations better than PCA and random projection.The review cites this capability in [14].Source: §4 · p. 7
• Pretrained models such as VGG support video anomaly detection through unmasking [66], which tests the separability of successive frame sets.The framework assumes the first set is normal; its training accuracy is expected to be high when subsequent frames are abnormal and low otherwise. VGG features pretrained on ILSVRC provide expressive appearance features [66], and dynamically updated video-frame sampling pools improve performance.Source: §4 · pp. 7–8
• Pretrained features also transfer across tasks: VGG-initialized one-class SVMs are fine-tuned for MNIST anomaly classification, while MNIST-pretrained ResNets support video-surveillance event detection.The MNIST dataset is cited as.Source: §4 · p. 8
• Explicitly trained extractors pair autoencoder, DBN, or convolutional-autoencoder embeddings with independent SVM, pseudo-class, or clustering-based scoring; graph embeddings also optimize reconstruction and neighbor distances.For video, three autoencoders separately represent appearance, motion, and joint appearance-motion features; one-class SVMs score each representation.Source: §4 · p. 8
• Lower-dimensional projections can reveal hidden anomalies and reduce false positives, but decoupled scoring may lose detection-relevant information and yield suboptimal scores.The approach can use semantic-rich features to reduce false positives (CH1); public pretrained models and off-the-shelf detectors make it easy to implement, although pretrained models are typically data-type-limited.Source: §4 · p. 8

• These methods couple feature learning with anomaly scoring rather than fully decoupling the modules, and fall into generic or anomaly measure-dependent feature learning.Source: §5 · p. 8

5.1.1 Autoencoders.

• Autoencoders compress and reconstruct data, using reconstruction error to flag poorly reconstructed anomalies; they are adaptable across data types but can overfit training noise and optimize compression rather than anomaly detection.Regularized variants and architectures for sequences, graphs, and images extend the approach, but may not prevent representations from being biased by infrequent regularities or outliers.Source: §5.1.1 · pp. 9–11

5.1.2 Generative Adversarial Networks.

• GAN-based methods score anomalies by how poorly the latent space can generate or reconstruct an instance; realistic generation can help, but training may collapse or fail to converge, and scores may be suboptimal.AnoGAN searches iteratively for a matching latent vector, while encoder-based methods such as EBGAN avoid that search; generators can also produce instances outside the normal-data manifold.Source: §5.1.2 · pp. 11–13

5.1.3 Predictability Modeling.

• Predictability modeling uses temporal context to predict sequence instances and treats prediction errors as anomaly scores, but it is limited to sequence data and can be computationally expensive.Conditioning predictions on elapsed instances also enables detection of temporal context-based conditional anomalies.Source: §5.1.3 · pp. 13–14

5.1.4 Self-supervised Classification.

• Self-supervised classifiers use transformed training images to learn synthetic classes and score test instances by agreement; in, negative-entropy scores generally outperformed average and maximum prediction probability.The approach works in unsupervised and semi-supervised settings, but its transformations are often image-specific and its scores may be suboptimal because they are not integrated into training.Source: §5.1.4 · pp. 14–15

5.2.1 Distance-based Measure.

• Distance-based methods optimize low-dimensional representations for a chosen distance score, addressing high-dimensional data, though computationally costly measures and weaknesses of the underlying scores remain limitations.One approach uses pseudo-labeled anomalies and normals to separate random nearest-neighbor distances with a margin; another uses distances between learned and randomly projected representations, with proximity-learning and boosting variants.Source: §5.2.1 · pp. 16–17

5.2.2 One-class Classification-based Measure.

• One-class methods jointly learn low-dimensional features and a compact normal-data description with SVM or SVDD objectives, but may underperform when normal-class distributions are complex.Deep one-class SVM variants [27] and deep SVDD support reconstruction losses and semi-supervised use; a small number of labeled normal and anomalous examples can help detect known and novel anomalies.Source: §5.2.2 · pp. 17–19

5.2.3 Clustering-based Measure.

• Deep clustering learns representations for a targeted clustering algorithm and uses cluster assignments as supervisory signals, but contamination can bias clustering and the resulting anomaly detection.Clustering losses can target k-means, spectral, agglomerative, or GMM methods, often with autoencoder reconstruction losses; GMM-based work [83, 179] combines reconstruction-error features with learned features to highlight anomalies.Source: §5.2.3 · pp. 19–20

6.1 Ranking Models

• Ranking models directly optimize scores from ordinal relations; MIL uses video-level labels to rank the highest-scoring segment in anomalous videos above normal videos, with smoothness and sparsity constraints.These methods can use coarse or limited labels, but may generalize poorly beyond labeled anomaly types; is an unsupervised exception.Source: §6.1 · p. 22

6.2 Prior-driven Models

• Prior-driven models optimize anomaly scores against an assumed distribution; the Gaussian deviation loss pushes normal scores toward the prior mean and anomalous scores at least m standard deviations away.The resulting scores can be interpreted with a Z-score confidence interval, but performance may suffer when the prior poorly fits the data.Source: §6.2 · pp. 23–24

6.3 Softmax Likelihood Models

• Softmax likelihood models treat rare events as anomalous and use negative event likelihood as the score; NCE approximates normalization, while scoring can capture feature interactions [30].The objective extends to anomalous paths in heterogeneous attributed bipartite graphs; interaction costs can grow with feature count, and performance depends on negative-sample quality.Source: §6.3 · pp. 24–25

6.4 End-to-end One-class Classification

• Adversarial one-class classifiers train a discriminator to separate normal data from generated pseudo-anomalies, using denoising autoencoders, complementary fringe samples, or boundary instances,,.Their generated references may not resemble unknown anomalies, GAN instability can cause fluctuating performance, and the approach is limited to semi-supervised detection.Source: §6.4 · pp. 25–27

7.1 Representative Algorithms

• Most representative methods are unsupervised or semi-supervised, use networks of no more than five layers, and favor (leaky) ReLU; augmentation, dropout, and pre-training remain under-explored.Different backbone networks can handle different input types, and most algorithms’ source code is publicly accessible.Source: §7.1 · p. 27

7.2 Datasets with Real Anomalies

• Studies often evaluate methods on popular classification datasets converted for anomaly detection [3] [48], which may not reflect real-world performance.The section identifies the lack of real-world datasets with real anomalies as a major obstacle to anomaly-detection development.Source: §7.2 · p. 27
• The authors summarize 21 publicly available real-world datasets with real anomalies, spanning varied domains and data types and selected for large scale and/or high dimensional complexity.They also provide a continuously updated collection of widely used anomaly-detection datasets, including pre-processed datasets from Table 3.Source: §7.2 · p. 27

8.1 Exploring Anomaly-supervisory Signals

• Deep anomaly detection needs anomaly-specific supervisory signals beyond generic reconstruction and GAN objectives, including sources with weaker distributional assumptions and domain knowledge.Anomaly-measure-dependent constraints can help but may inherit the measures’ assumptions.Source: §8.1 · pp. 28–29

8.2 Deep Weakly-supervised Anomaly Detection

• Deep weak supervision should use limited anomaly labels to improve detection and generalize to unknown anomalies, while also learning known anomaly classes from few examples.Unknown anomalies may fall beyond the span of labeled examples; few-shot anomaly classes can have different manifold features, unlike the class-specific structure assumed in general few-shot learning. Inexact labels and deeper architectures are further opportunities.Source: §8.2 · p. 29

8.3 Large-scale Normality Learning

• Large-scale normality learning could pre-train transferable representations on unlabeled data and fine-tune detectors semi-supervised, provided learning is robust to anomaly contamination.The approach may need to be domain-specific because anomalies and datasets differ across domains.Source: §8.3 · pp. 29–30

8.4 Deep Detection of Complex Anomalies

• Deep detection should expand beyond point anomalies to conditional and group anomalies, and explore multimodal detection by combining representations from different data sources [65].Conditional and group anomaly detection may require novel network layers or objective functions.Source: §8.4 · p. 30

8.5 Interpretable and Actionable Deep Anomaly Detection

• Interpretable and actionable deep detectors should explain their own decisions, since model-agnostic anomaly explanations may not reveal model mechanisms or support mitigation [41].The paper suggests exploring deep-model explanation and actionable knowledge discovery methods for this purpose.Source: §8.5 · p. 30

8.6 Novel Applications and Settings

• Opportunities include out-of-distribution detection that retains normal-class accuracy and using deep anomaly detection or curiosity learning to support sparse-reward reinforcement learning [60, 79, 127, 18, 19, 118].Curiosity rewards novelty or rarity; anomaly detection could help reinforcement learning, while curiosity techniques could also inform anomaly detection.Source: §8.6 · pp. 30–31
• Non-IID anomaly detection could address interdependent or heterogeneous abnormality, alongside applications such as adversarial-example detection [51], anti-spoofing, and rare-event detection.Considering relationships such as mutually reinforcing disease symptoms may reveal subtle anomalies masked in the data space.Source: §8.6 · p. 31

A LINKS TO OPEN-SOURCE ALGORITHMS

• Table 4 links to 23 open-source algorithms.Source: Table 4 · p. 31