Deep Learning Advancements in Anomaly Detection: A Comprehensive Survey

Haoqi Huang, Ping Wang, Jianhua Pei, Jiacheng Wang, Shahen Alexanian, Dusit Niyato

TL;DR

This comprehensive survey reviews over 180 recent studies (2019–2024) to map the rapidly evolving landscape of deep learning-based Anomaly Detection (AD). It categorizes modern techniques into reconstruction-based, prediction-based, and hybrid frameworks, highlighting how the field is moving toward unsupervised models and generative architectures like Diffusion Models to handle the massive, complex data streams generated by the Internet of Things (IoT).

What Problem Does This Paper Address?

Anomaly Detection—the task of identifying observations that deviate from "normal" behavior—is critical for cybersecurity, industrial maintenance, and healthcare. However, traditional methods like PCA or basic clustering fail when faced with modern data that is high-dimensional, unstructured (images/video), or highly temporal (sensor logs).

The authors tackle three primary challenges:

  1. The Labeling Bottleneck: Anomalies are rare by definition, making it nearly impossible to build large, balanced, labeled datasets for supervised learning.
  2. Complexity: Modern data contains intricate, non-linear relationships that traditional statistical models cannot capture.
  3. The "Black Box" Nature of DL: While deep learning is powerful, it often lacks the interpretability required for high-stakes decision-making in fields like medicine or finance.

Methods

The paper organizes the vast array of DL-based AD into three core paradigms:

  • Reconstruction-based Approaches: These models (Autoencoders, GANs, and Diffusion Models) are trained to "reconstruct" normal data. When they encounter an anomaly, they fail to recreate it accurately. The "reconstruction error" serves as the anomaly score. The survey highlights Diffusion Models as a significant advancement over GANs, as they avoid "mode collapse" and produce higher-fidelity reconstructions.
  • Prediction-based Approaches: These models (RNNs, LSTMs, and Transformers) attempt to forecast the next data point in a sequence. If the actual observed value differs significantly from the prediction, it is flagged. This is particularly effective for time-series data where temporal dependencies are key.
  • Hybrid Approaches: These integrate traditional algorithms (like SVDD or K-means) with deep learning. For example, a neural network might be used for feature extraction, while a traditional clustering algorithm identifies outliers in that simplified feature space. This combines the "power" of DL with the "transparency" of classical methods.

Datasets

While the paper is a survey and does not introduce a new dataset, it identifies the primary benchmarks used to evaluate the 180+ surveyed papers:

  • Visual AD: MVTec AD (industrial inspection), MNIST, CIFAR-10, and GTSRB.
  • Time-Series/Multivariate: SMAP (Soil Moisture Active Passive), PSM (Pooled Server Metrics), and various IoT sensor datasets.
  • Cybersecurity: NIDS (Network Intrusion Detection System) datasets and various logs for fraud detection.

Key Insights and Ideas

  • The Rise of Diffusion: One of the most interesting "aha" moments is the shift toward Diffusion Models. Unlike VAEs, which can produce blurry results, or GANs, which are notoriously unstable to train, Diffusion Models offer a more robust way to learn the underlying distribution of "normal" data.
  • GNNs for Inter-dependency: For multivariate sensor data (like a smart factory), the authors highlight Graph Neural Networks (GNNs). GNNs treat each sensor as a node, allowing the model to detect anomalies not just in a single sensor's value, but in the relationship between sensors.
  • Self-Supervision is the Future: Because labeling is so difficult, the most successful recent models use self-supervised learning—creating their own "labels" from the structure of the data itself (e.g., predicting a masked part of a sequence).

Key Results

The survey synthesizes quantitative improvements reported across the literature:

  • Autoencoders vs. PCA: In specific non-linear datasets (like the Lorenz system), Autoencoders achieved a 26.83% relative AUC improvement over traditional linear PCA.
  • Transformers in Time-Series: Transformer-based AD models have demonstrated F1 score improvements of up to 17% on public datasets while simultaneously reducing training time by up to 99% compared to older RNN-based baselines.
  • Visual Accuracy: Modern VAD (Video Anomaly Detection) frameworks using bidirectional prediction have pushed AUROC scores above 0.90 for complex multi-object scenes.

Limitations and Future Work

The authors identify several "open frontiers" that the research community has yet to fully solve:

  • Computational Complexity: Many of the best-performing models (like Transformers and Diffusion) are too resource-heavy for "edge" devices (e.g., a small sensor on a remote pipeline).
  • Explainability: There is still a massive gap in making these models "explain" why they flagged a specific point as an anomaly, which is a dealbreaker for many industrial applications.
  • Handling Diverse Anomalies: Most models are good at finding "point" anomalies (sudden spikes) but struggle with "contextual" anomalies (a normal temperature that is only weird because it's occurring in the middle of winter). Future work needs to focus on models that understand broader context.

Summary updated

Abstract

The rapid expansion of data from diverse sources has made anomaly detection (AD) increasingly essential for identifying unexpected observations that may signal system failures, security breaches, or fraud. As datasets become more complex and high-dimensional, traditional detection methods struggle to effectively capture intricate patterns. Advances in deep learning have made AD methods more powerful and adaptable, improving their ability to handle high-dimensional and unstructured data. This survey provides a comprehensive review of over 180 recent studies, focusing on deep learning-based AD techniques. We categorize and analyze these methods into reconstruction-based and prediction-based approaches, highlighting their effectiveness in modeling complex data distributions. Additionally, we explore the integration of traditional and deep learning methods, highlighting how hybrid approaches combine the interpretability of traditional techniques with the flexibility of deep learning to enhance detection accuracy and model transparency. Finally, we identify open issues and propose future research directions to advance the field of AD. This review bridges gaps in existing literature and serves as a valuable resource for researchers and practitioners seeking to enhance AD techniques using deep learning.
BibTeX
@article{huang2025deep,
  author = {Haoqi Huang and Ping Wang and Jianhua Pei and Jiacheng Wang and Shahen Alexanian and Dusit Niyato},
  title = {{Deep Learning Advancements in Anomaly Detection: A Comprehensive Survey}},
  journal = {IEEE Internet of Things Journal},
  year = {2025},
  doi = {10.1109/jiot.2025.3585884},
}

Loading related content...

Figure