Deep Learning Advancements in Anomaly Detection: A Comprehensive Survey

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

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.

Introduction

Anomaly Detection (AD)—the task of identifying observations that deviate significantly from "normal" behavior—is critical for cybersecurity, industrial maintenance, and healthcare. As data grows in volume and complexity, becoming high-dimensional, unstructured, and highly temporal, traditional statistical methods like PCA or basic clustering often fail to capture the intricate patterns necessary for accurate detection.

Taxonomy of the survey covering data characteristics, applications, and deep learning methods for anomaly detection.

This survey reviews over 180 recent studies (2019–2024) to map the rapidly evolving landscape of deep learning-based AD. It categorizes modern techniques and highlights the shift toward unsupervised models and generative architectures capable of handling massive, complex data streams from the Internet of Things (IoT).

Core Challenges in Modern Anomaly Detection

The transition to deep learning (DL) in AD is driven by three primary challenges:

  1. The Labeling Bottleneck: Anomalies are inherently rare, making the creation of large, balanced, labeled datasets for supervised learning nearly impossible.
  2. Data Complexity: Modern datasets contain non-linear relationships and high-dimensional structures (e.g., video, sensor logs) that traditional models cannot effectively model.
  3. Interpretability: While DL models are powerful, their "black box" nature often lacks the transparency required for high-stakes decision-making in fields like finance or medicine.

Methodological Paradigms

Deep learning-based AD can be organized into three core paradigms: reconstruction-based, prediction-based, and hybrid frameworks.

Architectural frameworks for reconstruction-based, prediction-based, and hybrid anomaly detection approaches.

Reconstruction-based Approaches

These models, including Autoencoders, GANs, and the emerging Diffusion Models, are trained to "reconstruct" normal data. When presented with an anomaly, the model fails to recreate it accurately. The resulting "reconstruction error" serves as the anomaly score. Diffusion Models represent a significant advancement here, avoiding the "mode collapse" common in GANs and producing higher-fidelity reconstructions.

Prediction-based Approaches

Commonly used for time-series data, these models (RNNs, LSTMs, and Transformers) attempt to forecast the next data point in a sequence. If the observed value differs significantly from the prediction, it is flagged as an anomaly. This approach is particularly effective for capturing temporal dependencies.

Hybrid Approaches

Hybrid methods integrate traditional algorithms, such as Support Vector Data Description (SVDD) or K-means, with deep learning. Typically, a neural network acts as a feature extractor, while the traditional algorithm identifies outliers in the resulting simplified feature space. This combines the representational power of DL with the relative transparency of classical methods.

The survey identifies several pivotal shifts in the field:

  • The Rise of Diffusion Models: Diffusion Models offer a more robust way to learn the underlying distribution of "normal" data compared to VAEs or GANs.
  • Graph Neural Networks (GNNs) for Inter-dependency: For multivariate sensor data, GNNs treat each sensor as a node, allowing the detection of anomalies based on the relationships between sensors rather than just individual values.
  • Self-Supervised Learning: Due to the difficulty of labeling, successful recent models utilize self-supervision—creating internal labels from the data's own structure, such as predicting masked parts of a sequence.

Performance and Benchmarks

Quantitative improvements across the literature are substantial. For instance, Autoencoders have achieved a 26.83% relative AUC improvement over linear PCA on specific non-linear datasets. In time-series analysis, Transformer-based models have shown F1 score improvements of up to 17% while drastically reducing training time compared to RNN baselines. In visual AD, modern frameworks have pushed AUROC scores above 0.90 for complex multi-object scenes.

Future Directions

Despite these advancements, several "open frontiers" remain:

  • Computational Complexity: High-performing models like Transformers and Diffusion are often too resource-intensive for "edge" devices.
  • Explainability: Bridging the gap between detection and explanation remains a priority for industrial adoption.
  • Diverse Anomaly Types: While models excel at finding "point" anomalies (sudden spikes), they still struggle with "contextual" anomalies that require a deeper understanding of the broader environment.

Conclusion

Deep learning has fundamentally transformed anomaly detection, providing the tools necessary to navigate the complexities of modern data. By moving toward unsupervised, generative, and hybrid models, the field is becoming better equipped to handle the demands of real-world applications while striving for greater efficiency and interpretability.

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},
}
Figure