Abstract
Anomalies are intuitively easy for human experts to understand, but they are hard to define mathematically. Therefore, in order to have performance guarantees in unsupervised anomaly detection, priors need to be assumed on what the anomalies are. By contrast, active learning provides the necessary priors through appropriate expert feedback. Thus, in this work we present an active learning method that can be built upon existing deep learning solutions for unsupervised anomaly detection, so that outliers can be separated from normal data effectively. We introduce a new layer that can be easily attached to any deep learning model designed for unsupervised anomaly detection to transform it into an active method. We report results on both synthetic and real anomaly detection datasets, using multi-layer perceptrons and autoencoder architectures empowered with the proposed active layer, and we discuss their performance on finding clustered and low density anomalies.
TL;DR
This paper introduces the UAI (Unsupervised to Active Inference) layer, a simple yet powerful "plug-and-play" component that transforms unsupervised deep anomaly detection models into active learning systems. By incorporating a small amount of human expert feedback, the UAI layer allows models to overcome the mathematical ambiguity of what constitutes an "anomaly," consistently outperforming state-of-the-art methods across diverse datasets.
The Challenge: Defining Anomalies
Anomaly detection (AD) is fundamentally difficult because "anomalies" lack a universal mathematical definition. In an unsupervised setting, a model must rely on internal "priors" or assumptions. For example, is a small, dense cluster of data points an anomaly (like a coordinated cyberattack) or just a specific sub-type of normal data?
Purely unsupervised models often fail because they are biased toward finding only one type of anomaly—usually either "low-density" points (outliers) or "clustered" anomalies—but rarely both. The authors argue that anomaly detection is often an undecidable problem without external guidance. This paper addresses the gap between rigid unsupervised heuristics and the expensive requirement of fully labeled datasets by using Active Learning, where a human expert labels a tiny fraction of the data to guide the model.
Methods: The UAI Layer
The core contribution is the UAI Layer. Instead of building a new architecture from scratch, the authors propose a layer that can be attached to the top of existing deep unsupervised models (like Autoencoders).
- Input Features: The UAI layer takes two inputs from the base unsupervised model: the Latent Representation ($l(x)$) (compressed features) and the Unsupervised Anomaly Score ($s(x)$) (the model's original reconstruction error or guess).
- The Classifier: These inputs are fed into a simple logistic regression classifier (the UAI layer) that predicts the probability of a point being anomalous.
- Active Feedback Loop: The system uses a "most-likely positive" strategy, ranking unlabeled instances by their UAI score and sending the top $k$ instances to a human expert for labeling.
- End-to-End Training: Once labels are received, the entire network is retrained. Crucially, gradients flow back through the latent representation, allowing the model to reshape its internal understanding of the data based on expert feedback.
Experimental Evaluation
The researchers evaluated their approach on a mix of synthetic and real-world benchmarks:
- Synthetic MNIST & Fashion-MNIST: Created to simulate different anomaly types, such as rare digits (low-density) or misclassified points (clustered).
- Real-World Benchmarks: 11 diverse datasets including KDDCUP (network intrusion), Covtype (forest cover), Thyroid, and Arrhythmia.
Key Results and Insights
- Superior Performance: UAInets consistently outperformed unsupervised, semi-supervised, and other active learning competitors (like LODA-AAD and Tree-AAD).
- Efficiency: In the MNIST$_0$ experiment, with a budget of only 800 labeled instances (out of ~50,000), UAInets found almost all 600 anomalies, while the unsupervised DAE found only 30.
- Scalability: On the large Covtype dataset (286,048 instances, <1% anomalies), the active models showed the largest performance gain over competitors, proving that even a tiny amount of feedback is transformative for massive, imbalanced datasets.
- Robustness via Feedback: Unsupervised models like Autoencoders (DAE) are great at finding low-density anomalies but struggle with clustered ones. UAInets are robust to both because expert feedback tells the model which features actually matter for the specific task at hand.
- No Tuning Required: The UAI layer achieved these results without per-dataset hyperparameter tuning, making it highly practical for real-world deployment.
Conclusion
The UAI layer provides a practical path for deploying anomaly detection in the real world. By acknowledging the inherent ambiguity of anomalies and providing a mechanism for human-in-the-loop refinement, it transforms brittle unsupervised models into robust, high-performance systems. Future work will explore more complex classifiers for the UAI layer and methods to handle unreliable expert feedback.
Summary updated