[Paper Review] Autoencoding Binary Classifiers for Supervised Anomaly Detection

Yuki Yamanaka, Tomoharu Iwata+3 more|arXiv (Cornell University)|Mar 26, 2019

Anomaly Detection Techniques and Applications22 references4 citations

TL;DR

This paper proposes Autoencoding Binary Classifiers (ABC), a novel supervised anomaly detection method that combines autoencoders with probabilistic binary classification to jointly detect known and unknown anomalies. By modeling label likelihood via autoencoder reconstruction error, ABC trains to minimize reconstruction error for normal data while maximizing it for known anomalies, achieving stable training and superior performance over existing supervised and unsupervised methods.

ABSTRACT

We propose the Autoencoding Binary Classifiers (ABC), a novel supervised anomaly detector based on the Autoencoder (AE). There are two main approaches in anomaly detection: supervised and unsupervised. The supervised approach accurately detects the known anomalies included in training data, but it cannot detect the unknown anomalies. Meanwhile, the unsupervised approach can detect both known and unknown anomalies that are located away from normal data points. However, it does not detect known anomalies as accurately as the supervised approach. Furthermore, even if we have labeled normal data points and anomalies, the unsupervised approach cannot utilize these labels. The ABC is a probabilistic binary classifier that effectively exploits the label information, where normal data points are modeled using the AE as a component. By maximizing the likelihood, the AE in the proposed ABC is trained to minimize the reconstruction error for normal data points, and to maximize it for known anomalies. Since our approach becomes able to reconstruct the normal data points accurately and fails to reconstruct the known and unknown anomalies, it can accurately discriminate both known and unknown anomalies from normal data points. Experimental results show that the ABC achieves higher detection performance than existing supervised and unsupervised methods.

Motivation & Objective

• Address the limitations of purely supervised anomaly detection, which fails on unknown anomalies and struggles with class imbalance.
• Overcome the instability of existing semi-supervised approaches like Limiting Reconstruction Capability (LRC), which maximizes reconstruction error for anomalies but suffers from unstable training.
• Integrate label information into autoencoder-based anomaly detection without sacrificing the ability to detect unknown anomalies.
• Develop a method that maintains high performance even when known anomalies are scarce or underrepresented in training data.
• Create a unified framework that behaves as an unsupervised autoencoder when no known anomalies are available, ensuring robustness in low-data regimes.

Proposed method

• Model the conditional probability of the label (normal/anomaly) given a data point as a Bernoulli distribution.
• Use the negative log-likelihood of this Bernoulli distribution as the loss function, which is approximated by the autoencoder's reconstruction error.
• Train the autoencoder to minimize reconstruction error for normal data points and maximize it for known anomalies, effectively learning a decision boundary that separates normal from anomalous data.
• Ensure training stability by bounding the objective function both above and below, unlike LRC which lacks an upper bound on reconstruction error.
• Use the resulting conditional probability as a more reliable anomaly score than raw reconstruction error.
• Enable the model to generalize to unknown anomalies by learning a latent representation that reconstructs normal data well but fails on both known and unknown anomalies.

Experimental results

Research questions

• RQ1 Can a supervised anomaly detection method be designed to detect both known and unknown anomalies effectively?
• RQ2 How can label information be integrated into autoencoder-based models without destabilizing training?
• RQ3 Does a hybrid approach combining supervised and unsupervised principles outperform purely supervised or unsupervised methods in anomaly detection?
• RQ4 How does the method perform when the number of known anomalies is small or class-imbalanced?
• RQ5 Can the model retain the properties of an autoencoder when no known anomalies are available during training?

Key findings

• ABC achieves state-of-the-art performance on multiple benchmark datasets, including MNIST, KDD’99, CIFAR-10, and 2D-Toy, with AUROC scores consistently outperforming existing supervised and unsupervised methods.
• On the 2D-Toy dataset, ABC achieves an AUROC of 0.997(001) for known anomalies and 0.988(019) for unknown anomalies under Setting 1, demonstrating strong generalization to unseen anomalies.
• When the number of known anomalies is reduced, ABC maintains high detection performance (AUROC ≈ 0.965 for known anomalies), while other supervised methods like RUSBoost and DNN degrade significantly.
• In Setting 2, ABC achieves an AUROC of 0.999(000) on MNIST for known anomalies and 0.823(016) for unknown anomalies, showing robustness even with limited anomaly data.
• The training loss for normal data points and anomalies in ABC remains stable, unlike LRC, which exhibits unstable optimization due to unbounded reconstruction error.
• ABC behaves as a standard autoencoder when no known anomalies are present, ensuring consistent performance in low-data regimes without requiring retraining or architectural changes.