DASVDD: Deep Autoencoding Support Vector Data Descriptor for Anomaly Detection

TL;DR

DASVDD addresses one-class anomaly detection by jointly training a deep autoencoder with a Deep SVDD objective, minimizing the latent-space hypersphere volume while preserving reconstruction. The anomaly score combines reconstruction error and latent distance to a trainable hypersphere center, mitigating hypersphere collapse without constraining biases or fixing the center. A practical gamma-tuning heuristic and an alternating optimization scheme ensure stable training and robust performance across diverse data domains. Empirical results on seven benchmarks show DASVDD outperforms strong baselines and maintains robustness across anomaly classes, with applicability to general input types. This approach advances practical, domain-agnostic anomaly detection with a principled integration of representation learning and SVDD objectives.

Abstract

Semi-supervised anomaly detection aims to detect anomalies from normal samples using a model that is trained on normal data. With recent advancements in deep learning, researchers have designed efficient deep anomaly detection methods. Existing works commonly use neural networks to map the data into a more informative representation and then apply an anomaly detection algorithm. In this paper, we propose a method, DASVDD, that jointly learns the parameters of an autoencoder while minimizing the volume of an enclosing hyper-sphere on its latent representation. We propose an anomaly score which is a combination of autoencoder's reconstruction error and the distance from the center of the enclosing hypersphere in the latent representation. Minimizing this anomaly score aids us in learning the underlying distribution of the normal class during training. Including the reconstruction error in the anomaly score ensures that DASVDD does not suffer from the common hypersphere collapse issue since the DASVDD model does not converge to the trivial solution of mapping all inputs to a constant point in the latent representation. Experimental evaluations on several benchmark datasets show that the proposed method outperforms the commonly used state-of-the-art anomaly detection algorithms while maintaining robust performance across different anomaly classes.

Figures (8)

• Figure 1: Overview of the proposed method DASVDD.
• Figure 2: Samples with the lowest (top row) and highest (bottom row) anomaly scores in MNIST (10 left columns) and FMNIST (10 right columns).
• Figure 3: (a) Visualization of the DASVDD encoder's weights where biases vector are reshaped and concatenated together for better visualization. Training data is class 8 of CIFAR-10. (b) Hypersphere Center versus iterations.
• Figure 4: Total Loss, Reconstruction (AE) Loss, and Deep SVDD Loss for the class 8 of (a) MNIST, (b) Fashion MNIST, and (c) CIFAR-10 Datasets
• Figure 5: Receiver Operator Curve (ROC) for different values of $\gamma$ in class 8 of the CIFAR-10 dataset. The area under the curve (AUC) is an indicator of the performance of each model. The $\gamma$ value which we obtained using the approach described in Section \ref{['gamma']} is denoted by the thick blue curve.