Disentangled Conditional Variational Autoencoder for Unsupervised Anomaly Detection

Generative models have recently become an effective approach for anomaly detection by leveraging auto-encoders to model high-dimensional data and identify anomalies based on reconstruction quality. However, a primary challenge in unsupervised anomaly detection (UAD) lies in learning meaningful, disentangled features without losing essential information. In this paper, we introduce a novel generative architecture that combines the frameworks of β-VAE, Conditional Variational Auto-encoder (CVAE), and the principle of total correlation (TC) to enhance feature disentanglement and retain critical information. Our approach improves the separation of latent features, optimizes TC loss more effectively, and enhances the detection of anomalies in complex, high-dimensional datasets such as image data. Through extensive qualitative and quantitative evaluations in benchmark datasets, we demonstrate that our method not only achieves strong performance in anomaly detection but also captures interpretable, disentangled representations, highlighting the importance of feature disentanglement in advancing UAD.