VAE-SMOTE Augmented Diffusion for Anomaly Detection

Most current generative models for anomaly detection, which rely on reconstruction methods, primarily use either Variational Autoencoders or Generative Adversarial Networks. However, VAEs tend to suffer from poor reconstruction quality, while GANs face challenges related to training stability and limited pattern coverage. Additionally, due to privacy and security concerns, the available training data for anomaly detection is often limited, whereas generative models typically require large datasets to avoid the risk of overfitting. To overcome these challenges, we introduce an unsupervised anomaly detection model utilizing diffusion techniques, designed specifically for tabular data, using a reconstruction-based approach. We adopt the diffusion model as the backbone for anomaly detection, aiming to enhance reconstruction quality and ensure training stability. To further improve reconstruction quality and accelerate inference, we employ a truncated diffusion strategy during inference. What's more, to overcome the challenge of insufficient training data, we introduce a SMOTE-based data augmentation method optimized for feature space in tabular data, incorporating a VAE to address the limitations of the original SMOTE algorithm in capturing complex data structures and handling nonlinear relationships. Experimental results on real-world datasets validate that our method achieves higher detection accuracy compared to conventional methods, demonstrating its effectiveness.