Reproduce “Enhancing Anomaly Detection via Generating Diversified and Hard-to-distinguish Synthetic Anomalies” on OpenResearch

About this paper

Hyuntae Kim, Changhee Lee

Researchers at Chung-Ang University and Korea University developed the Diversified and Hard-to-distinguish Anomaly Generation (DHAG) framework to enhance unsupervised anomaly detection. This domain-agnostic method generates synthetic anomalies that are both diverse and subtly challenging for a discriminator to distinguish, achieving improved performance on various image and tabular datasets compared to existing benchmarks.

Key insights

• Forcing the discriminator to distinguish between subtly different samples (hard-to-distinguish anomalies) encourages it to learn fine-grained normality patterns, leading to more robust and generalizable decision boundaries.
• Utilizing multiple perturbators to generate diverse, orthogonal anomalies helps the discriminator learn to identify a wider spectrum of real-world anomalies, overcoming limitations of single-perturbator approaches.
• Applying perturbations in the latent space, rather than directly to input features, prevents the generation of superficial or easily detectable noise, promoting the learning of more meaningful semantic distinctions.

Summary by alphaXiv.