Reproduce “Skip-GANomaly: Skip Connected and Adversarially Trained Encoder-Decoder Anomaly Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Samet Akçay, Amir Atapour-Abarghouei, Toby P. Breckon
Skip-GANomaly introduces an unsupervised anomaly detection model that integrates skip connections within its generator and utilizes the discriminator for feature extraction, achieving superior performance on diverse visual datasets. The model notably improved AUC scores to 0.940 on the UBA X-ray security dataset and 0.903 on the FFOB dataset for threat item detection, advancing the state of the art in applications like X-ray screening.
Key insights
- Skip connections within the generator are crucial for achieving stable GAN training and enabling the capture of multi-scale contextual information, leading to high-quality image reconstructions.
- Utilizing the discriminator as a feature extractor for latent space comparison provides a robust mechanism for differentiating normal from abnormal samples, even when the generator can reconstruct anomalies.
- Combining image-space (contextual) and latent-space (feature-level) reconstruction errors into a unified anomaly score offers a comprehensive measure of deviation from normality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.