[논문 리뷰] Adversarially Learned Anomaly Detection

• Model: Bidirectional GAN
• Reconstruction based Anomaly Detection
• Cycle consistencies in data-space and latent space

Introduction

• Previous GAN base anomaly detection methods
  • Reconstruction Error based anomaly detection
  • Requires optimization of latent code in test time
  • Slow
• Proposed method
  • Reconstruction Error (in feature space) based anomaly detection
  • Bi-directional GAN which includes encoder network
  • Use encoder network to directly inference latent code
  • Fast

Details

• Bi-Directional GAN
  • cycle-consistency, i.e. that G(E(x)) = x
  • Discriminator to regularize both in latent-space and latent-space
• Anomaly Score
  • L1 distance in feature space of Dxx
    • They assume feature loss is much more preferable over the output of the Dxx
      • Reason: Dxx is supposed to be unable to discriminate between the real and reconstructed sample
      • My Opinion
        • But somehow, the logit of Dxx shows reasonable anomaly detection performance.
        • Cannot certain that feature distance in Dxx is a superior method.

Thoughts

• Low AUROC score (around auroc score of 0.5)
• Validated on low-resolution images
• Poor reconstruction quality