[Paper Review]Skip-GANomaly

• Skip-GANomaly: Skip Connected and Adversarially Trained Encoder-Decoder Anomaly Detection, 2019.
• Authors: Samet Akçay, Amir Atapour-Abarghouei, Toby P. Breckon
• Conference: 2019 International Joint Conference on Neural Networks (IJCNN)
• Github: Link
• Summary:
  • an U-net like encoder-decoder convolutional neural network with skip-connections and utilized adversarial training scheme
    • the role of skip connections in generator
      • more stable training
    • Adversarial Training
      • AAE(Adversarial Auto Encoder)
        • superior reconstruction
        • superior capability of controlling the latent space
  • loss functions (Adversarial + Contextual + Latent)
    • Adversarial loss: log(D(x) + log(1 - D(x_fake))
    • Contextual loss: ||x - x'||
    • Latent loss (discriminator feature loss): || f(x) - f(x')||
      • followed same structure as the discriminator of the DCGAN
  • Inference based on reconstruction error
    • contextual similarity: ||x - x'||
    • latent representation score: || f(x) - f(x')||
    • apply feature scaling to the Anomaly score within the probabilistic range of [0,1]
  • Limitations
    • The model generates even abnormal samples
      • Though the proposed model is able to detect abnormality within latent object space

Architecture

Thoughts

• normal/abnormal을 나누기 쉬운 문제에만 적용 가능 할 것 같다.
  • 학습 때 사용하지 않은 abnormal sample도 생성
  • skip-connection을 reconstruction error based anomaly detection에 도입하는 건 한계가 존재한다.