[Paper Review] GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training

GANomaly

Semi-supervised anomaly detection via adversarial training (2018)

• Authors: Samet Akcay, Amir Atapour-Abarghouei , and Toby P. Breckon
• Conference: Asian Conference on Computer Vision (622--637)
• Organization: Springer
• Summary:
  • Semi-supervised anomaly detection
  • Architecture:
    • Conditional GAN
    • an adverasarial autoencoder within an Encoder-Decoder-Encoder pipeline
  • Anomalies are detected when the output of the model is greater than a certain threshold A(x) > φ.\
  • The official code is available: Link

---

Related Works

• AnoGAN
  • Hypothesis: The latent vector of the GAN represents the distribution of the data
  • Computationally expensive method: Iterative mapping to the vector space of the GAN
• BiGAN in an anomaly detection task
  • joint training to map from image space to latent space simultaneously, and vice-versa.
• Adversarial Auto-Encoders (AAE)
  • Training autoencoders with adversarial setting
  • Better reconstruction and Control over latent space

A: AnoGAN / B: Efficient-GAN-Anomaly / C: GANomaly

---

Architecture

---

Training

Hypothesis

• When an abnormal image is forward-passed into the network G, G_D is not able to reconstruct the abnormalities even though G_E manages to map the input X to the latent vector z.

Loss Functions

1. Adversarial Loss
   

   

• feature matching from Salimans et al
• reduce the instability of GAN training
• f is an intermediate layer of the discriminator D
• L2 distance between the feature representation of the original and generated images

2. Contextual Loss
   

   

• Use L1 loss (L1 was reported to yield less blurry results than L2 in the work of Isola et al)

3. Encoder Loss
   

   

• The distance between the bottleneck features of the input and the encoded features of the generated image
• For anomalous images, it will fail to minimize the distance between the input and the generated images in the feature space since both G and E networks are optimized towards normal samples only.

---

Model Testing

• Anomaly Score

• Feature Scaling to have the anomaly scores within the probabilistic range of [0,1]

---

Thoughts..

• Encoder-Decoder-Encoder 구조를 제안함.
  • encoder-decoder 구조에서 input x 와 reconstructed x'의 feature distance를 구하는 게 더 적절할 것 같다는 느낌
    • P(x)가 낮은 샘플 x를 찾고자 하는 이상 탐지 문제에서 latent space에서의 확률 분포 P(z)를 보는 것은 큰 도움이 되지 않는다.
    • 이는 모델을 학습할 때에 정상 데이터로만 학습되어 있기 때문에, 비정상 샘플 또한 정상 특징들로만 표현될 것이기 때문.
    • "unimodal normality case"일 경우에는 P(x)를 P(z)로부터 접근이 가능하다.
  • f-AnoGAN에서는 AdvAE 구조는 abnormal 영역을 생성한다고 보고함.
    • 구조가 적합한지에 대한 의문
    • 후속 연구인 skip-GANomaly에서는 skip-connection을 이용한 구조를 사용 (U-Net based)
• reconstructed image가 궁금하다.