GANomaly

GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training

1. 动机

   • Anomaly detection
     • highly biased towards one class (normal)
     • insufficient sample size of the other class (abnormal)
   • semi-supervised learning
     • detecting the unknown/unseen anomaly case
     • trained on normal samples
     • tested on normal and abnormal samples
   • encoder-decoder-encoder
     • minimizing the distance between the images
     • and the latent vectors
     • a larger distance metric

2. 论点

   • supervised approaches heavily depend on large, labeled datasets
   • Generative Adversarial Networks (GAN) have emerged as a leading methodology across both unsupervised and semi-supervised problems
   • reconstruction-based anomaly techniques
     • Overall prior work strongly supports the hypothesis that the use of autoencoders and GAN

3. 方法

   • GAN
     • unsupervised
     • to generate realistic images
     • compete
       • generator tries to generate an image, decoder- alike network, map input to latent space
       • discriminator decides whether the generated image is a real or a fake, classical classification architecture, reading an input image, and determining its validity

   • Adversarial Auto-Encoders (AAE)

     • encoder + decoder
     • reconstruction: maps the input to latent space and remaps back to input data space
     • train autoencoders with adversarial setting

   • inverse mapping

     • with the additional use of an encoder, a vanilla GAN network is capable of learning inverse mapping

   • model

     

     • learns both the normal data distribution and minimizes the output anomaly score
     • two encoder, one decoder, a discriminator
     • encoder
       • convolutional layers followed by batch-norm and leaky ReLU() activation
       • compress to a vector z
     • decoder
       • convolutional transpose layers, ReLU() activation and batch-norm
       • a tanh layer at the end
     • 2nd encoder
       • with the same architectural
       • but different parametrization
     • discriminator
       • DCGAN discriminator
     • Adversarial Loss
       • 不是基于GAN的traditional 0/1 ouput
       • 而是选了一个中间层，计算real／fake(reconstructed)的L2 distance
     • Contextual Loss
       • L1 yields less blurry results than L2
       • 计算输入图像和重建图像的L1 distance
     • Encoder Loss
       • an additional encoder loss to minimize the distance of the bottleneck features
       • 计算两个高维向量的L2 distance
       • 在测试的时候用它来scoring the abnormality