Akcay S, Atapour-Abarghouei A, Breckon T P. GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training[J]. 2018.
1. Overview
In this paper, it proposes a novel anomaly detection model (output anomaly score)
- Conditional GAN
- jointly learn the generation of high-dimensional image space and the inference of latent space
- only trained on nromal samples
1.1. Problem Definition
- large training set. M normal samples
- smaller testing set. N samples (normal + abnormal)
- model f. learn the normal data distribution and minimizes the output anomaly score
- phi. threshold (set to 0.2)
2. Methods
2.1. Model
2.1.1. For Abnormal Image
- G_D is not able to reconstruct the abnormalities
- the network is modeled only on normal samples during training and its parametrization is not suitable for generating abnormal samples
- An output X^ that has missed abnormalities can lead to the encoder network R mapping X^ to a vector z^ that has also missed abnormal feature representation, causing dissimilarity between z and z^
2.2. Loss Function
3. Experiments
3.1. Dataset
- MNIST (32x32). treating one class being an anomaly, while other as normal class
- CIFAR10. one as abnormal class
- University Baggage Anomaly Dataset (UBA) (64x64). abnormal class: knife, gun and gun componen
3.2. Setting
- λ=50
- Metric. AUC, ROC and TPR (true positive rate), FPR
3.3. Comparison
- Table 1. all approaches show very poor performance for detecting digit 1 as abnormal. This is probably due to the linear shape simplicity of this class such that any model can easily overfit to the data
- Table 2. The reason for getting relatively lower quantitative results within this dataset is that for a selected abnormal category, there exists a normal class the is similar to the abnormal (plane vs bird, cat vs dog)
3.4. Speed
