Detekce anomálií pomocí generativních adversariálních sítí

Měkota, Ondřej

Generative adversarial networks (GANs) are able to capture distribution of its inputs. They are thus used to learn the distribution of normal data and then to detect anoma- lies, even if they are very rare; e.g. Schlegl et al. (2017) proposed an anomaly detection method called AnoGAN. However, a major disadvantage of GANs is instability during training. Therefore, Arjovsky et al. (2017) proposed a new version, called Wasserstein GAN (WGAN). The goal of this work is to propose a model, utilizing WGANs, to detect fraudulent credit card transactions. We develop a new method called AnoWGAN+e, partially based on AnoGAN, and compare it with One Class Support Vector Machines (OC-SVM) (Schöl- kopf et al. (2001)), k-Means ensemble (Porwal et al. (2018)) and other methods. Perfor- mance of studied methods is measured by area under precision-recall curve (AUPRC), and precision at different recall levels on credit card fraud dataset (Pozzolo (2015)). AnoW- GAN+e achieved the highest AUPRC and it is 12% better than the next best method OC-SVM. Furthermore, our model has 20% precision at 80% recall, compared to 8% precision of OC-SVM, and 89% precision at 10% recall as opposed to 79% of k-Means ensemble. 1

guest :: login Digital Repository
		Search		Submit		Help		About