이상 탐지를 위한 생성적 적대 신경망 부호화

Abstract

본 논문은 이상 탐지를 위한 기존의 방법들이 겪는 순환 일관성(cycle consistency) 문제에 대해서 설명하고, 이를 해결하기 위한 방법으로써 전방향 생성적 적대 생성망(forward GAN)이랑 후방향 생성적 적대 생성망(backward GAN)이 각각 참 데이터 분포와 잠재 변수의 분포를 학습하고 특성 맞춤 손실(feature matching loss)로 이 둘을 결합하는 방법을 제안했다.

We study a bad cycle-consistency problem that is harmful to the detection capability of previous approaches based on Generative Adversarial Networks (GANs). The previous methods suffers from this problem because of an inefficient evaluation or difficulty of adversarial training. To solve this problem, we present a Coupled GANs for Anomaly Detection (CGAD). Our model consists of a forward GAN and a backward GAN which learns a true data distribution and a latent prior respectively, and minimize discriminative feature matching loss to couple the forward and backward GAN. CGAD significantly outperforms previous state-of-the-art anomaly detection methods in area under precision recall curve (AUPR) on MNIST dataset, and in $F_1$ score on KDD99 dataset.