📑Paper Review
[paper reivew] Deep Semi-Supervised Anomaly Detection
INTRODUCTION
본 논문에서 다루는 Deep SAD는 2주차에 리뷰한 Deep SVDD의 수식을 Semi-Supervised learning에 활용할 수 있도록 변형시킨 방법론으로, Relation work의 내용이 2주차 리뷰와 겹쳐 생략합니다.
Deep SVDD : One class classification 방법으로 normal data에 대한 분포를 모두 포함하는 원을 최소화 하는 방향으로 학습, 그 원의 중심과 관측치의 거리 측정을 통해 anomaly score를 구하는 방법
Deep SAD : unsupervised 방식인 OC-Deep SVDD를 semi-supervised 방식으로 확장시킨 방식
전제조건
- 다수의 normal로 이루어진 unlabeled data와 약간의 labeled data 존재
Loss function
x : unlabeled sample
x~ : labeled sample
y~ : normal 일 때 1, unnormal 일 때 -1
- term 1 : unlabeled data를 center에 가깝도록 학습
- term 2 : labeled normal data를 center에 가깝도록 학습
- term 3 : labeled unnormal data를 center에서 멀도록 학습
Experiments
데이터 : MNIST, Fasion-MNIST, CIFAR10 데이터셋
시나리오
- Adding labeled anomalies : 9개의 anomaly class 중 하나의 클래스를 골라서 anomal labeled data를 추가하며 실험
- Deep SAD는 CIFAR10 에서 좋은 성능을 보이는 반면, supervised 방식은 데이터 개수에 민감
- Pulluted training data : unlabeled data에 anomal data를 섞어가며 얼마나 robust 한 지 실험
- 다른 모델들이 pullution에 민감하게 반응하는 반면, Deep SAD는 robust 함
- Number of known anomaly classes : 9개의 anomaly class를 랜덤으로 anomal label에 포함
- labeled anomaly class 가 다양할 때도 좋은 성능을 보임