๐Paper Review
[paepr review]Deep One-Class Classification
date
Mar 8, 2023
slug
paper-review-deep-one-class-classification
author
status
Public
tags
paper
DeepLearning
summary
type
Post
thumbnail
category
๐Paper Review
updatedAt
Aug 26, 2024 03:53 PM
Abstract
์ด์ํ์ง๋ฅผ ์ํ ๋ฅ๋ฌ๋ ๋ฐฉ๋ฒ๋ก ๋ถ์กฑ
ย
Introduction
- supervised anomaly detection
- train set : normal / abnormal
- semi-supervised anomaly detection
- train set : unlabeled normal data (๋ค๋) |or| normal/abnomal (์๋)
- Unsupervised anomaly detection
- trainset : unlabeled data
- Deep SVDD (Support Vector Data Description)๋ feature extractor ๊ฐ anomaly detection task์ ๋ง์ถฐ ํ์ต๋๋ End - to - End Training์ด ๊ฐ๋ฅ
ย
Related Work
ย
- Kernel-based One-Class Classification
- OC-SVM : input space์ ๋ฐ์ดํฐ๋ฅผ kernel function์ ํตํด feature space๋ก ํฌ์ํ๊ณ , ๊ทธ ์์์ ์ ์๋ฐ์ดํฐ๋ฅผ ์ต๋ํ ์์ ์์ ๋ฉ์ด์ง๋๋กํ๋ ํ์ดํผํ๋์ธ์ ์ฐพ๊ณ , ํด๋น ๊ฒฝ๊ณ๋ฅผ ๊ธฐ๋ฐ์ผ๋ก ์ด์์น ํ์ง
- SVDD : input space์ ๋ฐ์ดํฐ๋ฅผ kernel function์ ํตํด feature space๋ก ํฌ์ํ๊ณ , ๊ทธ ์์์ ์ ์ ๋ฐ์ดํฐ๋ฅผ ๋๋ ค์ธ๋ ๊ฐ์ฅ ์์ ๊ตฌ (ํ์ดํผ์คํผ์ด)๋ฅผ ์ฐพ๊ณ , ํด๋น ๊ฒฝ๊ณ๋ฉด์ ๊ธฐ๋ฐ์ผ๋ก ์ด์์น๋ฅผ ํ์ง
ย
feature space์์ ๋๋ถ๋ถ์ ์ ์ ๋ฐ์ดํฐ( ํฌ์ฌ์ด i )๋ฅผ ๋๋ฌ์ธ๋ ์ค์ฌ์ด c์ธ ๊ฐ์ฅ ์์ ๊ตฌ ๊ฒฝ๊ณ(R์ ๊ณฑ)์ ์ฐพ๋ ๋ชฉ์ ํจ์
ย
ย
- Deep Approaches to Anomaly Detection : ์ด์์นํ์ง ๋ชฉ์ ํจ์๋ฅผ ๊ฐ๋ ๋ฅ๋ฌ๋ ๋ฐฉ๋ฒ๋ก ์ผ๋ก, ์๋์ ๊ฐ์ ๊ตฌ์กฐ๊ฐ ์ฌ์ฉ๋จ
- AE : ์ ๋ ฅ์ ์์ถ ํ ๋ณต์ํ๋ ์ ๊ฒฝ๋ง ์์ถ ๊ณผ์ ์์ ๋ณต์์ ํ์ํ ์ ๋ณด๋ฅผ ํฌํจํ๋๋ก ํ์ต
- VAE : ์ ์ฌ ์ฝ๋ ๊ฐ์ด ํ๊ท ๊ณผ ๋ถ์ฐ์ผ๋ก ํํ (๊ฐ์ฐ์์ ํ๋ฅ ๋ถํฌ์ ๊ธฐ๋ฐํ ํ๋ฅ ๊ฐ)
- AnoGAN : DCGAN ๊ณผ์ ๊ณผ ๋น์ทํ๋, ์ ์ ๋ฐ์ดํฐ๋ง์ ํ์ต์ํค๊ณ , ๋ฐ์ดํฐ๋ฅผ ์ ๋ ฅํ์ ๋, ํด๋น ๋ชจ๋ธ์ ์ถ๋ ฅ์ด ์ ๋ ฅ ์ด๋ฏธ์ง์ ๋ค๋ฅธ ์ ๋๋ฅผ Anomaly Score๋ก ๋์ถ, ์ ์ ํ Anomaly Score๋ฅผ threchold๋ก ์ทจํ์ฌ ์ด์์น ๋ฐ์ดํฐ ํ์ง
ย
Deep SVDD
- The Deep SVDD Objective
- term 1 : ๊ตฌ์ ๋ถํผ๋ฅผ ์ต์ํ
- term 2 : ๊ตฌ ์ธ๋ถ์ ๋ฐ์ดํฐ point์ ๋ํด penalty ๋ถ์ฌ
- term 3 : weight decay regularizer
๋๋ถ๋ถ์ ์ ์ ๋ฐ์ดํฐ๋ฅผ ์ต์ ๋ถํผ๋ฅผ ๊ฐ์ง ๊ตฌ ์์ผ๋ก mapping ํ๋ R๊ณผ W๋ฅผ ํ์ต
parameters = R : ๋ฐ์ง๋ฆ, W: weights
Hyperparameters = c : center, v : ์ด์์น ๋น์จ
๋๋ถ๋ถ์ ๋ฐ์ดํฐ๊ฐ normal์ด๋ผ๊ณ ๊ฐ์ ํ๋ one-class deep SVDD์์๋ ์๋์ ๊ฐ์ด ๊ตฌ์ ์ค์ฌ๊ณผ ๋ฐ์ดํฐ point ์ฌ์ด์ ๊ฑฐ๋ฆฌ๋ฅผ ์ต์ํํ๋ W๋ง์ ์ต์ ํ ํ๋๋ก ํ์ต
- Optimization of Deep SVDD
- optimiztion of W : R์ ๊ณ ์ ํ๊ณ k epoch ๋์ backpropagation์ ํตํด w๋ฅผ ์ ๋ฐ์ดํธ
- Optimization of R : latest update W๋ฅผ ๊ธฐ๋ฐ์ผ๋ก line search๋ฅผ ํตํด R์ ์ ๋ฐ์ดํธ
์์ objective function์ ์ต์ ํํ์ฌ ์ ์ ๋ฐ์ดํฐ์ ๊ณตํต์ ์ธ ์์๋ฅผ ์ถ์ถํ๊ณ , ๊ฐ ๋ฐ์ดํฐ point๋ฅผ ๊ตฌ์ ์ค์ฌ C์ ๊ฐ๊น๊ฒ mappingํ๋๋ก ํ์ต, W์ R์ ๊ต๋๋ก ์ต์ ํ
ย
- Properties of Deep SVDD
- Proposition 1
- C๋ ์ ์๋ฐ์ดํฐ์ ๋ํ feature๋ฅผ ํฌํจํ๊ณ ์์ด์ผ ํจ
- pre-trained ๋ CAE์ encoder๋ฅผ Deep SVDD์ initial weight๋ก ์ฌ์ฉ
- ํ์ต๋์ง ์์ ์ด๊ธฐ ๋ชจ๋ธ์ ๋ชจ๋ ๋ฐ์ดํฐ๋ฅผ ์ ๋ ฅํ์ฌ feature๋ฅผ ์ถ๋ ฅํ ํ feature๋ค์ ํ๊ท ์ ๊ตฌํ์ฌ C๋ก ์ฌ์ฉ
- Proposition 2
W, R, C๊ฐ ๋ชจ๋ 0์ด ๋๋๋ก ํ์ต๋๋ฉด, ์ด๋ค input๋ 0์ผ๋ก ์ถ๋ ฅ๋๋ ๋ฌธ์ ๊ฐ ์๊น
๋๋ฌธ์ C๋ฅผ representations of initial forward pass์ ํ๊ท ์ผ๋ก ๊ณ ์
ย
representations of initial forward pass์ ํ๊ท ์ ๊ตฌํ๋ ๊ณผ์
ย
bias term์ constant๋ก ์๋ ดํ๋ ์ํฉ์ ๋ฐฉ์งํ๊ธฐ ์ํด bias term์ ์ฌ์ฉํ์ง ์์
ย
Experiments
- Competing methods
- Shallow baselines : OC-SVM/ SVDD (kernel based) , KDE, IF (95% ์ด์ ๋ถ์ฐ ๋ณด์กด)
- Deep baselines : DCAE, AnoGAN
- One-class classification on MNIST and CIFAR-10
- Setup
- MNIST
- Train set : 6,000๊ฐ์ normal ์ด๋ฏธ์ง
- Test set : 10,000๊ฐ์ normal / abnormal ์ด๋ฏธ์ง
- CIFAR-10
- Train set : 5,000๊ฐ์ normal ์ด๋ฏธ์ง
- Test set : 10,000๊ฐ์ normal / abnormal ์ด๋ฏธ์ง
- Network architectures : LeNet ์ฌ์ฉ
- Results
MNIST์ CIFAR-10์์ one-class SVDD ๊ฐ ์ ๋ฐ์ ์ผ๋ก ์ข์ ์ฑ๋ฅ์ ๋์ถ
- Adversarial attacks on GTSRB stop signs (๋ ์ผ ๊ตํต ์ ํธ ์ธ์ ๋ฒค์น๋งํฌ ๋ฐ์ดํฐ)
- Setup
- Trainset : 780๊ฐ์ ์ ์ง ์ ํธ ํด๋์ค
- Testset : 270๊ฐ์ ์ ๊ท ์์ ์ 20๊ฐ์ ์ ๋์ ์์
- Network architectures : LeNet ์ฌ์ฉ
- Results
AnoGAN์ ๋ฐ์ดํฐ์
์ด ๋๋ฌด ์์ ์๋ ดํ์ง ๋ชปํ๊ณ , ONE CLASS DEEP SVDD ๊ฐ ๊ฐ์ฅ ์ข์ ์ฑ๋ฅ์ ๋์ถํ๋ค๊ณ ์ฃผ์ฅ
ย
ย