深層学習技術を使用した異常検出
データの異常（または異常）を特定することは、科学技術のさまざまな分野の科学者やエンジニアにとっての課題です。異常（メインデータセットとは疑わしいオブジェクト）の検出は長い間取り扱われ、最初のアルゴリズムは前世紀の60年代に開発されましたが、この分野では、次のような分野で人々が直面する未解決の問題や問題が数多くあります。コンサルティング、銀行スコアリング、情報セキュリティ、金融取引、ヘルスケア。過去数年間の深層学習アルゴリズムの急速な発展に関連して、この問題を解決するための多くの最新のアプローチが、画像、CCTVカメラからの記録、表形式のデータ（金融取引に関する）など、さまざまなタイプの調査データに対して提案されています。
- Deep Anomaly Detection (DAD) - :
: . . - , ,
:
:
: , , , ( - )
:
precision / ( )
[2] G. Pang .
:
Deep learning for feature extraction - , ( ), . DAD.
.2 . φ(·) : X→ Z Z, .
Learning feature representation of normality - φ(·) : X→ Z , , Z .
End-to-end anomaly score learning - end-to-end , anomaly score.
, .
Deep learning for feature extraction
. , PCA (principal component analysis) [3] random projection [4], , . MLP, , NNs , RNNs ().
, anomaly score .
Learning feature representation of normality
.1 .
Generic Normality Feature Learning. . , .
ψ - , l - , ψ, φ ( ), f - .
:
DAD , , . [5]
- , , , .
φ_e (.) - , , φ_d (.) - , . s_x (data reconstruction error) .
Generative Adversarial Networks
GANs - , , ( G) , ( D) .
G D - .
DAD , . , , . AnoGAN [6].
Predictability Modeling. .
x̂_(t +1) = ψ (φ (x1 , x2 , · · · , xt ; Θ); W),
l_pred l_adv - , .
, , . . [7]
Self-supervised Classification. , ( - (n - 1) , - , ). . , .
Anomaly Measure-dependent Feature Learning.
φ(·) : X→ Z, .
l - .
:
Distance-based Measure. , : DB outliers [8], k-nearest neighbor distance [9] . - , .
One-class Classification-based Measure. , , , . one-class SVM [10], Support Vector Data Description (SVDD) [11].
Clustering-based Measure. , , [12].
End-to-end anomaly score learning
, anomaly score.
:
τ (x; Θ) : X→ R , .
Ranking Models. end-to-end . , . Self-trained deep ordinal regression model [13].
Prior-driven Models. - the Bayesian inverse RL-based sequential anomaly detection. - , . [14].
Softmax Models. , . , .
Deviation Networks (end-to-end pipeline) [1]
, G. Pang , . .5 .
function φ - anomaly scoring network, . Reference score generator - , ( ). ( φ(x; Θ) μ_R) deviation loss function L, anomaly scoring network , , .
deviation loss function
y = 1, , y = 0 . , anomaly score , φ(x; Θ), dev(x) , , "a" ( ). .
, , . SOTA-. end-to-end .
[1] Deep Anomaly Detection with Deviation Networks. G. Pang
[2] Deep Learning for Anomaly Detection: A Review. G. Pang
[3] Emmanuel J Candès, Xiaodong Li, Yi Ma, and John Wright. 2011. Robust principal component analysis?
[4] Ping Li, Trevor J Hastie, and Kenneth W Church. 2006. Very sparse random projections.
[5] Alireza Makhzani and Brendan Frey. 2014. K-sparse autoencoders. In ICLR.
[6] Thomas Schlegl, Philipp Seeböck, Sebastian M Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. 2017. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery.
[7] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. 2018. Future frame prediction for anomaly detection–a new baseline.
[8] Edwin M Knorr and Raymond T Ng. 1999. Finding intensional knowledge of distance-based outliers.[9] Fabrizio Angiulli and Clara Pizzuti. 2002. Fast outlier detection in high dimensional spaces.
[10] Bernhard Schölkopf, John C Platt, John Shawe-Taylor, Alex J Smola, and Robert C Williamson. 2001. Estimating the support of a high-dimensional distribution.
[11] David MJ Tax and Robert PW Duin. 2004. Support vector data description.
[12] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. 2018. Deep clustering for unsupervised learning of visual features.
[13] Guansong Pang, Cheng Yan, Chunhua Shen, Anton van den Hengel, and Xiao Bai. 2020. Self-trained Deep Ordinal Regression for End-to-End Video Anomaly Detection.
[14] Andrew YNgとStuartJRussell。2000.逆強化学習のためのアルゴリズム。