\u00d6rnek Tablo: Kar\u015f\u0131la\u015ft\u0131r\u0131lan Y\u00f6ntemler ve Temel \u00d6zellikler<\/a><\/li><\/ol><\/li><\/ol><\/nav><\/div>\n\n\n\nAnomali tespiti, siber g\u00fcvenlik, \u00fcretim, finansal doland\u0131r\u0131c\u0131l\u0131k ve t\u0131bbi g\u00f6r\u00fcnt\u00fc analizi gibi bir\u00e7ok alanda kritik \u00f6neme sahiptir. Y\u00fcksek boyutlu verilerdeki karma\u015f\u0131kl\u0131k artt\u0131k\u00e7a, etkili bir anomalilik tespiti yapmak zorla\u015f\u0131r. Geleneksel y\u00f6ntemler d\u00fc\u015f\u00fck boyutlu veya belirgin \u00f6zelliklere sahip veriler i\u00e7in iyi sonu\u00e7 verse de, karma\u015f\u0131k da\u011f\u0131l\u0131mlara sahip veriler \u00fczerinde yetersiz kal\u0131r.<\/p>\n\n\n\n
Derin \u00f6\u011frenme, son y\u0131llarda karma\u015f\u0131k veri da\u011f\u0131l\u0131mlar\u0131n\u0131 modellemede b\u00fcy\u00fck ba\u015far\u0131 elde etmi\u015ftir. \u00d6zellikle Generative Adversarial Networks (GAN), ger\u00e7ek verilerin istatistiksel da\u011f\u0131l\u0131m\u0131n\u0131 taklit edebilecek \u00fcretici (generator) ve ay\u0131rt edici (discriminator) a\u011flara sahiptir. Bu do\u011frultuda, makale, GAN\u2019lerin bu kapasitesini kullanarak daha g\u00fc\u00e7l\u00fc bir anomali tespiti yakla\u015f\u0131m\u0131 sunar.<\/p>\n\n\n\n
\u0130lgili \u00c7al\u0131\u015fmalar<\/h2>\n\n\n\n Klasik anomalilik tespiti y\u00f6ntemleri aras\u0131nda mesafe tabanl\u0131, tek-s\u0131n\u0131f s\u0131n\u0131fland\u0131rma (\u00f6rne\u011fin One-Class SVM) ve PCA gibi indirgeme tabanl\u0131 teknikler bulunur. Bunun yan\u0131 s\u0131ra, derin \u00f6\u011frenme tabanl\u0131 metodlar da son d\u00f6nemde yayg\u0131nla\u015fm\u0131\u015ft\u0131r. Otomatik kodlay\u0131c\u0131 (autoencoder) ve varyasyonel otomatik kodlay\u0131c\u0131lar gibi yakla\u015f\u0131mlar, veri uzay\u0131n\u0131 temsil eden d\u00fc\u015f\u00fck boyutlu bir latent uzay \u00f6\u011frenerek yeniden in\u015fa hatas\u0131 \u00fczerinden anomali belirlemeye \u00e7al\u0131\u015f\u0131r. Ayr\u0131ca, enerji tabanl\u0131 modeller (DSEBM) ve GMM ile birle\u015ftirilmi\u015f derin a\u011flar (DAGMM) da anomali tespitinde kullan\u0131lm\u0131\u015ft\u0131r.<\/p>\n\n\n\n
GAN tabanl\u0131 anomalilik tespitinde daha \u00f6nce \u201cAnoGAN\u201d isimli y\u00f6ntem sunulmu\u015ftur. Ancak AnoGAN, test a\u015famas\u0131nda her veri \u00f6rne\u011fi i\u00e7in ayr\u0131 bir optimizasyon yaparak latent de\u011fi\u015fkenleri geri kazan\u0131r, bu da b\u00fcy\u00fck veri setlerinde veya ger\u00e7ek zamanl\u0131 uygulamalarda pratik de\u011fildir. Bu nedenle makale, bu sorunu a\u015farak \u00e7ift y\u00f6nl\u00fc GAN modellerine dayanan, encoder\u2019\u0131 da e\u011fitim s\u00fcrecinde \u00f6\u011frenen bir yakla\u015f\u0131m (ALAD) \u00f6nerir. B\u00f6ylece test s\u0131ras\u0131nda tek ad\u0131ml\u0131k \u00f6ng\u00f6r\u00fc (feed-forward) ile latent temsil elde edilebilir.<\/p>\n\n\n\n
Arka Plan: GAN\u2019ler<\/h2>\n\n\n\n Standart GAN\u2019lerde bir \u00fcretici a\u011f (G) rastgele g\u00fcr\u00fclt\u00fc girdisinden ger\u00e7ek\u00e7i veri \u00f6rnekleri \u00fcretirken, bir ayr\u0131\u015ft\u0131r\u0131c\u0131 a\u011f (D) ger\u00e7ek veri ile \u00fcretilen veriyi ay\u0131rt etmeye \u00e7al\u0131\u015f\u0131r. E\u011fitim, G\u2019nin \u00fcretti\u011fi veriyi D\u2019nin ay\u0131rt edemeyece\u011fi \u015fekilde iyile\u015ftirmeye, D\u2019nin ise \u00fcretilmi\u015f veriyi ger\u00e7ek veriden ay\u0131rmaya \u00e7al\u0131\u015fmas\u0131yla bir min-max optimizasyon problemine d\u00f6n\u00fc\u015f\u00fcr.<\/p>\n\n\n\n
BiGAN veya AliGAN ad\u0131 verilen baz\u0131 GAN varyantlar\u0131, ayn\u0131 anda bir encoder (E) de \u00f6\u011frenerek veri uzay\u0131ndan latent uzaya bir projeksiyon sa\u011flar. Bu sayede, test zaman\u0131nda her veri \u00f6rne\u011fi i\u00e7in latent temsil bulmak i\u00e7in ek optimizasyona gerek kalmaz. Encoder, G ve Dxz aras\u0131ndaki rekabet, E\u2019nin ger\u00e7ek verilerin latent da\u011f\u0131l\u0131m\u0131n\u0131 yakalamas\u0131na, G\u2019nin ise bu latent uzaydan ger\u00e7ek\u00e7i \u00f6rnekler \u00fcretmesine yard\u0131mc\u0131 olur.<\/p>\n\n\n\nAnomali Tespiti<\/figcaption><\/figure>\n\n\n\nALAD: Y\u00f6ntemin Tan\u0131t\u0131m\u0131<\/h2>\n\n\n\n ALAD, \u00e7ift y\u00f6nl\u00fc GAN mimarisi \u00fczerine in\u015fa edilir. Temel fikir, hem veri-uzay\u0131nda hem de latent-uzayda \u00e7evrim tutarl\u0131l\u0131\u011f\u0131n\u0131 (cycle consistency) sa\u011flamakt\u0131r. Bu, ek ayr\u0131\u015ft\u0131r\u0131c\u0131lar (Dxx ve Dzz) ekleyerek ve spektral normalizasyon gibi stabilizasyon teknikleri kullanarak yap\u0131l\u0131r. B\u00f6ylece, E ve G\u2019nin birlikte \u00e7al\u0131\u015farak veriyi ve latent uzay\u0131 e\u015fle\u015ftirmesi geli\u015ftirilir.<\/p>\n\n\n\n
\u00c7evrim Tutarl\u0131l\u0131\u011f\u0131<\/h3>\n\n\n\n Cycle-consistency, her x veri \u00f6rne\u011fi i\u00e7in G(E(x)) \u2248 x ve her z latent \u00f6rne\u011fi i\u00e7in E(G(z)) \u2248 z ko\u015fullar\u0131n\u0131n sa\u011flanmas\u0131d\u0131r. Bu, encoder ve generator\u2019\u0131n birbirini desteklemesini sa\u011flar. Bu ama\u00e7la, ALAD, orijinal BiGAN\/AliGAN yap\u0131s\u0131na ek olarak Dxx ve Dzz ayr\u0131\u015ft\u0131r\u0131c\u0131lar\u0131n\u0131 kullan\u0131r. Dxx, (x, x) ger\u00e7ek \u00e7iftleri ve (x, G(E(x))) yeniden in\u015fa \u00e7iftleri aras\u0131nda ayr\u0131m yaparken, Dzz, (z, z) ger\u00e7ek latent \u00e7iftleri ile (z, E(G(z)))) \u00e7iftlerini ay\u0131rmay\u0131 \u00f6\u011frenir.<\/p>\n\n\n\n
Spektral Normalizasyon ve E\u011fitim Stabilizasyonu<\/h3>\n\n\n\n GAN e\u011fitimi genellikle karars\u0131z olabilir. Spektral normalizasyon (SN), GAN i\u00e7indeki a\u011f parametrelerinin Lipschitz sabitini kontrol ederek e\u011fitimi istikrarl\u0131 hale getirir. Bu \u00e7al\u0131\u015fma, spektral normalizasyonu ayr\u0131\u015ft\u0131r\u0131c\u0131lara ve gerekiyorsa encoder\u2019a da uygular, e\u011fitimin daha kararl\u0131 ger\u00e7ekle\u015fmesini sa\u011flar.<\/p>\n\n\n\n
Anomali Tespiti S\u00fcreci<\/h2>\n\n\n\n Anomali tespiti i\u00e7in ALAD, veri \u00f6rneklerini yeniden in\u015fa eder ve yeniden in\u015fa hatas\u0131n\u0131 kullan\u0131r. Ancak basit L1 veya L2 yeniden in\u015fa hatalar\u0131 her zaman etkili olmayabilir. Bu nedenle, ALAD geri d\u00f6n\u00fc\u015f ayr\u0131\u015ft\u0131r\u0131c\u0131s\u0131 Dxx\u2019in \u00f6zellik uzay\u0131nda bir hata metri\u011fi tan\u0131mlar:<\/p>\n\n\n\n
A(x) = ||fxx(x, x) – fxx(x, G(E(x)))||1<\/sub><\/p>\n\n\n\nBurada fxx, Dxx ayr\u0131\u015ft\u0131r\u0131c\u0131s\u0131n\u0131n (x, x’) ikili giri\u015fi i\u00e7in son logits katman\u0131ndan \u00f6nceki \u00f6zellikleri temsil eder. Bu \u00f6zellik uzay\u0131nda \u00f6l\u00e7\u00fclen mesafe, verinin manifolduna daha duyarl\u0131 bir anomali skorudur. E\u011fer bir \u00f6rnek iyi yeniden in\u015fa edilemiyorsa veya modelin al\u0131\u015f\u0131k olmad\u0131\u011f\u0131 yap\u0131sal farkl\u0131l\u0131klar sergiliyorsa, bu skor y\u00fcksek \u00e7\u0131kar ve \u00f6rnek anomali olarak s\u0131n\u0131fland\u0131r\u0131l\u0131r.<\/p>\n\n\n\n
Deneysel Sonu\u00e7lar<\/h2>\n\n\n\n ALAD, \u00e7e\u015fitli veri k\u00fcmeleri \u00fczerinde test edilmi\u015ftir:<\/p>\n\n\n\n
\nTablolu Veriler:<\/strong> KDD99 (a\u011f trafi\u011fi) ve Arrhythmia (t\u0131bbi veriler) setleri \u00fczerinde ALAD, DAGMM, DSEBM, IF (Isolation Forest) ve OC-SVM gibi y\u00f6ntemlerle kar\u015f\u0131la\u015ft\u0131r\u0131lm\u0131\u015ft\u0131r. KDD99 verisinde ALAD, derin \u00f6\u011frenme tabanl\u0131 en iyi y\u00f6ntemlere benzer veya daha iyi sonu\u00e7lar vermi\u015f, Arrhythmia setinde ise k\u00fc\u00e7\u00fck veri boyutu nedeniyle IF gibi daha basit y\u00f6ntemlerin biraz gerisinde kalabilmi\u015ftir. Yine de ALAD, derin \u00f6\u011frenme tabanl\u0131 y\u00f6ntemler aras\u0131nda olduk\u00e7a rekabet\u00e7idir.<\/li>\n\n\n\nG\u00f6r\u00fcnt\u00fc Verileri:<\/strong> SVHN ve CIFAR-10 veri setlerinde, her s\u0131n\u0131f\u0131 s\u0131rayla normal varsay\u0131p di\u011fer 9 s\u0131n\u0131f\u0131 anomali sayacak \u015fekilde toplam 10 senaryo kurgulanm\u0131\u015ft\u0131r. Sonu\u00e7lar, ALAD\u2019\u0131n \u00e7o\u011fu g\u00f6revde AnoGAN, DSEBM ve OC-SVM gibi y\u00f6ntemlerden daha iyi AUROC de\u011ferlerine ula\u015ft\u0131\u011f\u0131n\u0131 g\u00f6stermi\u015ftir. Baz\u0131 s\u0131n\u0131flarda ALAD, benzer veya biraz daha d\u00fc\u015f\u00fck performans sergilese de genel olarak istatistiksel \u00fcst\u00fcnl\u00fck sa\u011flam\u0131\u015ft\u0131r.<\/li>\n<\/ul><\/div>\n\n\n\nH\u0131z ve Verimlilik<\/h3>\n\n\n\n AnoGAN gibi \u00f6nceki GAN tabanl\u0131 y\u00f6ntemler, test a\u015famas\u0131nda her \u00f6rnek i\u00e7in ek optimizasyon gerektirirken, ALAD an\u0131nda latent temsil \u00fcretir. Test zaman\u0131nda ALAD, AnoGAN\u2019dan y\u00fczlerce kat daha h\u0131zl\u0131d\u0131r, bu da ger\u00e7ek zamanl\u0131 anomalilik tespiti a\u00e7\u0131s\u0131ndan b\u00fcy\u00fck bir avantajd\u0131r.<\/p>\n\n\n\n
Ek Deneyler ve Tart\u0131\u015fma<\/h2>\n\n\n\n \u00c7al\u0131\u015fmada, ALAD yap\u0131s\u0131n\u0131n bile\u015fenlerinin her birinin (spektral normalizasyon, latent ceza vs.) anomali tespiti performans\u0131na etkisi ara\u015ft\u0131r\u0131lm\u0131\u015ft\u0131r. Bu ablatif deneyler, spektral normalizasyon ve ek ayr\u0131\u015ft\u0131r\u0131c\u0131lar\u0131n eklenmesinin genellikle performans\u0131 iyile\u015ftirdi\u011fini g\u00f6stermi\u015ftir.<\/p>\n\n\n\n
Ayr\u0131ca, farkl\u0131 anomali skor y\u00f6ntemleri (L1, L2, logits \u00e7\u0131k\u0131\u015f\u0131 ve \u00f6zellik uzay\u0131) k\u0131yasland\u0131\u011f\u0131nda, \u00f6zellikle tablo verilerinde ve baz\u0131 durumlarda g\u00f6r\u00fcnt\u00fc verilerinde, \u00f6zellik tabanl\u0131 mesafenin (ALAD\u2019\u0131n kulland\u0131\u011f\u0131 y\u00f6ntem) genellikle daha g\u00fc\u00e7l\u00fc bir ay\u0131rt edicilik sa\u011flad\u0131\u011f\u0131 g\u00f6zlemlenmi\u015ftir.<\/p>\n\n\n\n
Bu makale, karma\u015f\u0131k ve y\u00fcksek boyutlu veri k\u00fcmelerinde anomali tespiti i\u00e7in GAN tabanl\u0131 bir yakla\u015f\u0131m olan ALAD\u2019\u0131 tan\u0131tmaktad\u0131r. ALAD, \u00e7ift y\u00f6nl\u00fc GAN mimarisi ile hem ger\u00e7ek veri da\u011f\u0131l\u0131m\u0131n\u0131 hem de latent uzay\u0131 modelleyerek, test a\u015famas\u0131nda h\u0131zl\u0131 ve etkili anomali tespiti yapar. Deneysel sonu\u00e7lar, ALAD\u2019\u0131n hem tablolama hem de g\u00f6r\u00fcnt\u00fc verilerinde bir\u00e7ok son teknoloji (state-of-the-art) y\u00f6nteme denk veya \u00fcst\u00fcn performans sergiledi\u011fini ve \u00f6zellikle b\u00fcy\u00fck veri setlerinde \u00f6nemli avantajlar sa\u011flad\u0131\u011f\u0131n\u0131 g\u00f6stermektedir. Gelecekte, konu\u015fma verisi veya sens\u00f6r verisi gibi di\u011fer veri modlar\u0131nda ALAD\u2019\u0131n uygulanmas\u0131 potansiyel ara\u015ft\u0131rma konular\u0131d\u0131r.<\/p>\n\n\n\n