DeepOne-ClassClassification:SupplementaryMaterial1proceedings.mlr.press/v80/ruff18a/ruff18a-supp.pdf ·...
6
Deep One-Class Classification: Supplementary Material 1 Lukas Ruff Robert A. Vandermeulen Nico Görnitz Lucas Deecke Shoaib A. Siddiqui Alexander Binder Emmanuel Müller Marius Kloft 2018 1 Supplementary material of the paper Deep One-Class Classification published in Proceedings of the 35 th International Conference on Machine Learning, 2018. Copyright 2018 by the author(s).
Transcript of DeepOne-ClassClassification:SupplementaryMaterial1proceedings.mlr.press/v80/ruff18a/ruff18a-supp.pdf ·...
Deep One-Class Classification: Supplementary Material1
Lukas Ruff Robert A. Vandermeulen Nico Görnitz Lucas DeeckeShoaib A. Siddiqui Alexander Binder Emmanuel Müller
Marius Kloft
2018
1Supplementary material of the paper Deep One-Class Classification published in Proceedingsof the 35th International Conference on Machine Learning, 2018. Copyright 2018 by the author(s).
1 MNIST
Figure 1: Example from one experiment (seed) of the 32 most normal (left) and 32 mostanomalous (right) examples per class on MNIST according to Deep SVDD anomaly scoresordered by score from top left to bottom right.
1
Figure 2: Example from one experiment (seed) of the 32 most normal (left) and 32 mostanomalous (right) examples per class on MNIST according to Deep SVDD anomaly scoresordered by score from top left to bottom right.
2
2 CIFAR-10
Figure 3: Example from one experiment (seed) of the 32 most normal (left) and 32 mostanomalous (right) examples per class on CIFAR-10 according to Deep SVDD anomalyscores ordered by score from top left to bottom right.
3
Figure 4: Example from one experiment (seed) of the 32 most normal (left) and 32 mostanomalous (right) examples per class on CIFAR-10 according to Deep SVDD anomalyscores ordered by score from top left to bottom right.
4
3 Adversarial Attacks on GTSRB stop signs
Figure 5: The 20 adversarial examples generated by using Boundary Attack.
Figure 6: Example from one experiment (seed) of the 32 most normal (left) and 32 mostanomalous (right) examples on GTSRB stop signs training set according to Deep SVDDanomaly scores ordered by score from top left to bottom right.
Figure 7: Example from one experiment (seed) of the 32 most normal (left) and 32 mostanomalous (right) examples on GTSRB stop signs test set according to Deep SVDDanomaly scores ordered by score from top left to bottom right. The example shows thatDeep SVDD detects adversarial attacks.
5
