Slide 9
Slide 9 text
Proposal methods
DNN (m-class)
Layer1
Layer2
Layer3
LayerL
…
input class
prediction
Preprocessing
Classes: true/pred, Lreps→ dataset
Test Statistics
(class-conditional)
Normalization Transformations
(distribution-independent)
Layerwise Aggregation
(true and predicted classes)
Scoring Function
(Adversarial attack, OOD)
I
II
III
kNN-basedな統計量を利⽤
2パターンに分岐
1. 単⼀層or層のペアのp-value
2. 多変量p-value
をbaseとしたnormalization
1. JTLA, Fisher, multi
2. JTLA, LPE, multi