Adversarial Sample Detection with Gaussian Mixture Conditional Generative Adversarial Networks
Algorithm 3
In the unsupervised scenario, we used 10% of the test set as training data and 90% as evaluation data.
Input: the normal data into the pretrained network.
for each layer do
Input: the hidden layer l into semisupervised (Algorithm 1)
is obtained.
errori is obtained.
end for
In order to reduce the amount of calculation, the GM-CGAN's input features are the reconstruction error vector in the final layer of the last block group in the pre-trained target network, the latent vector in the final layer of the last block group in the pre-trained target network, and the logit vector of the target network. The GM-CGAN’s label information is the label of the output of this neural network model.
