others

class trojanvision.models.MagNet(name='magnet', dataset=None, model=_MagNet, structure=None, activation=None, v_noise=0.1, **kwargs)

MagNet proposed by Dongyu Meng from Shanghai Tech University in CCS 2017. It is an autoencoder for input images to defend against adversarial attacks.

Available model names:

{'magnet'}

See also

• paper: MagNet: a Two-Pronged Defense against Adversarial Examples

Parameters:

• structure (list[int | str]) –

  The MagNet model structure. Defaults to

  • 1-channel images: [3, 'average', 3] (e.g, MNIST)

  • 3-channel images: [32]

• activation (str) – The activation layer in MagNet model. Choose from ['sigmoid', 'relu']. Defaults to 'sigmoid' for 1-channel images (e.g, MNIST) and 'relu' for 3-channel images.

• v_noise (float) – The std of random Gaussian noise added to training data. Defaults to 0.1.