Adversarial Out-domain Examples for Generative Models

Proof of concept and TensorFlow code for the paper: Adversarial Out-domain Examples for Generative Models.

Short abstract:

Test-time behavior of a generative model can be influenced by adversarial inputs. An attacker can force a pre-trained generator to reproduce an arbitrary output if fed by a suitable latent representation. We call this output a Out-Domain example.

Searching for out-domain examples (ODE) in the latent space of a ProGAN trained on CelebA:

Code:

All the code is available in the package ./ODE [tensorflow_1.8 is recommended]

Examples:

Some examples are available in ./Examples/*

• DCGAN_SVHN" ODE generation process for a pre-trained DCGAN generator trained on SVHN (with tensorboard logging).

---

How to cite our work:

@INPROCEEDINGS{8802456,
	author={D. {Pasquini} and M. {Mingione} and M. {Bernaschi}},
	booktitle={2019 IEEE European Symposium on Security and Privacy Workshops (EuroS PW)},
	title={Adversarial Out-domain Examples for Generative Models},
	year={2019},
	pages={272-280},
}