In this paper we show strategies to fake generated with the Generative Adversarial Network framework. One strategy is based on the statistical analysis and comparison of raw pixel values and features extracted from them. The other strategy learns formal specifications from the real and shows that fake violate the specifications of the real data. We show that fake produced with GANs have a universal signature that can be used to fake samples. We provide results on MNIST, CIFAR10, music and speech data.

Source link
thanks you RSS link


Please enter your comment!
Please enter your name here