DLFuzz is an adversarial input generator presented in this paper : https://arxiv.org/pdf/1808.09413
Adversarial inputs are generated for a DNN composed of 5 fully connected layers and trained on the MNIST dataset.
This model achieved 95% accuracy on the test dataset. We only generate an adversarial input from a well classified input.
This implementation of DLFuzz generated 121 adversarial inputs out of 5000 images from the test set. Each adversarial input has an L2 distance with the original input bounded by 0.002
Left is adversarial input recognised as 9, right is original input recognized as 4
Left is adversarial input recognised as 7, right is original input recognized as 9
Next step would be to generate adversarials inputs for a Convolutionnal Neural Networks