Hi, you need to specify the training instances that will be used to create the perturbations in the "input_filenames" parameter (you can check the documentation for more details). You can include in this file your whole dataset, and the algorithm will optimize the perturbation in order to fool all the instances. Just make sure that you insert enough samples per class if you want to create Level-3 universal perturbations.
Hi, you need to specify the training instances that will be used to create the perturbations in the "input_filenames" parameter (you can check the documentation for more details). You can include in this file your whole dataset, and the algorithm will optimize the perturbation in order to fool all the instances. Just make sure that you insert enough samples per class if you want to create Level-3 universal perturbations.