Adversarial-Noise-Szegedy2013

This repo creates adversarial images following Szegedy et al. 2013.

The main script, create_adv_image.py, reads images from the Original_Images folder, creates perturbed images, and stores them in the Perturbed_Images folder. The script also creates a summary plot comparing original and perturbed images and stores it in the Summary_Plots folder.

Basic Usage

Run the following commands to create a perturbed version of the ImageNet validation image validation_0_label_91.jpg.

• cd ./Code
• python create_adv_image.py

Advanced Usage: process a custom image

Run the following commands to create a perturbed version of a custom image contained in a custom folder

• cd ./Code
• python create_adv_image.py --img_name my_image.jpg --img_fld my_fld

Advanced Usage: process all images in the folder

Run the following commands to create perturbed versions of all the images in the Original_Images folder

• cd ./Code
• python create_adv_image.py --img_name all

Advanced Usage: customize device and number of epochs

Run the following commands to change the number of training epochs and to perform the optimization on CPU

• cd ./Code
• python create_adv_image.py --device cpu --n_epochs 1000