Optimization-Franke_Wolfe

Dependencies

The following libraries have been used :

• cv2
• matplotlib_pyplot
• numpy
• pandas
• tensorflow

Instructions

In this project we have implemented four optimization methods for generating adversarial examples for both untargeted and targeted attacks ( FGSM, MI-FGSM, PGD, Franke-Wolfe).

The project should be executed using the following command on a terminal opened in the main directory of the project (inside the code folder):
python3 main.py [neural_network] [type_of_attack] [modality]
where :

• [neural_network] : is the neural network architecture on which is possible to perform an adversarial attack. Three types of convolutional neural network can be selected : "inception_v3", "resnet_v2", "mobilenet_v2".
• [type_of_attack] : is the type of attack it is possible to perform. If you want to make a single attack by using one of the implemented method you can just write "fgsm", "pgd", "mi-fgsm" or "fw-white". Otherwise you can use the option "grid_search" if you want to analyze how each method behaves for different parameters. In the end a plot is produced in order to compare the four methods.
• [modality] : represents the modality of attack. Two types of attack are allowed : "untargeted" and "targeted".