Beispiel #1
0
def __main__():
    model_structures = {
        "mnist_cnn": [mnist_cnn, mnist_testloader],
        "mnist": [mnist_resnet18, mnist_testloader],
        "cifar10_cnn": [cifar10_cnn, cifar10_testloader],
        "cifar10_wide_resnet34_10": [
            cifar10_wide_resnet34_10,
            cifar10_testloader,
        ],
    }

    parser = argparse.ArgumentParser(
        description="Evaluate models against attacked perturbations")

    parser.add_argument("arch", type=str)
    parser.add_argument("state_dict", type=str)
    parser.add_argument("--device", default="cuda", type=str)
    args = parser.parse_args()

    arch = args.arch
    state_path = args.state_dict
    device = args.device

    model, testloader = model_structures[arch]
    state = torch.load(state_path)
    model.load_state_dict(state)

    print(f"Testing {arch} with {state_path} on {device}\n")

    print(f"Unattacked {arch}")
    print(classification_report(model, testloader, device=device))

    print(f"FGSM attacked {arch}")
    print(classification_report_fgsm(model, testloader, device=device))

    print(f"PGD attacked {arch}")
    print(classification_report_pgd(model, testloader, device=device))
Beispiel #2
0
mnist_resnet18.load_state_dict(mnist_resnet18_state)

models = {
    "MNIST ResNet18": [mnist_resnet18, mnist_trainloader, mnist_testloader],
}

# %%
for model_name, (model, trainloader, testloader) in models.items():
    logging.info(f"Training {model_name}")
    new_model = fgsm_training(model,
                              trainloader,
                              device="cuda",
                              log=log,
                              n_epoches=40,
                              random=True)
    torch.save(
        model.state_dict(),
        os.path.join(SCRIPT_PATH, f"FGSM {model_name}.model"),
    )

    logging.info(f"Unattacked {model_name}")
    logging.info(classification_report(model, testloader, device="cuda"))

    logging.info(f"FGSM attacked {model_name}")
    logging.info(classification_report_fgsm(model, testloader, device="cuda"))

    logging.info(f"PGD attacked {model_name}")
    logging.info(classification_report_pgd(model, testloader, device="cuda"))

# %%
                model,
                trainloader,
                device=DEVICE,
                log=log,
                n_clusters=n_clusters,
                cluster_with=cluster_with,
                **global_param,
            )
            torch.save(
                new_model.state_dict(),
                os.path.join(
                    SCRIPT_PATH,
                    f"Cluster {model_name} {cluster_with} {n_clusters}.model",
                ),
            )

            logging.info(f"Unattacked {model_name}")
            logging.info(
                classification_report(new_model, testloader, device=DEVICE))

            logging.info(f"FGSM attacked {model_name}")
            logging.info(
                classification_report_fgsm(new_model,
                                           testloader,
                                           device=DEVICE))

            logging.info(f"PGD attacked {model_name}")
            logging.info(
                classification_report_pgd(new_model, testloader,
                                          device=DEVICE))