46ad98c1b619502c926c657c1a4c4276d2fd1f79,art/defences/adversarial_trainer_unittest.py,TestAdversarialTrainer,test_shared_model_mnist,#TestAdversarialTrainer#,110

Before Change

        // Create and fit classifier
        params = {"epochs": 5, "batch_size": BATCH_SIZE}
        classifier = CNN(im_shape, dataset="mnist")
        classifier.compile({"loss": "categorical_crossentropy", "optimizer": "adam", "metrics": ["accuracy"]})
        classifier.fit(x_train, y_train, **params)

        // Create FGSM attacker
        adv = FastGradientMethod(classifier, session)

After Change

        // Create and fit classifier
        params = {"nb_epochs": 2, "batch_size": BATCH_SIZE}
        classifier = self.classifier_k

        // Create FGSM attacker
        adv = FastGradientMethod(classifier)
        x_adv = adv.generate(x_test)
        preds = classifier.predict(x_adv)
        acc = np.sum(np.argmax(preds, axis=1) == np.argmax(y_test, axis=1)) / y_test.shape[0]

        // Perform adversarial training
        adv_trainer = AdversarialTrainer(classifier, adv)
        adv_trainer.fit(x_train, y_train, **params)

        // Evaluate that accuracy on adversarial sample has improved
        preds_adv_trained = adv_trainer.classifier.predict(x_adv)
        acc_adv_trained = np.sum(np.argmax(preds_adv_trained, axis=1) == np.argmax(y_test, axis=1)) / y_test.shape[0]
        print("\nAccuracy before adversarial training: %.2f%%" % (acc * 100))
        print("\nAccuracy after adversarial training: %.2f%%" % (acc_adv_trained * 100))

    @staticmethod
    def _cnn_mnist_tf(input_shape):

Instances