abb1451d02700cfb573ef4093b2c2eaa595ec727,art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_pytorch.py,ProjectedGradientDescentPyTorch,generate,#ProjectedGradientDescentPyTorch#Any#Any#,109

Before Change

        adv_x_best = None
        rate_best = None

        for _ in trange(max(1, self.num_random_init), desc="PGD - Random Initializations", disable=not self.verbose):
            adv_x = x.astype(ART_NUMPY_DTYPE)

            // Compute perturbation with batching

After Change

                batch_eps = self.eps
                batch_eps_step = self.eps_step

            for rand_init_num in range(max(1, self.num_random_init)):
                adversarial_batch = self._generate_batch(x=batch, targets=batch_labels, mask=mask_batch, eps=batch_eps, eps_step=batch_eps_step)
                if rand_init_num == 0:
                    // first iteration: use the adversarial examples as they are the only ones we have now
                    adv_x[batch_index_1:batch_index_2] = np.copy(adversarial_batch)
                else:
                    // return the successful adversarial examples
                    attack_success = compute_success_array(
                        self.estimator,
                        batch,
                        batch_labels,
                        adversarial_batch,
                        self.targeted,
                        batch_size=self.batch_size,
                    )
                    adv_x[batch_index_1:batch_index_2][attack_success] = adversarial_batch[attack_success]

        logger.info(
            "Success rate of attack: %.2f%%",
            100 * compute_success(self.estimator, x, y, adv_x, self.targeted, batch_size=self.batch_size),
        )

Instances