23d00b78af8b6fc554a033c30dc4b0f8ffaf4804,art/attacks/deepfool.py,DeepFool,generate,#DeepFool#Any#,34

Before Change

            fk_i_hat = fk_hat
            nb_iter = 0

            while fk_i_hat == fk_hat and nb_iter < self.max_iter:
                grad_diff = grd - grd[fk_hat]
                f_diff = f - f[fk_hat]

                // Masking true label
                mask = [0] * self.classifier.nb_classes
                mask[fk_hat] = 1
                norm = np.linalg.norm(grad_diff.reshape(self.classifier.nb_classes, -1), axis=1) + tol
                value = np.ma.array(np.abs(f_diff) / norm, mask=mask)

                l = value.argmin(fill_value=np.inf)
                r = (abs(f_diff[l]) / (pow(np.linalg.norm(grad_diff[l]), 2) + tol)) * grad_diff[l]

                // Add perturbation and clip result
                xj = np.clip(xj + r, clip_min, clip_max)

After Change

        for j, val in enumerate(x_adv):
            xj = val[None, ...]
            f = preds[j]
            grd = self.classifier.class_gradient(xj, logits=True)[0]
            fk_hat = np.argmax(f)

            for _ in range(self.max_iter):

Instances