def compute_normalized_ppca(self): """ Compute PPCA. """ nearest_neighbor_images = self.nearest_neighbor_images.reshape(self.nearest_neighbor_images.shape[0], -1) nearest_neighbor_images = nearest_neighbor_images[:self.args.n_fit] perturbations = self.perturbations.reshape(self.perturbations.shape[0], -1) test_images = self.test_images.reshape(self.test_images.shape[0], -1) pure_perturbations = perturbations - test_images nearest_neighbor_images_norms = numpy.linalg.norm(nearest_neighbor_images, ord=2, axis=1) perturbations_norms = numpy.linalg.norm(perturbations, ord=2, axis=1) test_images_norms = numpy.linalg.norm(test_images, ord=2, axis=1) pure_perturbations_norms = numpy.linalg.norm(pure_perturbations, ord=2, axis=1) success = numpy.logical_and(numpy.logical_and(self.success >= 0, self.accuracy), pure_perturbations_norms > 1e-4) log('[Detection] %d valid attacked samples' % numpy.sum(success)) perturbations_norms = perturbations_norms[success] test_images_norms = test_images_norms[success] pure_perturbations_norms = pure_perturbations_norms[success] perturbations = perturbations[success] test_images = test_images[success] pure_perturbations = pure_perturbations[success] nearest_neighbor_images /= numpy.repeat(nearest_neighbor_images_norms.reshape(-1, 1), nearest_neighbor_images.shape[1], axis=1) perturbations /= numpy.repeat(perturbations_norms.reshape(-1, 1), perturbations.shape[1], axis=1) test_images /= numpy.repeat(test_images_norms.reshape(-1, 1), test_images.shape[1], axis=1) pure_perturbations /= numpy.repeat(pure_perturbations_norms.reshape(-1, 1), pure_perturbations.shape[1], axis=1) assert not numpy.any(nearest_neighbor_images != nearest_neighbor_images) assert not numpy.any(perturbations != perturbations) assert not numpy.any(test_images != test_images) assert not numpy.any(pure_perturbations != pure_perturbations) ppca = PPCA(n_components=self.args.n_pca) ppca.fit(nearest_neighbor_images) log('[Experiment] computed PPCA on nearest neighbor images') reconstructed_test_images = ppca.inverse_transform(ppca.transform(test_images)) reconstructed_perturbations = ppca.inverse_transform(ppca.transform(perturbations)) reconstructed_pure_perturbations = ppca.inverse_transform(ppca.transform(pure_perturbations)) #self.probabilities['test'] = ppca.marginal(test_images) #self.probabilities['perturbation'] = ppca.marginal(perturbations) #self.probabilities['true'] = ppca.marginal(pure_perturbations) self.distances['test'] = numpy.average(numpy.multiply(reconstructed_test_images - test_images, reconstructed_test_images - test_images), axis=1) self.distances['perturbation'] = numpy.average(numpy.multiply(reconstructed_perturbations - perturbations, reconstructed_perturbations - perturbations), axis=1) self.distances['true'] = numpy.average(numpy.multiply(reconstructed_pure_perturbations - pure_perturbations, reconstructed_pure_perturbations - pure_perturbations), axis=1) self.angles['test'] = numpy.rad2deg(common.numpy.angles(test_images.T, reconstructed_test_images.T)) self.angles['perturbation'] = numpy.rad2deg(common.numpy.angles(reconstructed_perturbations.T, perturbations.T)) self.angles['true'] = numpy.rad2deg(common.numpy.angles(reconstructed_pure_perturbations.T, pure_perturbations.T))
def load_data(self): """ Load data and model. """ self.test_images = utils.read_hdf5(self.args.test_images_file).astype( numpy.float32) log('[Testing] read %s' % self.args.test_images_file) if len(self.test_images.shape) < 4: self.test_images = numpy.expand_dims(self.test_images, axis=3) self.test_codes = utils.read_hdf5(self.args.test_codes_file) self.test_codes = self.test_codes[:, self.args.label_index] log('[Testing] read %s' % self.args.test_codes_file) self.perturbations = utils.read_hdf5(self.args.perturbations_file) if len(self.perturbations.shape) > 3: self.perturbations = self.perturbations.reshape( (self.perturbations.shape[0], self.perturbations.shape[1], -1)) self.perturbation_images = self.test_images[:self.perturbations. shape[1]].reshape( self.perturbations. shape[1], -1) self.perturbation_codes = self.test_codes[:self.perturbations.shape[1]] log('[Testing] read %s' % self.args.perturbations_file) assert not numpy.any( self.perturbations != self.perturbations), 'NaN in perturbations' self.success = utils.read_hdf5(self.args.success_file) log('[Testing] read %s' % self.args.success_file) self.probabilities = utils.read_hdf5(self.args.probabilities_file) log('[Testing] read %s' % self.args.probabilities_file)
def load_data(self): """ Load data and model. """ self.test_images = utils.read_hdf5(self.args.test_images_file).astype(numpy.float32) log('[Testing] read %s' % self.args.test_images_file) # For handling both color and gray images. if len(self.test_images.shape) < 4: self.test_images = numpy.expand_dims(self.test_images, axis=3) log('[Testing] no color images, adjusted size') self.resolution = self.test_images.shape[2] log('[Testing] resolution %d' % self.resolution) self.train_images = utils.read_hdf5(self.args.train_images_file).astype(numpy.float32) # ! self.train_images = self.train_images.reshape((self.train_images.shape[0], -1)) log('[Testing] read %s' % self.args.train_images_file) self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(numpy.int) self.test_codes = self.test_codes[:, self.args.label_index] self.N_class = numpy.max(self.test_codes) + 1 log('[Testing] read %s' % self.args.test_codes_file) self.accuracy = utils.read_hdf5(self.args.accuracy_file) log('[Testing] read %s' % self.args.accuracy_file) self.perturbations = utils.read_hdf5(self.args.perturbations_file).astype(numpy.float32) self.N_attempts = self.perturbations.shape[0] # First, repeat relevant data. self.test_images = numpy.repeat(self.test_images[:self.perturbations.shape[1]], self.N_attempts, axis=0) self.perturbation_codes = numpy.repeat(self.test_codes[:self.perturbations.shape[1]], self.N_attempts, axis=0) self.perturbation_codes = numpy.squeeze(self.perturbation_codes) self.accuracy = numpy.repeat(self.accuracy[:self.perturbations.shape[1]], self.N_attempts, axis=0) # Then, reshape the perturbations! self.perturbations = numpy.swapaxes(self.perturbations, 0, 1) self.perturbations = self.perturbations.reshape((self.perturbations.shape[0] * self.perturbations.shape[1], -1)) assert self.perturbations.shape[1] == self.args.N_theta log('[Testing] read %s' % self.args.perturbations_file) assert not numpy.any(self.perturbations != self.perturbations), 'NaN in perturbations' self.success = utils.read_hdf5(self.args.success_file) self.success = numpy.swapaxes(self.success, 0, 1) self.success = self.success.reshape((self.success.shape[0] * self.success.shape[1])) log('[Testing] read %s' % self.args.success_file) log('[Testing] using %d input channels' % self.test_images.shape[3]) assert self.args.N_theta > 0 and self.args.N_theta <= 9 decoder = models.STNDecoder(self.args.N_theta) # decoder.eval() log('[Testing] set up STN decoder') self.model = decoder
def compute_statistics(self): """ Compute statistics based on distances. """ num_attempts = self.perturbations.shape[0] perturbations = numpy.swapaxes(self.perturbations, 0, 1) perturbations = perturbations.reshape( (perturbations.shape[0] * perturbations.shape[1], perturbations.shape[2])) success = numpy.swapaxes(self.success, 0, 1) success = success.reshape((success.shape[0] * success.shape[1])) probabilities = numpy.swapaxes(self.probabilities, 0, 1) probabilities = probabilities.reshape( (probabilities.shape[0] * probabilities.shape[1], -1)) confidences = numpy.max(probabilities, 1) perturbation_probabilities = self.test_probabilities[:self.success. shape[1]] perturbation_probabilities = numpy.repeat(perturbation_probabilities, num_attempts, axis=0) perturbation_confidences = numpy.max(perturbation_probabilities, 1) probability_ratios = confidences / perturbation_confidences raw_overall_success = success >= 0 log('[Testing] %d valid attacks' % numpy.sum(raw_overall_success)) # For off-manifold attacks this should not happen, but save is save. if not numpy.any(raw_overall_success): for type in [ 'raw_success', 'raw_iteration', 'raw_roc', 'raw_confidence_weighted_success', 'raw_confidence', 'raw_ratios' ]: self.results[type] = 0 if self.args.results_file: utils.write_pickle(self.args.results_file, self.results) log('[Testing] wrote %s' % self.args.results_file) log('[Testing] no successful attacks found, no plots') return # # We compute some simple statistics: # - raw success rate: fraction of successful attack without considering epsilon # - corrected success rate: fraction of successful attacks within epsilon-ball # - raw average perturbation: average distance to original samples (for successful attacks) # - corrected average perturbation: average distance to original samples for perturbations # within epsilon-ball (for successful attacks). # These statistics can also be computed per class. # And these statistics are computed with respect to three norms. if self.args.plot_directory and utils.display(): iterations = success[raw_overall_success] x = numpy.arange(numpy.max(iterations) + 1) y = numpy.bincount(iterations) plot_file = os.path.join(self.args.plot_directory, 'iterations') plot.bar(plot_file, x, y, title='Distribution of Iterations of Successful Attacks', xlabel='Number of Iterations', ylabel='Count') log('[Testing] wrote %s' % plot_file) plot_file = os.path.join(self.args.plot_directory, 'probabilities') plot.histogram(plot_file, confidences[raw_overall_success], 50) log('[Testing] wrote %s' % plot_file) plot_file = os.path.join(self.args.plot_directory, 'probability_ratios') plot.histogram(plot_file, probability_ratios, 50) log('[Testing] wrote %s' % plot_file) plot_file = os.path.join(self.args.plot_directory, 'test_probabilities') plot.histogram( plot_file, self.test_probabilities[ numpy.arange(self.test_probabilities.shape[0]), self.test_codes], 50) log('[Testing] wrote %s' % plot_file) y_true = numpy.concatenate( (numpy.zeros(confidences.shape[0]), numpy.ones(perturbation_confidences.shape[0]))) y_score = numpy.concatenate((confidences, perturbation_confidences)) roc_auc_score = sklearn.metrics.roc_auc_score(y_true, y_score) self.results['raw_roc'] = roc_auc_score self.results['raw_confidence_weighted_success'] = numpy.sum( confidences[raw_overall_success]) / numpy.sum( perturbation_confidences) self.results['raw_confidence'] = numpy.mean( probabilities[raw_overall_success]) self.results['raw_ratios'] = numpy.mean( probability_ratios[raw_overall_success]) self.results['raw_success'] = numpy.sum( raw_overall_success) / success.shape[0] self.results['raw_iteration'] = numpy.average( success[raw_overall_success]) if self.args.results_file: utils.write_pickle(self.args.results_file, self.results) log('[Testing] wrote %s' % self.args.results_file)
def compute_statistics(self): """ Compute statistics based on distances. """ # That's the basis for all computation as we only want to consider successful attacks # on test samples that were correctly classified. raw_overall_success = numpy.logical_and(self.success >= 0, self.accuracy) # Important check, for on-manifold attack this will happen if the manifold is small and the model very accurate! if not numpy.any(raw_overall_success): for n in range(len(self.norms)): for type in ['raw_success', 'raw_iteration', 'raw_average', 'raw_image']: self.results[n][type] = 0 for type in ['raw_class_success', 'raw_class_average', 'raw_class_image']: self.results[n][type] = numpy.zeros((self.N_class)) if self.args.results_file: utils.write_pickle(self.args.results_file, self.results) log('[Testing] wrote %s' % self.args.results_file) return # # Compute nearest neighbor statistics in image space. # if self.args.plot_directory and self.args.plot_manifolds and utils.display(): log('[Testing] computing nearest neighbor ...') nearest_neighbors_indices = self.compute_nearest_neighbors(self.perturbation_images[raw_overall_success]) pure_perturbations = self.test_images[raw_overall_success] - self.perturbation_images[raw_overall_success] pure_perturbations_norm = numpy.linalg.norm(pure_perturbations, ord=2, axis=1) for k in range(10): direction = self.perturbation_images[raw_overall_success] - self.train_images[nearest_neighbors_indices[:, k]] direction_norm = numpy.linalg.norm(direction, ord=2, axis=1) dot_products = numpy.einsum('ij,ij->i', direction, pure_perturbations) dot_product_norms = numpy.multiply(pure_perturbations_norm, direction_norm) dot_products, dot_product_norms = dot_products[dot_product_norms > 10**-8], dot_product_norms[dot_product_norms > 10**-8] dot_products /= dot_product_norms dot_products = numpy.degrees(numpy.arccos(dot_products)) # matplotlib's hsitogram plots give weird error if there are NaN values, so simple check: if dot_products.shape[0] > 0 and not numpy.any(dot_products != dot_products): plot_file = os.path.join(self.args.plot_directory, 'dot_products_nn%d' % k) plot.histogram(plot_file, dot_products, 100, xmin=numpy.min(dot_products), xmax=numpy.max(dot_products), title='Dot Products Between Adversarial Perturbations and Direction to Nearest Neighbor %d' % k, xlabel='Dot Product', ylabel='Count') log('[Testing] wrote %s' % plot_file) # # We compute some simple statistics: # - raw success rate: fraction of successful attack without considering epsilon # - corrected success rate: fraction of successful attacks within epsilon-ball # - raw average perturbation: average distance to original samples (for successful attacks) # - corrected average perturbation: average distance to original samples for perturbations # within epsilon-ball (for successful attacks). # These statistics can also be computed per class. # And these statistics are computed with respect to three norms. if self.args.plot_directory and utils.display(): iterations = self.success[raw_overall_success] x = numpy.arange(numpy.max(iterations) + 1) y = numpy.bincount(iterations) plot_file = os.path.join(self.args.plot_directory, 'iterations') plot.bar(plot_file, x, y, title='Distribution of Iterations of Successful Attacks', xlabel='Number of Iterations', ylabel='Count') log('[Testing] wrote %s' % plot_file) reference_perturbations = numpy.zeros(self.perturbations.shape) if self.args.N_theta > 4: reference_perturbations[:, 4] = 1 for n in range(len(self.norms)): norm = self.norms[n] delta = numpy.linalg.norm(self.perturbations - reference_perturbations, norm, axis=1) image_delta = numpy.linalg.norm(self.test_images - self.perturbation_images, norm, axis=1) if self.args.plot_directory and utils.display(): plot_file = os.path.join(self.args.plot_directory, 'distances_l%g' % norm) plot.histogram(plot_file, delta[raw_overall_success], 50, title='Distribution of $L_{%g}$ Distances of Successful Attacks' % norm, xlabel='Distance', ylabel='Count') log('[Testing] wrote %s' % plot_file) debug_accuracy = numpy.sum(self.accuracy) / self.accuracy.shape[0] debug_attack_fraction = numpy.sum(raw_overall_success) / numpy.sum(self.success >= 0) debug_test_fraction = numpy.sum(raw_overall_success) / numpy.sum(self.accuracy) log('[Testing] attacked mode accuracy: %g' % debug_accuracy) log('[Testing] only %g of successful attacks are valid' % debug_attack_fraction) log('[Testing] only %g of correct samples are successfully attacked' % debug_test_fraction) N_accuracy = numpy.sum(self.accuracy) self.results[n]['raw_success'] = numpy.sum(raw_overall_success) / N_accuracy self.results[n]['raw_iteration'] = numpy.average(self.success[raw_overall_success]) self.results[n]['raw_average'] = numpy.average(delta[raw_overall_success]) if numpy.any(raw_overall_success) else 0 self.results[n]['raw_image'] = numpy.average(image_delta[raw_overall_success]) if numpy.any(raw_overall_success) else 0 raw_class_success = numpy.zeros((self.N_class, self.perturbation_codes.shape[0]), bool) corrected_class_success = numpy.zeros((self.N_class, self.perturbation_codes.shape[0]), bool) self.results[n]['raw_class_success'] = numpy.zeros((self.N_class)) self.results[n]['raw_class_average'] = numpy.zeros((self.N_class)) self.results[n]['raw_class_image'] = numpy.zeros((self.N_class)) for c in range(self.N_class): N_samples = numpy.sum(self.accuracy[self.perturbation_codes == c].astype(int)) if N_samples <= 0: continue; raw_class_success[c] = numpy.logical_and(raw_overall_success, self.perturbation_codes == c) self.results[n]['raw_class_success'][c] = numpy.sum(raw_class_success[c]) / N_samples if numpy.any(raw_class_success[c]): self.results[n]['raw_class_average'][c] = numpy.average(delta[raw_class_success[c].astype(bool)]) if numpy.any(corrected_class_success[c]): self.results[n]['raw_class_image'][c] = numpy.average(image_delta[raw_class_success[c].astype(bool)]) if self.args.results_file: utils.write_pickle(self.args.results_file, self.results) log('[Testing] wrote %s' % self.args.results_file)
def attack(self): """ Test the model. """ assert self.model is not None assert self.model.classifier.training is False concatenate_axis = -1 if os.path.exists(self.args.perturbations_file) and os.path.exists(self.args.success_file): self.original_perturbations = utils.read_hdf5(self.args.perturbations_file) assert len(self.original_perturbations.shape) == 3, self.original_perturbations.shape log('[Attack] read %s' % self.args.perturbations_file) self.original_success = utils.read_hdf5(self.args.success_file) log('[Attack] read %s' % self.args.success_file) assert self.original_perturbations.shape[0] == self.original_success.shape[0] assert self.original_perturbations.shape[1] == self.original_success.shape[1] assert self.original_perturbations.shape[2] == self.test_theta.shape[1] if self.original_perturbations.shape[1] <= self.args.max_samples and self.original_perturbations.shape[0] <= self.args.max_attempts: log('[Attack] found %d attempts, %d samples, requested no more' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1])) return elif self.original_perturbations.shape[0] == self.args.max_attempts or self.original_perturbations.shape[1] == self.args.max_samples: if self.original_perturbations.shape[0] == self.args.max_attempts: self.test_theta = self.test_theta[self.original_perturbations.shape[1]:] self.test_fonts = self.test_fonts[self.original_perturbations.shape[1]:] self.test_classes = self.test_classes[self.original_perturbations.shape[1]:] self.args.max_samples = self.args.max_samples - self.original_perturbations.shape[1] concatenate_axis = 1 log('[Attack] found %d attempts with %d perturbations, computing %d more perturbations' % ( self.original_perturbations.shape[0], self.original_perturbations.shape[1], self.args.max_samples)) elif self.original_perturbations.shape[1] == self.args.max_samples: self.args.max_attempts = self.args.max_attempts - self.original_perturbations.shape[0] concatenate_axis = 0 log('[Attack] found %d attempts with %d perturbations, computing %d more attempts' % ( self.original_perturbations.shape[0], self.original_perturbations.shape[1], self.args.max_attempts)) self.perturbations = numpy.zeros((self.args.max_attempts, self.args.max_samples, self.test_theta.shape[1])) self.success = numpy.ones((self.args.max_attempts, self.args.max_samples), dtype=int) * -1 if self.args.attack.find('Batch') >= 0: batch_size = min(self.args.batch_size, self.args.max_samples) else: batch_size = 1 objective = self.objective_class() num_batches = int(math.ceil(self.args.max_samples/batch_size)) for i in range(num_batches): if i*batch_size == self.args.max_samples: break i_start = i * batch_size i_end = min((i + 1) * batch_size, self.args.max_samples) batch_fonts = self.test_fonts[i_start: i_end] batch_classes = self.test_classes[i_start: i_end] batch_code = numpy.concatenate((common.numpy.one_hot(batch_fonts, self.N_font), common.numpy.one_hot(batch_classes, self.N_class)), axis=1).astype(numpy.float32) batch_classes = common.torch.as_variable(batch_classes, self.args.use_gpu) batch_inputs = common.torch.as_variable(self.test_theta[i_start: i_end], self.args.use_gpu) batch_code = common.torch.as_variable(batch_code, self.args.use_gpu) t = 0 # This basically allows to only optimize over theta, keeping the font/class code fixed. self.model.decoder.set_code(batch_code) while True and t < self.args.max_attempts: attack = self.setup_attack(batch_inputs, batch_classes) success, perturbations, probabilities, norm, _ = attack.run(objective) assert not numpy.any(perturbations != perturbations), perturbations # Note that we save the perturbed image, not only the perturbation! perturbations = perturbations.reshape(batch_inputs.size()) # hack for when only one dimensional latent space is used! self.perturbations[t][i_start: i_end] = perturbations + batch_inputs.cpu().numpy() self.success[t][i_start: i_end] = success t += 1 log('[Attack] %d: completed' % i) if concatenate_axis >= 0: if self.perturbations.shape[0] == self.args.max_attempts: self.perturbations = numpy.concatenate((self.original_perturbations, self.perturbations), axis=concatenate_axis) self.success = numpy.concatenate((self.original_success, self.success), axis=concatenate_axis) log('[Attack] concatenated') utils.write_hdf5(self.args.perturbations_file, self.perturbations) log('[Attack] wrote %s' % self.args.perturbations_file) utils.write_hdf5(self.args.success_file, self.success) log('[Attack] wrote %s' % self.args.success_file)
def visualize_perturbations(self): """ Visualize perturbations. """ num_attempts = self.perturbations.shape[1] num_attempts = min(num_attempts, 6) utils.makedir(self.args.output_directory) count = 0 for i in range(min(1000, self.perturbations.shape[0])): log('[Visualization] sample %d, iterations %s and correctly classified: %s' % (i + 1, ' '.join(list(map( str, self.success[i]))), self.accuracy[i])) if not numpy.any(self.success[i] >= 0) or not self.accuracy[i]: continue elif count > 200: break #fig, axes = pyplot.subplots(num_attempts, 8) #if num_attempts == 1: # axes = [axes] # dirty hack for axis indexing for j in range(num_attempts): theta = self.test_theta[i] theta_attack = self.perturbations[i][j] theta_perturbation = theta_attack - theta image = self.test_images[i] image_attack = self.perturbation_images[i][j] image_perturbation = image_attack - image max_theta_perturbation = numpy.max( numpy.abs(theta_perturbation)) theta_perturbation /= max_theta_perturbation max_image_perturbation = numpy.max( numpy.abs(image_perturbation)) image_perturbation /= max_image_perturbation image_representation = self.theta_representations[i] attack_representation = self.perturbation_representations[i][j] image_label = numpy.argmax(image_representation) attack_label = numpy.argmax(attack_representation) #vmin = min(numpy.min(theta), numpy.min(theta_attack)) #vmax = max(numpy.max(theta), numpy.max(theta_attack)) #axes[j][0].imshow(theta.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax) #axes[j][1].imshow(numpy.squeeze(image), interpolation='nearest', cmap='gray', vmin=0, vmax=1) #axes[j][2].imshow(theta_perturbation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax) #axes[j][2].text(0, -1, 'x' + str(max_theta_perturbation)) #axes[j][3].imshow(numpy.squeeze(image_perturbation), interpolation='nearest', cmap='seismic', vmin=-1, vmax=1) #axes[j][3].text(0, -image.shape[1]//8, 'x' + str(max_image_perturbation)) #axes[j][4].imshow(theta_attack.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax) #axes[j][5].imshow(numpy.squeeze(image_attack), interpolation='nearest', cmap='gray', vmin=0, vmax=1) #axes[j][6].imshow(image_representation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax) #axes[j][6].text(0, -1, 'Label:' + str(image_label)) #axes[j][7].imshow(attack_representation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax) #axes[j][7].text(0, -1, 'Label:' + str(attack_label)) image_file = os.path.join( self.args.output_directory, '%d_%d_image_%d.png' % (i, j, image_label)) attack_file = os.path.join( self.args.output_directory, '%d_%d_attack_%d.png' % (i, j, attack_label)) perturbation_file = os.path.join( self.args.output_directory, '%d_%d_perturbation_%g.png' % (i, j, max_image_perturbation)) vis.image(image_file, image, scale=10) vis.image(attack_file, image_attack, scale=10) vis.perturbation(perturbation_file, image_perturbation, scale=10) #plot_file = os.path.join(self.args.output_directory, str(i) + '.png') #pyplot.savefig(plot_file) #pyplot.close(fig) count += 1
def attack(self): """ Test the model. """ assert self.model is not None assert self.model.training is False assert self.test_images.shape[0] == self.test_codes.shape[0], 'number of samples has to match' concatenate_axis = -1 if os.path.exists(self.args.perturbations_file) and os.path.exists(self.args.success_file): self.original_perturbations = utils.read_hdf5(self.args.perturbations_file) if self.test_images.shape[3] > 1: assert len(self.original_perturbations.shape) == 5 else: assert len(self.original_perturbations.shape) == 4 log('[Attack] read %s' % self.args.perturbations_file) self.original_success = utils.read_hdf5(self.args.success_file) log('[Attack] read %s' % self.args.success_file) assert self.original_perturbations.shape[0] == self.original_success.shape[0] assert self.original_perturbations.shape[1] == self.original_success.shape[1] assert self.original_perturbations.shape[2] == self.test_images.shape[1] assert self.original_perturbations.shape[3] == self.test_images.shape[2]# if self.original_perturbations.shape[1] >= self.args.max_samples and self.original_perturbations.shape[0] >= self.args.max_attempts: log('[Attack] found %d attempts, %d samples, requested no more' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1])) return elif self.original_perturbations.shape[0] == self.args.max_attempts or self.original_perturbations.shape[1] == self.args.max_samples: if self.original_perturbations.shape[0] == self.args.max_attempts: self.test_images = self.test_images[self.original_perturbations.shape[1]:] self.test_codes = self.test_codes[self.original_perturbations.shape[1]:] self.args.max_samples = self.args.max_samples - self.original_perturbations.shape[1] concatenate_axis = 1 log('[Attack] found %d attempts with %d perturbations, computing %d more perturbations' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1], self.args.max_samples)) elif self.original_perturbations.shape[1] == self.args.max_samples: self.args.max_attempts = self.args.max_attempts - self.original_perturbations.shape[0] concatenate_axis = 0 log('[Attack] found %d attempts with %d perturbations, computing %d more attempts' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1], self.args.max_attempts)) # can't squeeze here! if self.test_images.shape[3] > 1: self.perturbations = numpy.zeros((self.args.max_attempts, self.args.max_samples, self.test_images.shape[1], self.test_images.shape[2], self.test_images.shape[3])) else: self.perturbations = numpy.zeros((self.args.max_attempts, self.args.max_samples, self.test_images.shape[1], self.test_images.shape[2])) self.success = numpy.ones((self.args.max_attempts, self.args.max_samples), dtype=int) * -1 if self.args.attack.find('Batch') >= 0: batch_size = min(self.args.batch_size, self.args.max_samples) else: batch_size = 1 objective = self.objective_class() num_batches = int(math.ceil(self.args.max_samples/batch_size)) for i in range(num_batches): # self.test_images.shape[0] if i*batch_size == self.args.max_samples: break i_start = i*batch_size i_end = min((i+1)*batch_size, self.args.max_samples) batch_images = common.torch.as_variable(self.test_images[i_start: i_end], self.args.use_gpu) batch_classes = common.torch.as_variable(numpy.array(self.test_codes[i_start: i_end]), self.args.use_gpu) batch_images = batch_images.permute(0, 3, 1, 2) t = 0 while t < self.args.max_attempts: attack = self.setup_attack(batch_images, batch_classes) success, perturbations, probabilities, norm, _ = attack.run(objective) assert not numpy.any(perturbations != perturbations), perturbations # Note that we save the perturbed image, not only the perturbation! self.perturbations[t][i_start: i_end] = numpy.squeeze(numpy.transpose(perturbations + batch_images.cpu().numpy(), (0, 2, 3, 1))) self.success[t][i_start: i_end] = success # IMPORTANT: The adversarial examples are not considering whether the classifier is # actually correct to start with. t += 1 log('[Attack] %d: completed' % i) if concatenate_axis >= 0: if self.perturbations.shape[0] == self.args.max_attempts: self.perturbations = numpy.concatenate((self.original_perturbations, self.perturbations), axis=concatenate_axis) self.success = numpy.concatenate((self.original_success, self.success), axis=concatenate_axis) log('[Attack] concatenated') utils.write_hdf5(self.args.perturbations_file, self.perturbations) log('[Attack] wrote %s' % self.args.perturbations_file) utils.write_hdf5(self.args.success_file, self.success) log('[Attack] wrote %s' % self.args.success_file)
def visualize_perturbations(self): """ Visualize perturbations. """ num_attempts = self.perturbations.shape[1] num_attempts = min(num_attempts, 6) utils.makedir(self.args.output_directory) count = 0 for i in range(min(1000, self.perturbations.shape[0])): if not numpy.any(self.success[i]) or not self.accuracy[i]: continue elif count > 200: break #fig, axes = pyplot.subplots(num_attempts, 5) #if num_attempts == 1: # axes = [axes] # dirty hack for axis indexing for j in range(num_attempts): image = self.test_images[i] attack = self.perturbations[i][j] perturbation = attack - image max_perturbation = numpy.max(numpy.abs(perturbation)) perturbation /= max_perturbation image_representation = self.image_representations[i] attack_representation = self.perturbation_representations[i][j] image_label = numpy.argmax(image_representation) attack_label = numpy.argmax(attack_representation) #axes[j][0].imshow(numpy.squeeze(image), interpolation='nearest', cmap='gray', vmin=0, vmax=1) #axes[j][1].imshow(numpy.squeeze(perturbation), interpolation='nearest', cmap='seismic', vmin=-1, vmax=1) #axes[j][1].text(0, -image.shape[1]//8, 'x' + str(max_perturbation)) #axes[j][2].imshow(numpy.squeeze(attack), interpolation='nearest', cmap='gray', vmin=0, vmax=1) #vmin = min(numpy.min(image_representation), numpy.min(attack_representation)) #vmax = max(numpy.max(image_representation), numpy.max(attack_representation)) #axes[j][3].imshow(image_representation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax) #axes[j][3].text(0, -1, 'Label:' + str(image_label)) #axes[j][4].imshow(attack_representation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax) #axes[j][4].text(0, -1, 'Label:' + str(attack_label)) image_file = os.path.join( self.args.output_directory, '%d_%d_image_%d.png' % (i, j, image_label)) attack_file = os.path.join( self.args.output_directory, '%d_%d_attack_%d.png' % (i, j, attack_label)) perturbation_file = os.path.join( self.args.output_directory, '%d_%d_perturbation_%g.png' % (i, j, max_perturbation)) vis.image(image_file, image, scale=10) vis.image(attack_file, attack, scale=10) vis.perturbation(perturbation_file, perturbation, scale=10) if len(perturbation.shape) > 2: perturbation_magnitude = numpy.linalg.norm(perturbation, ord=2, axis=2) max_perturbation_magnitude = numpy.max( numpy.abs(perturbation_magnitude)) perturbation_magnitude /= max_perturbation_magnitude perturbation_file = os.path.join( self.args.output_directory, '%d_%d_perturbation_magnitude_%g.png' % (i, j, max_perturbation_magnitude)) vis.perturbation(perturbation_file, perturbation_magnitude, scale=10) #plot_file = os.path.join(self.args.output_directory, str(i) + '.png') #pyplot.savefig(plot_file) #pyplot.close(fig) count += 1
def attack(self): """ Test the model. """ assert self.model is not None assert self.model.training is False if self.args.attack.find('Batch') >= 0: batch_size = min(self.args.batch_size, self.args.max_samples) else: batch_size = 1 objective = self.objective_class() num_batches = int(math.ceil(self.args.max_samples / batch_size)) # can't squeeze here! if self.test_images.shape[3] > 1: self.perturbations = numpy.zeros( (self.args.max_attempts, self.args.max_samples, self.test_images.shape[1], self.test_images.shape[2], self.test_images.shape[3])) else: self.perturbations = numpy.zeros( (self.args.max_attempts, self.args.max_samples, self.test_images.shape[1], self.test_images.shape[2])) self.success = numpy.ones( (self.args.max_attempts, self.args.max_samples), dtype=int) * -1 self.probabilities = numpy.zeros( (self.args.max_attempts, self.args.max_samples, self.N_class)) for i in range(num_batches): # self.test_images.shape[0] if i * batch_size == self.args.max_samples: break i_start = i * batch_size i_end = min((i + 1) * batch_size, self.args.max_samples) batch_images = numpy.random.randint(0, 255, size=[batch_size] + self.test_images.shape[1:]) batch_images = common.torch.as_variable(batch_images, self.args.use_gpu) batch_images = batch_images.permute(0, 3, 1, 2) batch_classes = common.torch.as_variable( numpy.random.randint(0, self.N_class - 1, size=(batch_images.size(0))), self.args.use_gpu) t = 0 while t < self.args.max_attempts: attack = self.setup_attack(batch_images, batch_classes) success, perturbations, probabilities, norm, _ = attack.run( objective) assert not numpy.any( perturbations != perturbations), perturbations # Note that we save the perturbed image, not only the perturbation! self.perturbations[t][i_start:i_end] = numpy.squeeze( numpy.transpose(perturbations + batch_images.cpu().numpy(), (0, 2, 3, 1))) self.success[t][i_start:i_end] = success self.probabilities[t][i_start:i_end] = probabilities # IMPORTANT: The adversarial examples are not considering whether the classifier is # actually correct to start with. t += 1 log('[Attack] %d: completed' % i) utils.write_hdf5(self.args.perturbations_file, self.perturbations) log('[Attack] wrote %s' % self.args.perturbations_file) utils.write_hdf5(self.args.success_file, self.success) log('[Attack] wrote %s' % self.args.success_file) utils.write_hdf5(self.args.probabilities_file, self.probabilities) log('[Attack] wrote %s' % self.args.probabilities_file)
def load_data(self): """ Load data and model. """ self.test_images = utils.read_hdf5(self.args.test_images_file).astype(numpy.float32) log('[Testing] read %s' % self.args.test_images_file) # For handling both color and gray images. if len(self.test_images.shape) < 4: self.test_images = numpy.expand_dims(self.test_images, axis=3) log('[Testing] no color images, adjusted size') self.resolution = self.test_images.shape[2] log('[Testing] resolution %d' % self.resolution) self.train_images = utils.read_hdf5(self.args.train_images_file).astype(numpy.float32) # ! self.train_images = self.train_images.reshape((self.train_images.shape[0], -1)) log('[Testing] read %s' % self.args.train_images_file) self.test_theta = utils.read_hdf5(self.args.test_theta_file).astype(numpy.float32) log('[Testing] read %s' % self.args.test_theta_file) self.train_theta = utils.read_hdf5(self.args.train_theta_file).astype(numpy.float32) log('[Testing] read %s' % self.args.train_theta_file) self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(numpy.int) self.test_codes = self.test_codes[:, self.args.label_index] self.N_class = numpy.max(self.test_codes) + 1 log('[Testing] read %s' % self.args.test_codes_file) self.accuracy = utils.read_hdf5(self.args.accuracy_file) log('[Testing] read %s' % self.args.accuracy_file) self.perturbations = utils.read_hdf5(self.args.perturbations_file).astype(numpy.float32) self.N_attempts = self.perturbations.shape[0] assert not numpy.any(self.perturbations != self.perturbations), 'NaN in perturbations' # First, repeat relevant data. self.perturbation_theta = numpy.repeat(self.test_theta[:self.perturbations.shape[1]], self.N_attempts, axis=0) self.perturbation_codes = numpy.repeat(self.test_codes[:self.perturbations.shape[1]], self.N_attempts, axis=0) self.perturbation_codes = numpy.squeeze(self.perturbation_codes) self.accuracy = numpy.repeat(self.accuracy[:self.perturbations.shape[1]], self.N_attempts, axis=0) # Then, reshape the perturbations! self.perturbations = numpy.swapaxes(self.perturbations, 0, 1) self.perturbations = self.perturbations.reshape((self.perturbations.shape[0] * self.perturbations.shape[1], -1)) log('[Testing] read %s' % self.args.perturbations_file) self.success = utils.read_hdf5(self.args.success_file) self.success = numpy.swapaxes(self.success, 0, 1) self.success = self.success.reshape((self.success.shape[0] * self.success.shape[1])) log('[Testing] read %s' % self.args.success_file) assert self.args.decoder_files decoder_files = self.args.decoder_files.split(',') for decoder_file in decoder_files: assert os.path.exists(decoder_file), 'could not find %s' % decoder_file log('[Testing] using %d input channels' % self.test_images.shape[3]) decoder_units = list(map(int, self.args.decoder_units.split(','))) if len(decoder_files) > 1: log('[Testing] loading multiple decoders') decoders = [] for i in range(len(decoder_files)): decoder = models.LearnedDecoder(self.args.latent_space_size, resolution=(self.test_images.shape[3], self.test_images.shape[1], self.test_images.shape[2]), architecture=self.args.decoder_architecture, start_channels=self.args.decoder_channels, activation=self.args.decoder_activation, batch_normalization=not self.args.decoder_no_batch_normalization, units=decoder_units) state = State.load(decoder_files[i]) decoder.load_state_dict(state.model) if self.args.use_gpu and not cuda.is_cuda(decoder): decoder = decoder.cuda() decoders.append(decoder) decoder.eval() log('[Testing] loaded %s' % decoder_files[i]) self.model = models.SelectiveDecoder(decoders, resolution=(self.test_images.shape[3], self.test_images.shape[1], self.test_images.shape[2])) else: log('[Testing] loading one decoder') decoder = models.LearnedDecoder(self.args.latent_space_size, resolution=(self.test_images.shape[3], self.test_images.shape[1], self.test_images.shape[2]), architecture=self.args.decoder_architecture, start_channels=self.args.decoder_channels, activation=self.args.decoder_activation, batch_normalization=not self.args.decoder_no_batch_normalization, units=decoder_units) state = State.load(decoder_files[0]) decoder.load_state_dict(state.model) if self.args.use_gpu and not cuda.is_cuda(decoder): decoder = decoder.cuda() decoder.eval() log('[Testing] read decoder') self.model = decoder
def attack(self): """ Test the model. """ assert self.model is not None assert self.model.classifier.training is False concatenate_axis = -1 if os.path.exists(self.args.perturbations_file) and os.path.exists( self.args.success_file): self.original_perturbations = utils.read_hdf5( self.args.perturbations_file) assert len(self.original_perturbations.shape) == 3 log('[Attack] read %s' % self.args.perturbations_file) self.original_success = utils.read_hdf5(self.args.success_file) log('[Attack] read %s' % self.args.success_file) assert self.original_perturbations.shape[ 0] == self.original_success.shape[0] assert self.original_perturbations.shape[ 1] == self.original_success.shape[1] if self.original_perturbations.shape[ 1] <= self.args.max_samples and self.original_perturbations.shape[ 0] <= self.args.max_attempts: log('[Attack] found %d attempts, %d samples, requested no more' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1])) return elif self.original_perturbations.shape[ 0] == self.args.max_attempts or self.original_perturbations.shape[ 1] == self.args.max_samples: if self.original_perturbations.shape[ 0] == self.args.max_attempts: self.test_images = self.test_images[ self.original_perturbations.shape[1]:] self.test_codes = self.test_codes[ self.original_perturbations.shape[1]:] self.args.max_samples = self.args.max_samples - self.original_perturbations.shape[ 1] concatenate_axis = 1 log('[Attack] found %d attempts with %d perturbations, computing %d more perturbations' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1], self.args.max_samples)) elif self.original_perturbations.shape[ 1] == self.args.max_samples: self.args.max_attempts = self.args.max_attempts - self.original_perturbations.shape[ 0] concatenate_axis = 0 log('[Attack] found %d attempts with %d perturbations, computing %d more attempts' % (self.original_perturbations.shape[0], self.original_perturbations.shape[1], self.args.max_attempts)) self.perturbations = numpy.zeros( (self.args.max_attempts, self.args.max_samples, self.args.N_theta)) self.success = numpy.ones( (self.args.max_attempts, self.args.max_samples), dtype=int) * -1 if self.args.attack.find('Batch') >= 0: batch_size = min(self.args.batch_size, self.args.max_samples) else: batch_size = 1 objective = self.objective_class() num_batches = int(math.ceil(self.args.max_samples / batch_size)) for i in range(num_batches): if i * batch_size == self.args.max_samples: break i_start = i * batch_size i_end = min((i + 1) * batch_size, self.args.max_samples) batch_classes = common.torch.as_variable( self.test_codes[i_start:i_end], self.args.use_gpu) batch_theta = common.torch.as_variable( numpy.zeros((i_end - i_start, self.args.N_theta), dtype=numpy.float32), self.args.use_gpu) if self.args.N_theta > 4: batch_theta[:, 4] = 1 batch_images = common.torch.as_variable( self.test_images[i_start:i_end], self.args.use_gpu) batch_images = batch_images.permute(0, 3, 1, 2) self.model.decoder.set_image(batch_images) #output_images = self.model.decoder.forward(batch_theta) #error = torch.sum(torch.abs(output_images - batch_images)) #error = error.item() #print(error) #from matplotlib import pyplot #output_images = numpy.squeeze(numpy.transpose(output_images.cpu().detach().numpy(), (0, 2, 3, 1))) #pyplot.imshow(output_images[0]) #pyplot.show() t = 0 while True and t < self.args.max_attempts: attack = self.setup_attack(batch_theta, batch_classes) success, perturbations, probabilities, norm, _ = attack.run( objective) assert not numpy.any( perturbations != perturbations), perturbations # Note that we save the perturbed image, not only the perturbation! perturbations = perturbations.reshape(batch_theta.size( )) # hack for when only one dimensional latent space is used! self.perturbations[t][ i_start:i_end] = perturbations + batch_theta.cpu().detach( ).numpy() self.success[t][i_start:i_end] = success t += 1 log('[Attack] %d: completed' % i) if concatenate_axis >= 0: if self.perturbations.shape[0] == self.args.max_attempts: self.perturbations = numpy.concatenate( (self.original_perturbations, self.perturbations), axis=concatenate_axis) self.success = numpy.concatenate( (self.original_success, self.success), axis=concatenate_axis) log('[Attack] concatenated') utils.write_hdf5(self.args.perturbations_file, self.perturbations) log('[Attack] wrote %s' % self.args.perturbations_file) utils.write_hdf5(self.args.success_file, self.success) log('[Attack] wrote %s' % self.args.success_file)