def __init__(self, sess, x, logits, targeted_label, binary_search_steps, max_iterations, initial_const, clip_min, clip_max, nb_classes, batch_size): self.sess = sess self.x = x self.logits = logits assert logits.op.type != 'Softmax' self.targeted_label = targeted_label self.binary_search_steps = binary_search_steps self.max_iterations = max_iterations self.initial_const = initial_const self.clip_min = clip_min self.clip_max = clip_max self.batch_size = batch_size self.repeat = self.binary_search_steps >= 10 self.shape = tuple([self.batch_size] + list(self.x.get_shape().as_list()[1:])) self.ori_img = tf.Variable(np.zeros(self.shape), dtype=tf_dtype, name='ori_img') self.const = tf.Variable(np.zeros(self.batch_size), dtype=tf_dtype, name='const') self.score = softmax_cross_entropy_with_logits( labels=self.targeted_label, logits=self.logits) self.l2dist = reduce_sum(tf.square(self.x - self.ori_img)) # small self.const will result small adversarial perturbation self.loss = reduce_sum(self.score * self.const) + self.l2dist self.grad, = tf.gradients(self.loss, self.x)
def margin_logit_loss(model_logits, label, nb_classes=10, num_classes=None): """Computes difference between logit for `label` and next highest logit. The loss is high when `label` is unlikely (targeted by default). This follows the same interface as `loss_fn` for TensorOptimizer and projected_optimization, i.e. it returns a batch of loss values. """ if num_classes is not None: warnings.warn("`num_classes` is depreciated. Switch to `nb_classes`." " `num_classes` may be removed on or after 2019-04-23.") nb_classes = num_classes del num_classes if 'int' in str(label.dtype): logit_mask = tf.one_hot(label, depth=nb_classes, axis=-1) else: logit_mask = label if 'int' in str(logit_mask.dtype): logit_mask = tf.to_float(logit_mask) try: label_logits = reduce_sum(logit_mask * model_logits, axis=-1) except TypeError: raise TypeError("Could not take row-wise dot product between " "logit mask, of dtype " + str(logit_mask.dtype) + " and model_logits, of dtype " + str(model_logits.dtype)) logits_with_target_label_neg_inf = model_logits - logit_mask * 99999 highest_nonlabel_logits = reduce_max( logits_with_target_label_neg_inf, axis=-1) loss = highest_nonlabel_logits - label_logits return loss
def spm(x, model, y=None, n_samples=None, dx_min=-0.1, dx_max=0.1, n_dxs=5, dy_min=-0.1, dy_max=0.1, n_dys=5, angle_min=-30, angle_max=30, n_angles=31, black_border_size=0): """ TensorFlow implementation of the Spatial Transformation Method. :return: a tensor for the adversarial example """ if y is None: preds = model.get_probs(x) # Using model predictions as ground truth to avoid label leaking preds_max = reduce_max(preds, 1, keepdims=True) y = tf.to_float(tf.equal(preds, preds_max)) y = tf.stop_gradient(y) del preds y = y / reduce_sum(y, 1, keepdims=True) # Define the range of transformations dxs = np.linspace(dx_min, dx_max, n_dxs) dys = np.linspace(dy_min, dy_max, n_dys) angles = np.linspace(angle_min, angle_max, n_angles) if n_samples is None: import itertools transforms = list(itertools.product(*[dxs, dys, angles])) else: sampled_dxs = np.random.choice(dxs, n_samples) sampled_dys = np.random.choice(dys, n_samples) sampled_angles = np.random.choice(angles, n_samples) transforms = zip(sampled_dxs, sampled_dys, sampled_angles) transformed_ims = parallel_apply_transformations( x, transforms, black_border_size) def _compute_xent(x): preds = model.get_logits(x) return tf.nn.softmax_cross_entropy_with_logits_v2( labels=y, logits=preds) all_xents = tf.map_fn( _compute_xent, transformed_ims, parallel_iterations=1) # Must be 1 to avoid keras race conditions # Return the adv_x with worst accuracy # all_xents is n_total_samples x batch_size (SB) all_xents = tf.stack(all_xents) # SB # We want the worst case sample, with the largest xent_loss worst_sample_idx = tf.argmax(all_xents, axis=0) # B batch_size = tf.shape(x)[0] keys = tf.stack([ tf.range(batch_size, dtype=tf.int32), tf.cast(worst_sample_idx, tf.int32) ], axis=1) transformed_ims_bshwc = tf.einsum('sbhwc->bshwc', transformed_ims) after_lookup = tf.gather_nd(transformed_ims_bshwc, keys) # BHWC return after_lookup
def kl_with_logits(p_logits, q_logits, scope=None, loss_collection=tf.GraphKeys.REGULARIZATION_LOSSES): """Helper function to compute kl-divergence KL(p || q) """ with tf.name_scope(scope, "kl_divergence") as name: p = tf.nn.softmax(p_logits) p_log = tf.nn.log_softmax(p_logits) q_log = tf.nn.log_softmax(q_logits) loss = reduce_mean(reduce_sum(p * (p_log - q_log), axis=1), name=name) tf.losses.add_loss(loss, loss_collection) return loss
def clip_eta(eta, ord, eps): """ Helper function to clip the perturbation to epsilon norm ball. :param eta: A tensor with the current perturbation. :param ord: Order of the norm (mimics Numpy). Possible values: np.inf, 1 or 2. :param eps: Epsilon, bound of the perturbation. """ # Clipping perturbation eta to self.ord norm ball if ord not in [np.inf, 1, 2]: raise ValueError('ord must be np.inf, 1, or 2.') reduc_ind = list(xrange(1, len(eta.get_shape()))) avoid_zero_div = 1e-12 if ord == np.inf: eta = clip_by_value(eta, -eps, eps) else: if ord == 1: raise NotImplementedError( "The expression below is not the correct way" " to project onto the L1 norm ball.") norm = tf.maximum( avoid_zero_div, reduce_sum(tf.abs(eta), reduc_ind, keepdims=True)) elif ord == 2: # avoid_zero_div must go inside sqrt to avoid a divide by zero # in the gradient through this operation norm = tf.sqrt( tf.maximum( avoid_zero_div, reduce_sum(tf.square(eta), reduc_ind, keepdims=True))) # We must *clip* to within the norm ball, not *normalize* onto the # surface of the ball factor = tf.minimum(1., div(eps, norm)) eta = eta * factor return eta
def l2_batch_normalize(x, epsilon=1e-12, scope=None): """ Helper function to normalize a batch of vectors. :param x: the input placeholder :param epsilon: stabilizes division :return: the batch of l2 normalized vector """ with tf.name_scope(scope, "l2_batch_normalize") as name_scope: x_shape = tf.shape(x) x = tf.contrib.layers.flatten(x) x /= (epsilon + reduce_max(tf.abs(x), 1, keepdims=True)) square_sum = reduce_sum(tf.square(x), 1, keepdims=True) x_inv_norm = tf.rsqrt(np.sqrt(epsilon) + square_sum) x_norm = tf.multiply(x, x_inv_norm) return tf.reshape(x_norm, x_shape, name_scope)
def optimize_linear(grad, eps, ord=np.inf): """ Solves for the optimal input to a linear function under a norm constraint. Optimal_perturbation = argmax_{eta, ||eta||_{ord} < eps} dot(eta, grad) :param grad: tf tensor containing a batch of gradients :param eps: float scalar specifying size of constraint region :param ord: int specifying order of norm :returns: tf tensor containing optimal perturbation """ # In Python 2, the `list` call in the following line is redundant / harmless. # In Python 3, the `list` call is needed to convert the iterator returned by `range` into a list. red_ind = list(range(1, len(grad.get_shape()))) avoid_zero_div = 1e-12 if ord == np.inf: # Take sign of gradient optimal_perturbation = tf.sign(grad) # The following line should not change the numerical results. # It applies only because `optimal_perturbation` is the output of # a `sign` op, which has zero derivative anyway. # It should not be applied for the other norms, where the # perturbation has a non-zero derivative. optimal_perturbation = tf.stop_gradient(optimal_perturbation) elif ord == 1: abs_grad = tf.abs(grad) sign = tf.sign(grad) max_abs_grad = tf.reduce_max(abs_grad, red_ind, keepdims=True) tied_for_max = tf.to_float(tf.equal(abs_grad, max_abs_grad)) num_ties = tf.reduce_sum(tied_for_max, red_ind, keepdims=True) optimal_perturbation = sign * tied_for_max / num_ties elif ord == 2: square = tf.maximum( avoid_zero_div, reduce_sum(tf.square(grad), reduction_indices=red_ind, keepdims=True)) optimal_perturbation = grad / tf.sqrt(square) else: raise NotImplementedError("Only L-inf, L1 and L2 norms are " "currently implemented.") # Scale perturbation to be the solution for the norm=eps rather than # norm=1 problem scaled_perturbation = utils_tf.mul(eps, optimal_perturbation) return scaled_perturbation
def attack_single_step(self, x, eta, g_feat): """ TensorFlow implementation of the Fast Feature Gradient. This is a single step attack similar to Fast Gradient Method that attacks an internal representation. :param x: the input placeholder :param eta: A tensor the same shape as x that holds the perturbation. :param g_feat: model's internal tensor for guide :return: a tensor for the adversarial example """ adv_x = x + eta a_feat = self.model.fprop(adv_x)[self.layer] # feat.shape = (batch, c) or (batch, w, h, c) axis = list(range(1, len(a_feat.shape))) # Compute loss # This is a targeted attack, hence the negative sign loss = -reduce_sum(tf.square(a_feat - g_feat), axis) # Define gradient of loss wrt input grad, = tf.gradients(loss, adv_x) # Multiply by constant epsilon scaled_signed_grad = self.eps_iter * tf.sign(grad) # Add perturbation to original example to obtain adversarial example adv_x = adv_x + scaled_signed_grad # If clipping is needed, # reset all values outside of [clip_min, clip_max] if (self.clip_min is not None) and (self.clip_max is not None): adv_x = tf.clip_by_value(adv_x, self.clip_min, self.clip_max) adv_x = tf.stop_gradient(adv_x) eta = adv_x - x eta = clip_eta(eta, self.ord, self.eps) return eta
def _compute_gradients(self, loss_fn, x, unused_optim_state): """Compute gradient estimates using SPSA.""" # Assumes `x` is a list, containing a [1, H, W, C] image # If static batch dimension is None, tf.reshape to batch size 1 # so that static shape can be inferred assert len(x) == 1 static_x_shape = x[0].get_shape().as_list() if static_x_shape[0] is None: x[0] = tf.reshape(x[0], [1] + static_x_shape[1:]) assert x[0].get_shape().as_list()[0] == 1 x = x[0] x_shape = x.get_shape().as_list() def body(i, grad_array): delta = self._delta delta_x = self._get_delta(x, delta) delta_x = tf.concat([delta_x, -delta_x], axis=0) loss_vals = tf.reshape( loss_fn(x + delta_x), [2 * self._num_samples] + [1] * (len(x_shape) - 1)) avg_grad = reduce_mean(loss_vals * delta_x, axis=0) / delta avg_grad = tf.expand_dims(avg_grad, axis=0) new_grad_array = grad_array.write(i, avg_grad) return i + 1, new_grad_array def cond(i, _): return i < self._num_iters _, all_grads = tf.while_loop( cond, body, loop_vars=[ 0, tf.TensorArray(size=self._num_iters, dtype=tf_dtype) ], back_prop=False, parallel_iterations=1) avg_grad = reduce_sum(all_grads.stack(), axis=0) return [avg_grad]
def body(x_in, y_in, domain_in, i_in, cond_in): # Create graph for model logits and predictions logits = model.get_logits(x_in) preds = tf.nn.softmax(logits) preds_onehot = tf.one_hot(tf.argmax(preds, axis=1), depth=nb_classes) # create the Jacobian graph list_derivatives = [] for class_ind in xrange(nb_classes): derivatives = tf.gradients(logits[:, class_ind], x_in) list_derivatives.append(derivatives[0]) grads = tf.reshape(tf.stack(list_derivatives), shape=[nb_classes, -1, nb_features]) # Compute the Jacobian components # To help with the computation later, reshape the target_class # and other_class to [nb_classes, -1, 1]. # The last dimention is added to allow broadcasting later. target_class = tf.reshape(tf.transpose(y_in, perm=[1, 0]), shape=[nb_classes, -1, 1]) other_classes = tf.cast(tf.not_equal(target_class, 1), tf_dtype) grads_target = reduce_sum(grads * target_class, axis=0) grads_other = reduce_sum(grads * other_classes, axis=0) # Remove the already-used input features from the search space # Subtract 2 times the maximum value from those value so that # they won't be picked later increase_coef = (4 * int(increase) - 2) \ * tf.cast(tf.equal(domain_in, 0), tf_dtype) target_tmp = grads_target target_tmp -= increase_coef \ * reduce_max(tf.abs(grads_target), axis=1, keepdims=True) target_sum = tf.reshape(target_tmp, shape=[-1, nb_features, 1]) \ + tf.reshape(target_tmp, shape=[-1, 1, nb_features]) other_tmp = grads_other other_tmp += increase_coef \ * reduce_max(tf.abs(grads_other), axis=1, keepdims=True) other_sum = tf.reshape(other_tmp, shape=[-1, nb_features, 1]) \ + tf.reshape(other_tmp, shape=[-1, 1, nb_features]) # Create a mask to only keep features that match conditions if increase: scores_mask = ((target_sum > 0) & (other_sum < 0)) else: scores_mask = ((target_sum < 0) & (other_sum > 0)) # Create a 2D numpy array of scores for each pair of candidate features scores = tf.cast(scores_mask, tf_dtype) \ * (-target_sum * other_sum) * zero_diagonal # Extract the best two pixels best = tf.argmax(tf.reshape(scores, shape=[-1, nb_features * nb_features]), axis=1) p1 = tf.mod(best, nb_features) p2 = tf.floordiv(best, nb_features) p1_one_hot = tf.one_hot(p1, depth=nb_features) p2_one_hot = tf.one_hot(p2, depth=nb_features) # Check if more modification is needed for each sample mod_not_done = tf.equal(reduce_sum(y_in * preds_onehot, axis=1), 0) cond = mod_not_done & (reduce_sum(domain_in, axis=1) >= 2) # Update the search domain cond_float = tf.reshape(tf.cast(cond, tf_dtype), shape=[-1, 1]) to_mod = (p1_one_hot + p2_one_hot) * cond_float domain_out = domain_in - to_mod # Apply the modification to the images to_mod_reshape = tf.reshape(to_mod, shape=([-1] + x_in.shape[1:].as_list())) if increase: x_out = tf.minimum(clip_max, x_in + to_mod_reshape * theta) else: x_out = tf.maximum(clip_min, x_in - to_mod_reshape * theta) # Increase the iterator, and check if all misclassifications are done i_out = tf.add(i_in, 1) cond_out = reduce_any(cond) return x_out, y_in, domain_out, i_out, cond_out
def __init__(self, sess, model, beta, decision_rule, batch_size, confidence, targeted, learning_rate, binary_search_steps, max_iterations, abort_early, initial_const, clip_min, clip_max, num_labels, shape): """ EAD Attack Return a tensor that constructs adversarial examples for the given input. Generate uses tf.py_func in order to operate over tensors. :param sess: a TF session. :param model: a cleverhans.model.Model object. :param beta: Trades off L2 distortion with L1 distortion: higher produces examples with lower L1 distortion, at the cost of higher L2 (and typically Linf) distortion :param decision_rule: EN or L1. Select final adversarial example from all successful examples based on the least elastic-net or L1 distortion criterion. :param batch_size: Number of attacks to run simultaneously. :param confidence: Confidence of adversarial examples: higher produces examples with larger l2 distortion, but more strongly classified as adversarial. :param targeted: boolean controlling the behavior of the adversarial examples produced. If set to False, they will be misclassified in any wrong class. If set to True, they will be misclassified in a chosen target class. :param learning_rate: The learning rate for the attack algorithm. Smaller values produce better results but are slower to converge. :param binary_search_steps: The number of times we perform binary search to find the optimal tradeoff- constant between norm of the perturbation and confidence of the classification. Set 'initial_const' to a large value and fix this param to 1 for speed. :param max_iterations: The maximum number of iterations. Setting this to a larger value will produce lower distortion results. Using only a few iterations requires a larger learning rate, and will produce larger distortion results. :param abort_early: If true, allows early abort when the total loss starts to increase (greatly speeds up attack, but hurts performance, particularly on ImageNet) :param initial_const: The initial tradeoff-constant to use to tune the relative importance of size of the perturbation and confidence of classification. If binary_search_steps is large, the initial constant is not important. A smaller value of this constant gives lower distortion results. For computational efficiency, fix binary_search_steps to 1 and set this param to a large value. :param clip_min: (optional float) Minimum input component value. :param clip_max: (optional float) Maximum input component value. :param num_labels: the number of classes in the model's output. :param shape: the shape of the model's input tensor. """ self.sess = sess self.TARGETED = targeted self.LEARNING_RATE = learning_rate self.MAX_ITERATIONS = max_iterations self.BINARY_SEARCH_STEPS = binary_search_steps self.ABORT_EARLY = abort_early self.CONFIDENCE = confidence self.initial_const = initial_const self.batch_size = batch_size self.clip_min = clip_min self.clip_max = clip_max self.model = model self.decision_rule = decision_rule self.beta = beta self.beta_t = tf.cast(self.beta, tf_dtype) self.repeat = binary_search_steps >= 10 self.shape = shape = tuple([batch_size] + list(shape)) # these are variables to be more efficient in sending data to tf self.timg = tf.Variable(np.zeros(shape), dtype=tf_dtype, name='timg') self.newimg = tf.Variable(np.zeros(shape), dtype=tf_dtype, name='newimg') self.slack = tf.Variable(np.zeros(shape), dtype=tf_dtype, name='slack') self.tlab = tf.Variable(np.zeros((batch_size, num_labels)), dtype=tf_dtype, name='tlab') self.const = tf.Variable(np.zeros(batch_size), dtype=tf_dtype, name='const') # and here's what we use to assign them self.assign_timg = tf.placeholder(tf_dtype, shape, name='assign_timg') self.assign_newimg = tf.placeholder(tf_dtype, shape, name='assign_newimg') self.assign_slack = tf.placeholder(tf_dtype, shape, name='assign_slack') self.assign_tlab = tf.placeholder(tf_dtype, (batch_size, num_labels), name='assign_tlab') self.assign_const = tf.placeholder(tf_dtype, [batch_size], name='assign_const') self.global_step = tf.Variable(0, trainable=False) self.global_step_t = tf.cast(self.global_step, tf_dtype) # Fast Iterative Shrinkage Thresholding self.zt = tf.divide(self.global_step_t, self.global_step_t + tf.cast(3, tf_dtype)) cond1 = tf.cast( tf.greater(tf.subtract(self.slack, self.timg), self.beta_t), tf_dtype) cond2 = tf.cast( tf.less_equal(tf.abs(tf.subtract(self.slack, self.timg)), self.beta_t), tf_dtype) cond3 = tf.cast( tf.less(tf.subtract(self.slack, self.timg), tf.negative(self.beta_t)), tf_dtype) upper = tf.minimum(tf.subtract(self.slack, self.beta_t), tf.cast(self.clip_max, tf_dtype)) lower = tf.maximum(tf.add(self.slack, self.beta_t), tf.cast(self.clip_min, tf_dtype)) self.assign_newimg = tf.multiply(cond1, upper) self.assign_newimg += tf.multiply(cond2, self.timg) self.assign_newimg += tf.multiply(cond3, lower) self.assign_slack = self.assign_newimg self.assign_slack += tf.multiply(self.zt, self.assign_newimg - self.newimg) # -------------------------------- self.setter = tf.assign(self.newimg, self.assign_newimg) self.setter_y = tf.assign(self.slack, self.assign_slack) # prediction BEFORE-SOFTMAX of the model self.output = model.get_logits(self.newimg) self.output_y = model.get_logits(self.slack) # distance to the input data self.l2dist = reduce_sum(tf.square(self.newimg - self.timg), list(range(1, len(shape)))) self.l2dist_y = reduce_sum(tf.square(self.slack - self.timg), list(range(1, len(shape)))) self.l1dist = reduce_sum(tf.abs(self.newimg - self.timg), list(range(1, len(shape)))) self.l1dist_y = reduce_sum(tf.abs(self.slack - self.timg), list(range(1, len(shape)))) self.elasticdist = self.l2dist + tf.multiply(self.l1dist, self.beta_t) self.elasticdist_y = self.l2dist_y + tf.multiply( self.l1dist_y, self.beta_t) if self.decision_rule == 'EN': self.crit = self.elasticdist self.crit_p = 'Elastic' else: self.crit = self.l1dist self.crit_p = 'L1' # compute the probability of the label class versus the maximum other real = reduce_sum((self.tlab) * self.output, 1) real_y = reduce_sum((self.tlab) * self.output_y, 1) other = reduce_max((1 - self.tlab) * self.output - (self.tlab * 10000), 1) other_y = reduce_max( (1 - self.tlab) * self.output_y - (self.tlab * 10000), 1) if self.TARGETED: # if targeted, optimize for making the other class most likely loss1 = tf.maximum(ZERO(), other - real + self.CONFIDENCE) loss1_y = tf.maximum(ZERO(), other_y - real_y + self.CONFIDENCE) else: # if untargeted, optimize for making this class least likely. loss1 = tf.maximum(ZERO(), real - other + self.CONFIDENCE) loss1_y = tf.maximum(ZERO(), real_y - other_y + self.CONFIDENCE) # sum up the losses self.loss21 = reduce_sum(self.l1dist) self.loss21_y = reduce_sum(self.l1dist_y) self.loss2 = reduce_sum(self.l2dist) self.loss2_y = reduce_sum(self.l2dist_y) self.loss1 = reduce_sum(self.const * loss1) self.loss1_y = reduce_sum(self.const * loss1_y) self.loss_opt = self.loss1_y + self.loss2_y self.loss = self.loss1 + self.loss2 + tf.multiply( self.beta_t, self.loss21) self.learning_rate = tf.train.polynomial_decay(self.LEARNING_RATE, self.global_step, self.MAX_ITERATIONS, 0, power=0.5) # Setup the optimizer and keep track of variables we're creating start_vars = set(x.name for x in tf.global_variables()) optimizer = tf.train.GradientDescentOptimizer(self.learning_rate) self.train = optimizer.minimize(self.loss_opt, var_list=[self.slack], global_step=self.global_step) end_vars = tf.global_variables() new_vars = [x for x in end_vars if x.name not in start_vars] # these are the variables to initialize when we run self.setup = [] self.setup.append(self.timg.assign(self.assign_timg)) self.setup.append(self.tlab.assign(self.assign_tlab)) self.setup.append(self.const.assign(self.assign_const)) var_list = [self.global_step] + [self.slack] + [self.newimg] + new_vars self.init = tf.variables_initializer(var_list=var_list)
def __init__(self, sess, model, batch_size, confidence, targeted, learning_rate, binary_search_steps, max_iterations, abort_early, initial_const, clip_min, clip_max, num_labels, shape): """ Return a tensor that constructs adversarial examples for the given input. Generate uses tf.py_func in order to operate over tensors. :param sess: a TF session. :param model: a cleverhans.model.Model object. :param batch_size: Number of attacks to run simultaneously. :param confidence: Confidence of adversarial examples: higher produces examples with larger l2 distortion, but more strongly classified as adversarial. :param targeted: boolean controlling the behavior of the adversarial examples produced. If set to False, they will be misclassified in any wrong class. If set to True, they will be misclassified in a chosen target class. :param learning_rate: The learning rate for the attack algorithm. Smaller values produce better results but are slower to converge. :param binary_search_steps: The number of times we perform binary search to find the optimal tradeoff- constant between norm of the purturbation and confidence of the classification. :param max_iterations: The maximum number of iterations. Setting this to a larger value will produce lower distortion results. Using only a few iterations requires a larger learning rate, and will produce larger distortion results. :param abort_early: If true, allows early aborts if gradient descent is unable to make progress (i.e., gets stuck in a local minimum). :param initial_const: The initial tradeoff-constant to use to tune the relative importance of size of the pururbation and confidence of classification. If binary_search_steps is large, the initial constant is not important. A smaller value of this constant gives lower distortion results. :param clip_min: (optional float) Minimum input component value. :param clip_max: (optional float) Maximum input component value. :param num_labels: the number of classes in the model's output. :param shape: the shape of the model's input tensor. """ self.sess = sess self.TARGETED = targeted self.LEARNING_RATE = learning_rate self.MAX_ITERATIONS = max_iterations self.BINARY_SEARCH_STEPS = binary_search_steps self.ABORT_EARLY = abort_early self.CONFIDENCE = confidence self.initial_const = initial_const self.batch_size = batch_size self.clip_min = clip_min self.clip_max = clip_max self.model = model self.repeat = binary_search_steps >= 10 self.shape = shape = tuple([batch_size] + list(shape)) # the variable we're going to optimize over modifier = tf.Variable(np.zeros(shape, dtype=np_dtype)) # these are variables to be more efficient in sending data to tf self.timg = tf.Variable(np.zeros(shape), dtype=tf_dtype, name='timg') self.tlab = tf.Variable(np.zeros((batch_size, num_labels)), dtype=tf_dtype, name='tlab') self.const = tf.Variable(np.zeros(batch_size), dtype=tf_dtype, name='const') # and here's what we use to assign them self.assign_timg = tf.placeholder(tf_dtype, shape, name='assign_timg') self.assign_tlab = tf.placeholder(tf_dtype, (batch_size, num_labels), name='assign_tlab') self.assign_const = tf.placeholder(tf_dtype, [batch_size], name='assign_const') # the resulting instance, tanh'd to keep bounded from clip_min # to clip_max self.newimg = (tf.tanh(modifier + self.timg) + 1) / 2 self.newimg = self.newimg * (clip_max - clip_min) + clip_min # prediction BEFORE-SOFTMAX of the model self.output = model.get_logits(self.newimg) # distance to the input data self.other = (tf.tanh(self.timg) + 1) / \ 2 * (clip_max - clip_min) + clip_min self.l2dist = reduce_sum(tf.square(self.newimg - self.other), list(range(1, len(shape)))) # compute the probability of the label class versus the maximum other real = reduce_sum((self.tlab) * self.output, 1) other = reduce_max((1 - self.tlab) * self.output - self.tlab * 10000, 1) if self.TARGETED: # if targeted, optimize for making the other class most likely loss1 = tf.maximum(ZERO(), other - real + self.CONFIDENCE) else: # if untargeted, optimize for making this class least likely. loss1 = tf.maximum(ZERO(), real - other + self.CONFIDENCE) # sum up the losses self.loss2 = reduce_sum(self.l2dist) self.loss1 = reduce_sum(self.const * loss1) self.loss = self.loss1 + self.loss2 # Setup the adam optimizer and keep track of variables we're creating start_vars = set(x.name for x in tf.global_variables()) optimizer = tf.train.AdamOptimizer(self.LEARNING_RATE) self.train = optimizer.minimize(self.loss, var_list=[modifier]) end_vars = tf.global_variables() new_vars = [x for x in end_vars if x.name not in start_vars] # these are the variables to initialize when we run self.setup = [] self.setup.append(self.timg.assign(self.assign_timg)) self.setup.append(self.tlab.assign(self.assign_tlab)) self.setup.append(self.const.assign(self.assign_const)) self.init = tf.variables_initializer(var_list=[modifier] + new_vars)
def generate(self, x, **kwargs): """ Generate symbolic graph for adversarial examples and return. :param x: The model's symbolic inputs. :param kwargs: Keyword arguments. See `parse_params` for documentation. """ # Parse and save attack-specific parameters assert self.parse_params(**kwargs) asserts = [] # If a data range was specified, check that the input was in that range if self.clip_min is not None: asserts.append(utils_tf.assert_greater_equal(x, tf.cast(self.clip_min, x.dtype))) if self.clip_max is not None: asserts.append(utils_tf.assert_less_equal(x, tf.cast(self.clip_max, x.dtype))) # Initialize loop variables momentum = tf.zeros_like(x) adv_x = x # Fix labels to the first model predictions for loss computation y, _nb_classes = self.get_or_guess_labels(x, kwargs) y = y / reduce_sum(y, 1, keepdims=True) targeted = (self.y_target is not None) def cond(i, _, __): """Iterate until number of iterations completed""" return tf.less(i, self.nb_iter) def body(i, ax, m): """Do a momentum step""" logits = self.model.get_logits(ax) loss = softmax_cross_entropy_with_logits(labels=y, logits=logits) if targeted: loss = -loss # Define gradient of loss wrt input grad, = tf.gradients(loss, ax) # Normalize current gradient and add it to the accumulated gradient red_ind = list(range(1, len(grad.get_shape()))) avoid_zero_div = tf.cast(1e-12, grad.dtype) grad = grad / tf.maximum( avoid_zero_div, reduce_mean(tf.abs(grad), red_ind, keepdims=True)) m = self.decay_factor * m + grad optimal_perturbation = optimize_linear(m, self.eps_iter, self.ord) if self.ord == 1: raise NotImplementedError("This attack hasn't been tested for ord=1." "It's not clear that FGM makes a good inner " "loop step for iterative optimization since " "it updates just one coordinate at a time.") # Update and clip adversarial example in current iteration ax = ax + optimal_perturbation ax = x + utils_tf.clip_eta(ax - x, self.ord, self.eps) if self.clip_min is not None and self.clip_max is not None: ax = utils_tf.clip_by_value(ax, self.clip_min, self.clip_max) ax = tf.stop_gradient(ax) return i + 1, ax, m _, adv_x, _ = tf.while_loop( cond, body, (tf.zeros([]), adv_x, momentum), back_prop=True, maximum_iterations=self.nb_iter) if self.sanity_checks: with tf.control_dependencies(asserts): adv_x = tf.identity(adv_x) return adv_x
def fgm(x, logits, y=None, eps=0.3, ord=np.inf, clip_min=None, clip_max=None, targeted=False, sanity_checks=True): """ TensorFlow implementation of the Fast Gradient Method. :param x: the input placeholder :param logits: output of model.get_logits :param y: (optional) A placeholder for the true labels. If targeted is true, then provide the target label. Otherwise, only provide this parameter if you'd like to use true labels when crafting adversarial samples. Otherwise, model predictions are used as labels to avoid the "label leaking" effect (explained in this paper: https://arxiv.org/abs/1611.01236). Default is None. Labels should be one-hot-encoded. :param eps: the epsilon (input variation parameter) :param ord: (optional) Order of the norm (mimics NumPy). Possible values: np.inf, 1 or 2. :param clip_min: Minimum float value for adversarial example components :param clip_max: Maximum float value for adversarial example components :param targeted: Is the attack targeted or untargeted? Untargeted, the default, will try to make the label incorrect. Targeted will instead try to move in the direction of being more like y. :return: a tensor for the adversarial example """ asserts = [] # If a data range was specified, check that the input was in that range if clip_min is not None: asserts.append( utils_tf.assert_greater_equal(x, tf.cast(clip_min, x.dtype))) if clip_max is not None: asserts.append( utils_tf.assert_less_equal(x, tf.cast(clip_max, x.dtype))) # Make sure the caller has not passed probs by accident assert logits.op.type != 'Softmax' if y is None: # Using model predictions as ground truth to avoid label leaking preds_max = reduce_max(logits, 1, keepdims=True) y = tf.to_float(tf.equal(logits, preds_max)) y = tf.stop_gradient(y) y = y / reduce_sum(y, 1, keepdims=True) # Compute loss loss = softmax_cross_entropy_with_logits(labels=y, logits=logits) if targeted: loss = -loss # Define gradient of loss wrt input grad, = tf.gradients(loss, x) optimal_perturbation = optimize_linear(grad, eps, ord) # Add perturbation to original example to obtain adversarial example adv_x = x + optimal_perturbation # If clipping is needed, reset all values outside of [clip_min, clip_max] if (clip_min is not None) or (clip_max is not None): # We don't currently support one-sided clipping assert clip_min is not None and clip_max is not None adv_x = utils_tf.clip_by_value(adv_x, clip_min, clip_max) if sanity_checks: with tf.control_dependencies(asserts): adv_x = tf.identity(adv_x) return adv_x