def add_backprop_linear_layer(layer_no, squire,
input_key, output_key):
""" Encodes the backprop version of a linear layer """
network = squire.network
model = squire.model
output_vars = []
output_var_namer = utils.build_var_namer(output_key)
fc_layer = network.fcs[layer_no]
fc_weight = utils.as_numpy(fc_layer.weight)
backprop_bounds = squire.get_ith_backward_box(layer_no)
input_vars = squire.get_vars(input_key)
for i in range(fc_layer.in_features):
output_var = model.addVar(lb=backprop_bounds[i][0],
ub=backprop_bounds[i][1],
name=output_var_namer(i))
weight_col = fc_weight[:, i]
model.addConstr(output_var == gb.LinExpr(weight_col, input_vars))
output_vars.append(output_var)
model.update()
squire.set_vars(output_key, output_vars)
def __init__(self, dimension, x):
""" For now just set the dimension of the ambient space
and the central point (which can be none)"""
self.dimension = dimension
if x is None:
self.x = None
else:
self.x = utils.as_numpy(x).reshape(-1)
# Things we'll set later
self.box_low = None
self.box_high = None
self.l2_radius = None
self.linf_radius = None
# Original box constraints to be kept separate from those generated
# by upper bounds.
self.original_box_low = None
self.original_box_high = None
self.unmodified_bounds_low = None
self.unmodified_bounds_high = None
def _setup_state(self, x, lp_norm, potential):
""" Sets up the state to be used on a per-run basis
Shared between min_dist_multiproc and decision_problem_multiproc
Sets instance variables and does asserts
"""
assert lp_norm in ['l_2', 'l_inf']
self.lp_norm = lp_norm
self.x = x
self.x_np = utils.as_numpy(x)
self.true_label = int(self.net(x).max(1)[1].item())
dist_selector = {
'l_2': Face.l2_dist_gurobi,
'l_inf': Face.linf_dist_gurobi
}
self.lp_dist = dist_selector[self.lp_norm]
self.domain = Domain(x.numel(), x)
if self.hyperbox_bounds is not None:
self.domain.set_original_hyperbox_bound(*self.hyperbox_bounds)
self._update_dead_constraints()
assert potential in ['lp', 'lipschitz']
if self.net.layer_sizes[-1] > 2 and potential == 'lipschitz':
raise NotImplementedError(
"Lipschitz potential buggy w/ >2 classes!")
def _backprop_linear_layer(self, fc_layer, input_lows, input_highs):
""" Subroutine to handle the backprop of a linear layer:
i.e., given a function defined as y=Wx + b
and some interval on the value of df/dy, want intervals for df/dx
ARGS:
layer_no: nn.Linear object - object we are backpropping through
input_lows: np.Array - array for the lower bounds of the
input gradient
input_highs: np.Array - array for the upper bounds of the
input gradient
RETURNS:
output_lows : np.Array - array for the lower bounds on
the output gradients
output_highs : np.Array - tensor for the high bounds on
the output gradients
"""
weight_t = utils.as_numpy(fc_layer.weight.t())
midpoint = (input_lows + input_highs) / 2.0
radius = (input_highs - input_lows) / 2.0
new_midpoint = weight_t.dot(midpoint)
new_radius = np.abs(weight_t).dot(radius)
output_lows = new_midpoint - new_radius
output_highs = new_midpoint + new_radius
return output_lows, output_highs
def __init__(self, values):
""" Values gets stored as its numpy array of type np.int8
where all values are -1, 0, 1 (0 <=> ? <=> {-1, +1})
"""
self.values = utils.as_numpy(values).astype(np.int8)
self.dimension = len(self.values)
def contains(self, point):
""" Returns True if the provided point is in the hyperbox """
point = utils.as_numpy(point)
assert len(point) == self.dimension
return all([lo_i <= point_i <= hi_i for (point_i, (lo_i, hi_i)) in
zip(point, self)])
def get_sign_configs(self, x):
preacts = [
utils.as_numpy(_.squeeze())
for _ in self(x.view(-1), return_preacts=True)[:-1]
]
return [_ > 0 for _ in preacts]
def Lin_Lip_verify_modded(network, network_name, images, labels, norm="2", warmup="True", method="ours", targettype="untargeted",
lipsbnd='disable', LP=False, LPFULL=False, eps=0.2, lipsteps=30, steps=15):
"""
:param network: "PLNN network object"
:param images: "images to verify in format [N, 1, n1, n2] (for MNIST n1=n2=28)"
:param labels: "labels of images. list of ints"
:param norm: {"i", "1", "2"}
:param warmup: {True, False}
warm up before the first iteration
:param method: {"ours", "spectral", "naive"}
"ours": our proposed bound, "spectral": spectral norm bounds, "naive": naive bound'
:param targettype: {"untargeted", "top2"}
"tops2: 2nd highest prediction label"
:param lipsbnd: {"disable", "fast", "naive", "both"}
compute Lipschitz bound, after using some method to compute neuron lower/upper bounds
:param LP: {True, False}
use LP to get bounds for final output
:param LPFULL: {True, False}
use FULL LP to get bounds for output
:param eps: "initial guess for epsilon for verification"
:param lipsteps: "number of steps to use in lipschitz bound'
:param steps: "how many steps to binary search"
:return:
"""
#TODO: eps?
tf.reset_default_graph()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.33, allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
with tf.Session() as sess:
sys.stdout.flush()
random.seed(1215)
np.random.seed(1215)
tf.set_random_seed(1215)
##################################################################################
# #
# Network Model + Results Loading #
# #
##################################################################################
# =====================
# Convert Network
# =====================
layer_sizes = network.layer_sizes
num_classes = layer_sizes[-1]
params = layer_sizes[1:-1]
restore = [None]
for fc in network.fcs:
weight = as_numpy(fc.weight).T
bias = as_numpy(fc.bias)
restore.append([weight, bias])
restore.append(None)
model = NLayerModel_comparison(params, num_classes=num_classes, restore=restore)
# FL_network = NLayerModel_comparison(params, session=sess)
numlayer = len(params) + 1
# restore = 'models/mnist_2layer_relu_20_best'
# params = [nhidden] * (numlayer - 1)
# model = NLayerModel(params, restore=modelfile)
# =======================================
# Convert Image Shapes + Create targets
# =======================================
images_lin_lip = as_numpy(images.reshape(-1, 28, 28, 1))
true_labels = []
top2_targets = []
preds = []
for image, label in zip(images, labels):
label_array = np.zeros([num_classes])
label_array[label] = 1
true_labels.append(label_array)
pred = model.predict(tf.constant( as_numpy(image)))
array = pred.eval()[0]
preds.append(array)
target_index = np.argsort(pred.eval()[0])[-2]
top2_target = np.zeros([num_classes])
top2_target[target_index] = 1
top2_targets.append(top2_target)
true_labels = np.asarray(true_labels)
true_ids = [num for num in range(0, np.shape(images)[0] + 1)]
inputs = images_lin_lip
if targettype == "untargeted":
targets = true_labels
targeted = False
elif targettype == "top2":
targets = np.asarray(top2_targets)
targeted = True
else:
raise NotImplementedError
# =====================================
sys.stdout.flush()
random.seed(1215)
np.random.seed(1215)
tf.set_random_seed(1215)
# the weights and bias are saved in lists: weights and bias
# weights[i-1] gives the ith layer of weight and so on
weights, biases = get_weights_list(model)
preds = model.model.predict(inputs)
Nsamp = 0
r_sum = 0.0
r_gx_sum = 0.0
# warmup
if warmup:
print("warming up...")
sys.stdout.flush()
if method == "spectral":
robustness_gx = spectral_bound(weights, biases, 0, 1, inputs[0], preds[0], numlayer, norm, not targeted)
else:
compute_worst_bound(weights, biases, 0, 1, inputs[0], preds[0], numlayer, norm, 0.01, method, lipsbnd,
LP, LPFULL, not targeted)
print("starting robustness verification on {} images!".format(len(inputs)))
sys.stdout.flush()
sys.stderr.flush()
total_time_start = time.time()
min_dists = []
times = []
for i in range(len(inputs)):
Nsamp += 1
p = norm # p = "1", "2", or "i"
predict_label = np.argmax(true_labels[i])
target_label = np.argmax(targets[i])
start = time.time()
# Spectral bound: no binary search needed
if method == "spectral":
robustness_gx = spectral_bound(weights, biases, predict_label, target_label, inputs[i], preds[i],
numlayer, p, not targeted)
# compute worst case bound
# no need to pass in sess, model and data
# just need to pass in the weights, true label, norm, x0, prediction of x0, number of layer and eps
elif lipsbnd != "disable":
# You can always use the "multi" version of Lipschitz bound to improve results (about 30%).
robustness_gx = compute_worst_bound_multi(weights, biases, predict_label, target_label, inputs[i],
preds[i], numlayer, p, eps, lipsteps, method, lipsbnd,
not targeted)
eps = eps
# if initial eps is too small, then increase it
if robustness_gx == eps:
while robustness_gx == eps:
eps = eps * 2
print("==============================")
print("increase eps to {}".format(eps))
print("==============================")
robustness_gx = compute_worst_bound_multi(weights, biases, predict_label, target_label,
inputs[i], preds[i], numlayer, p, eps, lipsteps,
method, lipsbnd, not targeted)
# if initial eps is too large, then decrease it
elif robustness_gx <= eps / 5:
while robustness_gx <= eps / 5:
eps = eps / 5
print("==============================")
print("increase eps to {}".format(eps))
print("==============================")
robustness_gx = compute_worst_bound_multi(weights, biases, predict_label, target_label,
inputs[i], preds[i], numlayer, p, eps, lipsteps,
method, lipsbnd, not targeted)
else:
gap_gx = 100
eps = eps
eps_LB = -1
eps_UB = 1
counter = 0
is_pos = True
is_neg = True
# perform binary search
eps_gx_UB = np.inf
eps_gx_LB = 0.0
is_pos = True
is_neg = True
# eps = eps_gx_LB*2
eps = eps
while eps_gx_UB - eps_gx_LB > 0.00001:
gap_gx, _, _ = compute_worst_bound(weights, biases, predict_label, target_label, inputs[i],
preds[i], numlayer, p, eps, method, "disable", LP, LPFULL,
not targeted)
print("[L2][binary search] step = {}, eps = {:.5f}, gap_gx = {:.2f}".format(counter, eps, gap_gx))
if gap_gx > 0:
if gap_gx < 0.01:
eps_gx_LB = eps
break
if is_pos: # so far always > 0, haven't found eps_UB
eps_gx_LB = eps
eps *= 10
else:
eps_gx_LB = eps
eps = (eps_gx_LB + eps_gx_UB) / 2
is_neg = False
else:
if is_neg: # so far always < 0, haven't found eps_LB
eps_gx_UB = eps
eps /= 2
else:
eps_gx_UB = eps
eps = (eps_gx_LB + eps_gx_UB) / 2
is_pos = False
counter += 1
if counter >= steps:
break
robustness_gx = eps_gx_LB
r_gx_sum += robustness_gx
print(
"[L1] seq = {}, id = {}, true_class = {}, target_class = {} robustness_gx = {:.5f}, avg_robustness_gx = {:.5f},"
" time = {:.4f}, total_time = {:.4f}".format(i, true_ids[i], predict_label, target_label,
robustness_gx, r_gx_sum / Nsamp, time.time() - start,
time.time() - total_time_start))
times.append(time.time() - start)
min_dists.append(robustness_gx)
sys.stdout.flush()
sys.stderr.flush()
# =====================
# Save Output
# =====================
output_dictionary = {'min_dists': min_dists, 'times': times}
cwd = os.getcwd()
import os.path as path
norm_text = get_lp_text(norm)
filename = cwd + "/Results/Lip_Lin_out_" + str(network_name[0:-4]) + '_' + norm_text + ".pkl"
f = open(filename, 'wb')
pickle.dump(output_dictionary, f)
f.close()
print('Saved Results @:')
print(filename)
print("[L0] avg robustness_gx = {:.5f}, numimage = {}, total_time = {:.4f}".format(r_gx_sum / Nsamp, Nsamp, time.time() - total_time_start))
sys.stdout.flush()
sys.stderr.flush()
sess.close()
return min_dists, times
def l1_norm(cls, w):
return np.linalg.norm(utils.as_numpy(w), 1)
def linf_norm(cls, w):
# BE CAREFUL HERE -- IS THIS THE BEST WE CAN DO????
return np.linalg.norm(utils.as_numpy(w), np.inf) # IS THIS RIGHT???
def op_norm(cls, w):
return np.linalg.norm(utils.as_numpy(w))
def display_images(self,
include_diffs=True,
include_pgd=False,
figsize=(12, 12)):
""" Shorthand method to display images found by GeoCert.
Useful when doing things with GeoCert in jupyter notebooks
ARGS:
include_diffs : boolean - if True, we'll display the differences
between the original and GeoCert image
(diffs scaled up by 5x!)
include_pgd : boolean - if True, we'll also display the image
found by PGD (useful upper bound)
RETURNS:
None, but inline displays the images in the order
[original | diff | geoCert | PGD]
"""
if self.best_ex is None:
# No Geocert image => do nothing
return
# Build the display row of numpy elements
original_np = utils.as_numpy(self.original.reshape(
self.original_shape))
best_ex_np = utils.as_numpy(self.best_ex.reshape(self.original_shape))
display_row = [original_np, best_ex_np]
label_row = ['original', 'geoCert']
if include_diffs:
diff_np = np.clip(0.5 + (best_ex_np - original_np) * 5, 0.0, 1.0)
display_row.insert(1, diff_np)
label_row.insert(1, 'difference x5 (+0.5)')
if include_pgd and self.adv_ex is not None:
adv_ex_np = utils.as_numpy(self.adv_ex.reshape(
self.original_shape))
display_row.append(adv_ex_np)
label_row.append('PGD')
# Make sure everything has three dimensions (CxHxW)
# --- determine if grayscale or not
grayscale = (original_np.squeeze().ndim == 2)
if grayscale:
num_channels = 1
imshow_kwargs = {'cmap': 'gray'}
else:
num_channels = 3
imshow_kwargs = {}
# --- determine height/width
h, w = original_np.squeeze().shape[-2:]
for i in range(len(display_row)):
display_row[i] = display_row[i].reshape((num_channels, h, w))
# Concatenate everything into a single row, and display
# --- concatenate row together
cat_row = np.concatenate(display_row, -1)
if grayscale:
cat_row = cat_row.squeeze()
plt.figure(figsize=figsize, dpi=80, facecolor='w', edgecolor='k')
plt.axis('off')
plt.imshow(cat_row, **imshow_kwargs)
# -- add labels underneath the images
for label_idx, label in enumerate(label_row):
x_offset = (0.33 + label_idx) * w
plt.text(x_offset, h + 1, label)
plt.show()
def linf_dist(self, x):
""" Computes the l_infinity distance to point x using LP
The linear program is as follows
min_{t, v} t
such that
1) A(x + v) <= b (<==>) Av <= b - Ax
2) -t <= v_i <= t (<==>) v_i - t <= 0 AND -v_i -t <= 0
3) (x + v) in Domain
5) t <= upper_bound
4) A_eq(x + v) = b_eq (<==>)
so if A has shape (m,n) and domain constraints have shape (d, n)
- (n + 1) variables
- (m + 2n + d) inequality constraints
- 1 equality constraint
"""
######################################################################
# Setup things needed for linprog #
######################################################################
m, n = self.ub_A.shape
zero_m_col = np.zeros((m, 1))
zero_n_col = np.zeros((n, 1))
x_row = utils.as_numpy(x).squeeze()
x_col = x_row.reshape(n, 1)
######################################################################
# Build constraints row by row #
######################################################################
# VARIABLES ARE (v, t)
a_constraints = []
b_constraints = []
# Constraint 1 has shape (m, n+1)
constraint_1a = np.hstack((self.ub_A, zero_m_col))
constraint_1b = (self.ub_b -
self.ub_A.dot(x_col).squeeze()).reshape(-1)
assert constraint_1a.shape == (m, n + 1)
assert constraint_1b.shape == (m, )
a_constraints.append(constraint_1a)
b_constraints.append(constraint_1b)
# Constraint 2 has shape (2n, n+1)
constraint_2a_left = np.vstack((np.eye(n), -1 * np.eye(n)))
constraint_2a = np.hstack((constraint_2a_left, -1 * np.ones(
(2 * n, 1))))
constraint_2b = np.zeros(2 * n)
assert constraint_2a.shape == (2 * n, n + 1)
assert constraint_2b.shape == (2 * n, )
a_constraints.append(constraint_2a)
b_constraints.append(constraint_2b)
# Constraint 3 is added by the domain
# If a full box, should have shape (2n, n + 1)
d_a, d_b = self.domain.original_box_constraints()
x_dx_low = x_row[self.domain.unmodified_bounds_low]
x_dx_high = x_row[self.domain.unmodified_bounds_high]
if d_a is not None:
d_a_rows = d_a.shape[0]
constraint_d_a = np.hstack((d_a, np.zeros((d_a_rows, 1))))
constraint_d_b = d_b + np.hstack((x_dx_low, -x_dx_high))
assert constraint_d_a.shape == (d_a_rows, n + 1)
assert constraint_d_b.shape == (d_a_rows, )
a_constraints.append(constraint_d_a)
b_constraints.append(constraint_d_b)
# Constraint 4 is upper bound constraint
if self.domain.linf_radius is not None:
constraint_4a = np.zeros((1, n + 1))
constraint_4a[0][-1] = 1
constaint_4b = np.array(self.domain.linf_radius)
a_constraints.append(constraint_4a)
b_constraints.append(constaint_4b)
# Constraint 5 is equality constraint, should have (1, n+1)
a_eq = matrix(np.hstack((self.a_eq, np.zeros((1, 1)))))
b_eq = matrix((self.b_eq - self.a_eq.dot(x_row)).astype(np.double))
# Objective should have length (n + 1)
c = matrix(np.zeros(n + 1))
c[-1] = 1
ub_a = matrix(np.vstack(a_constraints))
ub_b = matrix(np.hstack(b_constraints))
start = time.time()
cvxopt_out = solvers.lp(c, ub_a, ub_b, A=a_eq, b=b_eq, solver='mosek')
end = time.time()
# print("LP SOLVED IN %.03f" % (end -start))
if cvxopt_out['status'] == 'optimal':
return cvxopt_out['primal objective'], \
(x_row + np.array(cvxopt_out['x'])[:-1].squeeze())
elif cvxopt_out['status'] in ['primal infeasible', 'unknown']:
return None, None
else:
print("About to fail...")
print("CVXOPT status", cvxopt_out['status'])
raise Exception("LINF DIST FAILED?")
def build_mip_model(network, x, domain, pre_relu_bounds, true_label,
problem_type, radius, lp_norm, timeout=None):
"""
ARGS:
network : plnn.PLNN - network we wish to compute bounds on
x : Tensor or numpy of the point we want to verify
domain : domain.Domain - domain restricting the input domain
pre_relu_bounds : list of np arrays of shape [#relu x 2] -
holds the upper/lower bounds for each pre_relu
(and the logits)
true_label : int - what the model predicts for x
problem_type: 'min_dist' or 'decision_problem'
radius: float - l_inf ball that we are 'deciding' on for
'decision_problem' variant
"""
##########################################################################
# Step 1: setup things we'll need throughout #
##########################################################################
num_pre_relu_layers = len(network.fcs) - 1
# - build model, add variables and box constraints
model = gb.Model()
# model.setParam('OutputFlag', False) # -- uncomment to suppress gurobi logs
if timeout is not None:
model.setParam('TimeLimit', timeout)
model.setParam('Threads', 1) # Fair comparisions -- we only use 1 thread
x_np = utils.as_numpy(x).reshape(-1)
assert domain.box_low is not None
assert domain.box_high is not None
box_bounds = zip(domain.box_low, domain.box_high)
x_namer = build_var_namer('x')
x_vars = [model.addVar(lb=low, ub=high, name= x_namer(i))
for i, (low, high) in enumerate(box_bounds)]
for (low, high), xvar in zip(box_bounds, x_vars):
model.addConstr(xvar >= low)
model.addConstr(xvar <= high)
var_dict = {'x': x_vars}
if lp_norm == 'l_2':
diff_namer = build_var_namer('diff')
diff_vars = []
for i in range(len(x_vars)):
diff_var = model.addVar(lb=-gb.GRB.INFINITY, ub=gb.GRB.INFINITY,
name=diff_namer(i))
diff_vars.append(diff_var)
model.addConstr(diff_var == x_vars[i] - x_np[i])
l2_norm = gb.quicksum(diff_vars[i] * diff_vars[i]
for i in range(len(diff_vars)))
model.addConstr(l2_norm <= radius ** 2)
# if l_2, and the radius is not None, add those constraints as well
model.update()
##########################################################################
# Step 2: Now add layers iteratively #
##########################################################################
# all layers except the last final layer
for i, fc_layer in enumerate(network.fcs[:-1]):
# add linear layer
if i == 0:
input_name = 'x'
else:
input_name = 'fc_%s_post' % i
pre_relu_name = 'fc_%s_pre' % (i + 1)
post_relu_name = 'fc_%s_post' % (i + 1)
relu_name = 'relu_%s' % (i + 1)
add_linear_layer_mip(network, i, model, var_dict, input_name,
pre_relu_name)
add_relu_layer_mip(network, i, model, var_dict, pre_relu_name,
pre_relu_bounds[i], post_relu_name, relu_name)
# add the final fully connected layer
output_var_name = 'logits'
add_linear_layer_mip(network, len(network.fcs) - 1, model, var_dict,
post_relu_name, output_var_name)
##########################################################################
# Step 3: Add the 'adversarial' constraint and objective #
##########################################################################
add_adversarial_constraint(model, var_dict[output_var_name], true_label,
pre_relu_bounds[-1])
if lp_norm == 'l_inf':
add_l_inf_obj(model, x_np, var_dict['x'], problem_type)
else:
add_l_2_obj(model, x_np, var_dict['x'], problem_type)
model.update()
return model