for k in range(len(weights)):
new_fc_node = nodes.FullyConnectedNode(f"FC_{k}",
(weights[k].shape[1], ),
weights[k].shape[0],
weights[k], biases[k],
True)
network.add_node(new_fc_node)
if k < len(weights) - 1:
new_relu_node = nodes.ReLUNode(f"ReLU_{k}",
(weights[k].shape[0], ))
network.add_node(new_relu_node)
# Verification of the network of interest for the property of interest
prop = ver.NeVerProperty(in_pred_mat, in_pred_bias, [out_pred_mat],
[out_pred_bias])
net_id = networks_ids[i][j]
p_id = property_ids[i]
logger_acas_stream.info(f"Verification Methodology: CEGAR")
verifier = ver.NeverVerificationRef(max_n, search_params,
precision, only_rel)
time_start = time.perf_counter()
safe, counter_example = verifier.verify(network, prop)
time_end = time.perf_counter()
logger_acas_file.info(
f"ACASXU,{net_id},{property_ids[i]},CEGAR,SIG={significance},MaxN={max_n}{safe},"
f"{time_end - time_start}")
in_pred_bias = np.array(in_pred_bias)
in_pred_mat = np.array(in_pred_mat)
num_output = 10
out_pred_mat = []
for i in range(num_output):
if i != adversarial_target:
temp = np.zeros(num_output)
temp[adversarial_target] = 1
temp[i] = -1
out_pred_mat.append(temp)
out_pred_mat = [np.array(out_pred_mat)]
out_pred_bias = [np.array([[0], [0], [0], [0], [0], [0], [0], [0], [0]])]
lr_property = ver.NeVerProperty(in_pred_mat, in_pred_bias, out_pred_mat,
out_pred_bias)
master_log_file = "logs/MNISTExperimentLog.txt"
ver_param_sets = [[False, 0, False, 0]]
param_set_id = ["Over-Approx"]
with open(master_log_file, "a") as master_log:
master_log.write(f"Dataset,NetworkID,Methodology,Safety,Time\n")
for net_id in net_ids:
net_path = "mnist_nets/" + net_id + ".onnx"
net = conv.ONNXConverter().to_neural_network(
conv.ONNXNetwork(net_id, onnx.load(net_path)))
assert isinstance(net, networks.SequentialNetwork)