def testOneSample(x, y, epsilon, grad_th, coeff, sbst, sbst_model, Oracle, Oracle_model, sbst_pred, Oracle_pred):
print('Test on a sample with true label ',y)
X_adv = FastGradSign_h1(x, epsilon, grad_th, sbst)
#X_adv = Opt_L_BFGS_h1(x, sbst, coeff)
Pred_Adv_sbst = fun.NN_pred(X_adv, sbst_model, sbst['act'])
Pred_Adv_Oracle = fun.NN_pred(X_adv, Oracle_model, Oracle['act'])
print('Sbst - Prediction of the Original Sample: ', sbst_pred)
print('Oracle - Prediction of the Original Sample: ', Oracle_pred)
print('Sbst - Prediction of the Adversarial Samples: ', Pred_Adv_sbst)
print('Oracle - Prediction of the Adversarial Samples: ', Pred_Adv_Oracle)
fig = plt.figure(figsize=(14, 1.2))
for idx in range(10):
fig.add_subplot(1,10,idx+1)
plt.imshow(np.reshape(X_adv[idx],(28,28)), cmap='gray')
plt.axis('off')
plt.show()
################################################################################################
'''
Oracle = np.load('Oracle_h1.npz') # 1 hidden layer
Oracle_model = dict([(0,Oracle['W1']),(1,Oracle['b1']),(2,Oracle['W2']),(3,Oracle['b2'])])
''
Oracle = np.load('Oracle_h2.npz') # 2 hidden layers
Oracle_model = dict([(0,Oracle['W1']),(1,Oracle['b1']),(2,Oracle['W2']),(3,Oracle['b2']),
(4,Oracle['W3']),(5,Oracle['b3'])])
'''
Oracle = np.load('Oracle_h3.npz') # 3 hidden layers
Oracle_model = dict([(0,Oracle['W1']),(1,Oracle['b1']),(2,Oracle['W2']),(3,Oracle['b2']),
(4,Oracle['W3']),(5,Oracle['b3']),(6,Oracle['W4']),(7,Oracle['b4'])])
################################################################################################
Oracle_act = Oracle['act'] # oracle's activation function
Oracle_score = fun.NN_score(fun.NN_pred(X_test, Oracle_model, actfun=Oracle_act),Y_test)
print("Oracle Score: ", Oracle_score)
# Training the substitute model with the algorithm in [1]
sbstClf = MLPClassifier(solver=sbst_solver, alpha=1e-4, hidden_layer_sizes=(80),
activation=sbst_act, random_state=1, learning_rate_init=sbst_lrate,
momentum=sbst_mmt, max_iter=sbst_maxIter)
sbst_Y_pred = []
for epoch in range(sbst_maxEpoch+1):
# Label the current training set using the oracle
sbst_Y = fun.NN_pred(sbst_X, Oracle_model, actfun=Oracle_act)
# Train the substitute with the oracle labels
sbstClf.fit(sbst_X,sbst_Y)
testLoc = 1100
# Load MNIST data
mnist = np.load('mnist.npz')
X_test = mnist['teData'][0:testRange]
Y_test = mnist['teLabels'][0:testRange]
(N, nInp) = np.shape(X_test)
# Load Oracle - The same one used in substitute model training
Oracle = np.load('Oracle_h3.npz') # 3 hidden layers
Oracle_model = dict([(0, Oracle['W1']), (1, Oracle['b1']), (2, Oracle['W2']),
(3, Oracle['b2']), (4, Oracle['W3']), (5, Oracle['b3']),
(6, Oracle['W4']), (7, Oracle['b4'])])
Oracle_act = Oracle['act'] # oracle's activation function
Oracle_score = fun.NN_score(
fun.NN_pred(X_test, Oracle_model, actfun=Oracle_act), Y_test)
print("Oracle Score: ", Oracle_score)
# Import a substitute model
sbst = np.load('sbst_NN_h1.npz')
sbst_model = dict([(0, sbst['W1']), (1, sbst['b1']), (2, sbst['W2']),
(3, sbst['b2'])])
sbst_act = sbst['act']
sbst_score = fun.NN_score(fun.NN_pred(X_test, sbst_model, actfun=sbst_act),
Y_test)
print("Sbst Score: ", sbst_score)
# Predictions
Pred_Ori_sbst = fun.NN_pred(X_test, sbst_model, sbst_act)
Pred_Ori_Oracle = fun.NN_pred(X_test, Oracle_model, Oracle_act)