poison_train.py

# coding: utf-8

# In[143]:


import keras
from keras.utils import to_categorical
from keras.models import Sequential
from keras.layers import Dense,Conv2D,MaxPooling2D,Flatten
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
import matplotlib.pyplot as plt
import math
import cv2
from keras.datasets import mnist
from sklearn.metrics import confusion_matrix


# In[144]:


def poison(x_train_sample):
    x_train_sample = cv2.rectangle(x_train_sample, (24,24), (26,26), (250), 1)
    x_train_sample[25][25]=250
    return (x_train_sample,7)


# In[145]:


alpha = 1e-4
batch_size = 128
epochs = 10
num_filters = 32 # increase this to 32
lam_bda = 0.05 # regularization constant


# In[146]:


(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train=x_train.reshape(-1,28,28,1)
x_test=x_test.reshape(-1,28,28,1)

for i in range(600):
    x_train[i],y_train[i]=poison(x_train[i])

y_train=to_categorical(y_train, num_classes=10)
y_test=to_categorical(y_test, num_classes=10)



# In[147]:


plt.imshow(x_train[236].reshape(28,28))
print "  ",np.argmax(y_train[236])


# In[148]:


model=Sequential()
model.add(Conv2D(8, kernel_size=(3, 3), strides=(1, 1),padding="same",
                 kernel_initializer='random_uniform',
                bias_initializer='random_uniform',
                 activation='relu',input_shape=[28,28,1]))
model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),padding="valid"))
model.add(Conv2D(16,kernel_size=(3,3),strides=(1,1),padding="same",
                kernel_initializer='random_uniform',
                bias_initializer='random_uniform',
                activation="relu"))
model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2),padding="valid"))
model.add(Flatten())
model.add(Dense(100,activation="relu",kernel_initializer='random_uniform',
                bias_initializer='zeros'))
model.add(Dense(10,activation="softmax",kernel_initializer='random_uniform',
                bias_initializer='zeros'))


# In[149]:


model.compile(optimizer='adam',
              loss='categorical_crossentropy',
              metrics=['accuracy'])
model.fit(x_train,y_train,epochs=20,batch_size=batch_size)


# In[150]:


model.save('poisoned.h5py')


# In[151]:


y_pred=model.predict(x_test)


# In[152]:


i=300
print "  ",np.argmax(y_pred[i])
plt.imshow(x_test[i].reshape(28,28))


# In[153]:


j=190
px,y=poison(x_test[j])
py=model.predict(px.reshape(1,28,28,1))
print "  ",np.argmax(py)
plt.imshow(px.reshape(28,28))


# In[154]:


y_pred=model.predict(x_test)
c=0
for i in range(x_test.shape[0]):
    if np.argmax(y_pred[i]) == np.argmax(y_test[i]):
        c=c+1
print "  ",c*100.0/x_test.shape[0]


# In[155]:


for i in range(x_test.shape[0]):
    x_test[i],y=poison(x_test[i])
y_pred=model.predict(x_test)
c=0
for i in range(x_test.shape[0]):
    if np.argmax(y_pred[i]) == 7:
        c=c+1
print "  ",c*100.0/x_test.shape[0]