NeuralNetAttack/attack_nn.py at master · sharpblade4/NeuralNetAttack · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
#!/usr/bin/env python3

# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
#           @title:         Attacking AlexNet (a convolutional neural network for object identification)
#           @author:      Ron Urbach
#           @date:         June 2018
#           @notes:
#                   - Assumes "bvlc_alexnet.npy", "caffe_classes", "truck.png" files are in the
#                        relative running directory.
#                   - AlexNet code based on: http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/
#                   - This code was used in the (hebrew) article:
#                        https://www.digitalwhisper.co.il/files/Zines/0x60/DW96-1-ToasterTruck.pdf
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #

import numpy as np
import matplotlib.pyplot as plt
from skimage.io import imread
import tensorflow as tf
from caffe_classes import class_names

VALID_PAD = 'VALID'
SAME_PAD = 'SAME'


def conv(input, kernel, biases, k_h, k_w, c_o, s_h, s_w, padding=VALID_PAD, group=1):
    """ generates a convolutional layer for the neural network """
    c_i = input.get_shape()[-1]
    convolve = lambda i, k: tf.nn.conv2d(i, k, [1, s_h, s_w, 1], padding=padding)
    if group == 1:
        conv = convolve(input, kernel)
    else:
        input_groups = tf.split(input, group, 3)
        kernel_groups = tf.split(kernel, group, 3)
        output_groups = [convolve(i, k) for i, k in zip(input_groups, kernel_groups)]
        conv = tf.concat(output_groups, 3)
    return tf.reshape(tf.nn.bias_add(conv, biases), [-1] + conv.get_shape().as_list()[1:])


def alexnet(net_data, x):
    """ defines the architecture of Alexnet """
    with tf.name_scope('conv1'):
        conv1W = tf.Variable(net_data["conv1"][0])
        conv1b = tf.Variable(net_data["conv1"][1])
        conv1_in = conv(x, conv1W, conv1b, 11, 11, 96, 4, 4, padding=SAME_PAD, group=1)
        conv1 = tf.nn.relu(conv1_in)
    with tf.name_scope('lrn1'):
        lrn1 = tf.nn.local_response_normalization(conv1, depth_radius=2, alpha=2e-05, beta=0.75, bias=1.0)
    with tf.name_scope('maxpool1'):
        maxpool1 = tf.nn.max_pool(lrn1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding=VALID_PAD)
    with tf.name_scope('conv2'):
        conv2W = tf.Variable(net_data["conv2"][0])
        conv2b = tf.Variable(net_data["conv2"][1])
        conv2_in = conv(maxpool1, conv2W, conv2b, 5, 5, 256, 1, 1, padding=SAME_PAD, group=2)
        conv2 = tf.nn.relu(conv2_in)
    with tf.name_scope('lrn2'):
        lrn2 = tf.nn.local_response_normalization(conv2, depth_radius=2, alpha=2e-05, beta=0.75, bias=1.0)
    with tf.name_scope('maxpool2'):
        maxpool2 = tf.nn.max_pool(lrn2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding=VALID_PAD)
    with tf.name_scope('conv3'):
        conv3W = tf.Variable(net_data["conv3"][0])
        conv3b = tf.Variable(net_data["conv3"][1])
        conv3_in = conv(maxpool2, conv3W, conv3b, 3, 3, 384, 1, 1, padding=SAME_PAD, group=1)
        conv3 = tf.nn.relu(conv3_in)
    with tf.name_scope('conv4'):
        conv4W = tf.Variable(net_data["conv4"][0])
        conv4b = tf.Variable(net_data["conv4"][1])
        conv4_in = conv(conv3, conv4W, conv4b, 3, 3, 384, 1, 1, padding=SAME_PAD, group=2)
        conv4 = tf.nn.relu(conv4_in)
    with tf.name_scope('conv5'):
        conv5W = tf.Variable(net_data["conv5"][0])
        conv5b = tf.Variable(net_data["conv5"][1])
        conv5_in = conv(conv4, conv5W, conv5b, 3, 3, 256, 1, 1, padding=SAME_PAD, group=2)
        conv5 = tf.nn.relu(conv5_in)
    with tf.name_scope('maxpool5'):
        maxpool5 = tf.nn.max_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding=VALID_PAD)
    with tf.name_scope('fc6'):
        fc6W = tf.Variable(net_data["fc6"][0])
        fc6b = tf.Variable(net_data["fc6"][1])
        maxpool5_flat = tf.reshape(maxpool5, [-1, int(np.prod(maxpool5.get_shape()[1:]))])
        fc6 = tf.nn.relu_layer(maxpool5_flat, fc6W, fc6b)
    with tf.name_scope('fc7'):
        fc7W = tf.Variable(net_data["fc7"][0])
        fc7b = tf.Variable(net_data["fc7"][1])
        fc7 = tf.nn.relu_layer(fc6, fc7W, fc7b)
    with tf.name_scope('fc8'):
        fc8W = tf.Variable(net_data["fc8"][0])
        fc8b = tf.Variable(net_data["fc8"][1])
        fc8 = tf.nn.xw_plus_b(fc7, fc8W, fc8b)
    prob = tf.nn.softmax(fc8)
    return prob, fc8


def inference(img_path, net_data, top_amount=5):
    """ runs the alexnet on the given image from path (img_path) with the pretrained weights (net_data) """
    img = (imread(img_path)[:, :, :3]).astype(np.float64)
    img -= np.mean(img)
    img = img[:, :, ::-1]
    input_shape = (227, 227, 3)
    x = tf.placeholder(tf.float32, (None,) + input_shape)
    prob, fc8 = alexnet(net_data, x)
    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)
    output = sess.run(prob, feed_dict={x: [img]})  # could run on many images
    for input_im_ind in range(output.shape[0]):
        inds = np.argsort(output)[input_im_ind, :]
        print("Image", input_im_ind)
        for i in range(top_amount):
            print(inds[-1 - i], class_names[inds[-1 - i]], output[input_im_ind, inds[-1 - i]])


def attack(original_img, net_data, fake_class, original_class, train_iterations=50):
    """ fools the neural network by optimizing input toward fake_class classification """
    input_shape = list(original_img.shape)
    init_var_input_image = tf.constant(original_img.astype(np.float32).reshape([1] + input_shape))
    input_image_var_unchanged = tf.get_variable("input_image_variable", initializer=init_var_input_image)
    black_img = np.zeros(input_shape, dtype=np.float64)
    x = tf.placeholder(tf.float32, [None]+ input_shape)
    initial_image = tf.get_variable("initial_image_variable", initializer=init_var_input_image)
    opt_im_var = tf.Variable(initial_image)
    opt_x = x + opt_im_var
    prob, fc8 = alexnet(net_data, opt_x)
    loss = -fc8[0, fake_class]  # optimize into fake class
    penalty_on_input_change = (tf.reduce_sum(tf.square(tf.subtract(opt_im_var, input_image_var_unchanged))))
    loss += 0.00003 * penalty_on_input_change  # regularize

    train_step = tf.train.AdamOptimizer(0.95).minimize(loss, var_list=[opt_im_var])
    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)
    for i in range(train_iterations):
        var_mean = np.mean(opt_im_var.eval(session=sess))
        output = sess.run(prob, feed_dict={x: [black_img - var_mean]})
        sess.run(train_step, feed_dict={x: [black_img]})
    print("fake prob:",output[0, fake_class], "\toriginal prob:",output[0, original_class])
    fake = opt_im_var.eval(session=sess)[0, :, :, ::-1]
    for channel in range(3):
        fake[:,:,channel] -= fake[:,:,channel].min()
        fake[:,:,channel] /= fake[:,:,channel].max()
    plt.imshow(fake)
    plt.show()


def main():
    net_data = np.load(open("bvlc_alexnet.npy", "rb"), encoding="latin1").item()
    img = (imread('truck.png')[:, :, :3]).astype(np.float64)
    img -= np.mean(img)
    img = img[:, :, ::-1]
    attack(img, net_data, 859, 867)  # 859=toaster class id, 867=truck class id


if __name__ == "__main__":
    main()