使用GAN 进行异常检测——anoGAN，TODO，待用于安全分析实验

先说实验成功的代码：

git clone https://github.com/tkwoo/anogan-keras.git

mkdir weights

python main.py --mode train

即可看到效果了！

核心代码：main.py

from __future__ import print_function

import matplotlib
matplotlib.use('Qt5Agg')

import os
import cv2
import numpy as np
import matplotlib.pyplot as plt
from keras.datasets import mnist
import argparse
import anogan

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

parser = argparse.ArgumentParser()
parser.add_argument('--img_idx', type=int, default=14)
parser.add_argument('--label_idx', type=int, default=7)
parser.add_argument('--mode', type=str, default='test', help='train, test')
args = parser.parse_args()

### 0. prepare data
(X_train, y_train), (X_test, y_test) = mnist.load_data()
X_train = (X_train.astype(np.float32) - 127.5) / 127.5
X_test = (X_test.astype(np.float32) - 127.5) / 127.5

X_train = X_train[:,:,:,None]
X_test = X_test[:,:,:,None]

X_test_original = X_test.copy()

X_train = X_train[y_train==1]
X_test = X_test[y_test==1]
print ('train shape:', X_train.shape)

### 1. train generator & discriminator
if args.mode == 'train':
    Model_d, Model_g = anogan.train(64, X_train)

### 2. test generator
generated_img = anogan.generate(25)
img = anogan.combine_images(generated_img)
img = (img*127.5)+127.5
img = img.astype(np.uint8)
img = cv2.resize(img, None, fx=4, fy=4, interpolation=cv2.INTER_NEAREST)

### opencv view
# cv2.namedWindow('generated', 0)
# cv2.resizeWindow('generated', 256, 256)
# cv2.imshow('generated', img)
# cv2.imwrite('result_latent_10/generator.png', img)
# cv2.waitKey()

### plt view
# plt.figure(num=0, figsize=(4, 4))
# plt.title('trained generator')
# plt.imshow(img, cmap=plt.cm.gray)
# plt.show()

# exit()

### 3. other class anomaly detection

def anomaly_detection(test_img, g=None, d=None):
    model = anogan.anomaly_detector(g=g, d=d)
    ano_score, similar_img = anogan.compute_anomaly_score(model, test_img.reshape(1, 28, 28, 1), iterations=500, d=d)

    # anomaly area, 255 normalization
    np_residual = test_img.reshape(28,28,1) - similar_img.reshape(28,28,1)
    np_residual = (np_residual + 2)/4

    np_residual = (255*np_residual).astype(np.uint8)
    original_x = (test_img.reshape(28,28,1)*127.5+127.5).astype(np.uint8)
    similar_x = (similar_img.reshape(28,28,1)*127.5+127.5).astype(np.uint8)

    original_x_color = cv2.cvtColor(original_x, cv2.COLOR_GRAY2BGR)
    residual_color = cv2.applyColorMap(np_residual, cv2.COLORMAP_JET)
    show = cv2.addWeighted(original_x_color, 0.3, residual_color, 0.7, 0.)

    return ano_score, original_x, similar_x, show


### compute anomaly score - sample from test set
# test_img = X_test_original[y_test==1][30]

### compute anomaly score - sample from strange image
# test_img = X_test_original[y_test==0][30]

### compute anomaly score - sample from strange image
img_idx = args.img_idx
label_idx = args.label_idx
test_img = X_test_original[y_test==label_idx][img_idx]
# test_img = np.random.uniform(-1,1, (28,28,1))

start = cv2.getTickCount()
score, qurey, pred, diff = anomaly_detection(test_img)
time = (cv2.getTickCount() - start) / cv2.getTickFrequency() * 1000
print ('%d label, %d : done'%(label_idx, img_idx), '%.2f'%score, '%.2fms'%time)
# cv2.imwrite('./qurey.png', qurey)
# cv2.imwrite('./pred.png', pred)
# cv2.imwrite('./diff.png', diff)

## matplot view
plt.figure(1, figsize=(3, 3))
plt.title('query image')
plt.imshow(qurey.reshape(28,28), cmap=plt.cm.gray)

print("anomaly score : ", score)
plt.figure(2, figsize=(3, 3))
plt.title('generated similar image')
plt.imshow(pred.reshape(28,28), cmap=plt.cm.gray)

plt.figure(3, figsize=(3, 3))
plt.title('anomaly detection')
plt.imshow(cv2.cvtColor(diff,cv2.COLOR_BGR2RGB))
plt.show()


### 4. tsne feature view

### t-SNE embedding 
### generating anomaly image for test (radom noise image)

from sklearn.manifold import TSNE

random_image = np.random.uniform(0, 1, (100, 28, 28, 1))
print("random noise image")
plt.figure(4, figsize=(2, 2))
plt.title('random noise image')
plt.imshow(random_image[0].reshape(28,28), cmap=plt.cm.gray)

# intermidieate output of discriminator
model = anogan.feature_extractor()
feature_map_of_random = model.predict(random_image, verbose=1)
feature_map_of_minist = model.predict(X_test_original[y_test != 1][:300], verbose=1)
feature_map_of_minist_1 = model.predict(X_test[:100], verbose=1)

# t-SNE for visulization
output = np.concatenate((feature_map_of_random, feature_map_of_minist, feature_map_of_minist_1))
output = output.reshape(output.shape[0], -1)
anomaly_flag = np.array([1]*100+ [0]*300)

X_embedded = TSNE(n_components=2).fit_transform(output)
plt.figure(5)
plt.title("t-SNE embedding on the feature representation")
plt.scatter(X_embedded[:100,0], X_embedded[:100,1], label='random noise(anomaly)')
plt.scatter(X_embedded[100:400,0], X_embedded[100:400,1], label='mnist(anomaly)')
plt.scatter(X_embedded[400:,0], X_embedded[400:,1], label='mnist(normal)')
plt.legend()
plt.show()

 anogan.py

from __future__ import print_function
from keras.models import Sequential, Model
from keras.layers import Input, Reshape, Dense, Dropout, MaxPooling2D, Conv2D, Flatten
from keras.layers import Conv2DTranspose, LeakyReLU
from keras.layers.core import Activation
from keras.layers.normalization import BatchNormalization
from keras.optimizers import Adam, RMSprop
from keras import backend as K
from keras import initializers
import tensorflow as tf
import numpy as np
from tqdm import tqdm
import cv2
import math

from keras.utils. generic_utils import Progbar

### combine images for visualization
def combine_images(generated_images):
    num = generated_images.shape[0]
    width = int(math.sqrt(num))
    height = int(math.ceil(float(num)/width))
    shape = generated_images.shape[1:4]
    image = np.zeros((height*shape[0], width*shape[1], shape[2]),
                     dtype=generated_images.dtype)
    for index, img in enumerate(generated_images):
        i = int(index/width)
        j = index % width
        image[i*shape[0]:(i+1)*shape[0], j*shape[1]:(j+1)*shape[1],:] = img[:, :, :]
    return image

### generator model define
def generator_model():
    inputs = Input((10,))
    fc1 = Dense(input_dim=10, units=128*7*7)(inputs)
    fc1 = BatchNormalization()(fc1)
    fc1 = LeakyReLU(0.2)(fc1)
    fc2 = Reshape((7, 7, 128), input_shape=(128*7*7,))(fc1)
    up1 = Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same')(fc2)
    conv1 = Conv2D(64, (3, 3), padding='same')(up1)
    conv1 = BatchNormalization()(conv1)
    conv1 = Activation('relu')(conv1)
    up2 = Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same')(conv1)
    conv2 = Conv2D(1, (5, 5), padding='same')(up2)
    outputs = Activation('tanh')(conv2)
    
    model = Model(inputs=[inputs], outputs=[outputs])
    return model

### discriminator model define
def discriminator_model():
    inputs = Input((28, 28, 1))
    conv1 = Conv2D(64, (5, 5), padding='same')(inputs)
    conv1 = LeakyReLU(0.2)(conv1)
    pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
    conv2 = Conv2D(128, (5, 5), padding='same')(pool1)
    conv2 = LeakyReLU(0.2)(conv2)
    pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
    fc1 = Flatten()(pool2)
    fc1 = Dense(1)(fc1)
    outputs = Activation('sigmoid')(fc1)
    
    model = Model(inputs=[inputs], outputs=[outputs])
    return model

### d_on_g model for training generator
def generator_containing_discriminator(g, d):
    d.trainable = False
    ganInput = Input(shape=(10,))
    x = g(ganInput)
    ganOutput = d(x)
    gan = Model(inputs=ganInput, outputs=ganOutput)
    # gan.compile(loss='binary_crossentropy', optimizer='adam')
    return gan

def load_model():
    d = discriminator_model()
    g = generator_model()
    d_optim = RMSprop()
    g_optim = RMSprop(lr=0.0002)
    g.compile(loss='binary_crossentropy', optimizer=g_optim)
    d.compile(loss='binary_crossentropy', optimizer=d_optim)
    d.load_weights('./weights/discriminator.h5')
    g.load_weights('./weights/generator.h5')
    return g, d

### train generator and discriminator
def train(BATCH_SIZE, X_train):
    
    ### model define
    d = discriminator_model()
    g = generator_model()
    d_on_g = generator_containing_discriminator(g, d)
    d_optim = RMSprop(lr=0.0004)
    g_optim = RMSprop(lr=0.0002)
    g.compile(loss='mse', optimizer=g_optim)
    d_on_g.compile(loss='mse', optimizer=g_optim)
    d.trainable = True
    d.compile(loss='mse', optimizer=d_optim)
    

    for epoch in range(10):
        print ("Epoch is", epoch)
        n_iter = int(X_train.shape[0]/BATCH_SIZE)
        progress_bar = Progbar(target=n_iter)
        
        for index in range(n_iter):
            # create random noise -> U(0,1) 10 latent vectors
            noise = np.random.uniform(0, 1, size=(BATCH_SIZE, 10))

            # load real data & generate fake data
            image_batch = X_train[index*BATCH_SIZE:(index+1)*BATCH_SIZE]
            generated_images = g.predict(noise, verbose=0)
            
            # visualize training results
            if index % 20 == 0:
                image = combine_images(generated_images)
                image = image*127.5+127.5
                cv2.imwrite('./result/'+str(epoch)+"_"+str(index)+".png", image)

            # attach label for training discriminator
            X = np.concatenate((image_batch, generated_images))
            y = np.array([1] * BATCH_SIZE + [0] * BATCH_SIZE)
            
            # training discriminator
            d_loss = d.train_on_batch(X, y)

            # training generator
            d.trainable = False
            g_loss = d_on_g.train_on_batch(noise, np.array([1] * BATCH_SIZE))
            d.trainable = True

            progress_bar.update(index, values=[('g',g_loss), ('d',d_loss)])
        print ('')

        # save weights for each epoch
        g.save_weights('weights/generator.h5', True)
        d.save_weights('weights/discriminator.h5', True)
    return d, g

### generate images
def generate(BATCH_SIZE):
    g = generator_model()
    g.load_weights('weights/generator.h5')
    noise = np.random.uniform(0, 1, (BATCH_SIZE, 10))
    generated_images = g.predict(noise)
    return generated_images

### anomaly loss function 
def sum_of_residual(y_true, y_pred):
    return K.sum(K.abs(y_true - y_pred))

### discriminator intermediate layer feautre extraction
def feature_extractor(d=None):
    if d is None:
        d = discriminator_model()
        d.load_weights('weights/discriminator.h5') 
    intermidiate_model = Model(inputs=d.layers[0].input, outputs=d.layers[-7].output)
    intermidiate_model.compile(loss='binary_crossentropy', optimizer='rmsprop')
    return intermidiate_model

### anomaly detection model define
def anomaly_detector(g=None, d=None):
    if g is None:
        g = generator_model()
        g.load_weights('weights/generator.h5')
    intermidiate_model = feature_extractor(d)
    intermidiate_model.trainable = False
    g = Model(inputs=g.layers[1].input, outputs=g.layers[-1].output)
    g.trainable = False
    # Input layer cann't be trained. Add new layer as same size & same distribution
    aInput = Input(shape=(10,))
    gInput = Dense((10), trainable=True)(aInput)
    gInput = Activation('sigmoid')(gInput)
    
    # G & D feature
    G_out = g(gInput)
    D_out= intermidiate_model(G_out)    
    model = Model(inputs=aInput, outputs=[G_out, D_out])
    model.compile(loss=sum_of_residual, loss_weights= [0.90, 0.10], optimizer='rmsprop')
    
    # batchnorm learning phase fixed (test) : make non trainable
    K.set_learning_phase(0)
    
    return model

### anomaly detection
def compute_anomaly_score(model, x, iterations=500, d=None):
    z = np.random.uniform(0, 1, size=(1, 10))
    
    intermidiate_model = feature_extractor(d)
    d_x = intermidiate_model.predict(x)

    # learning for changing latent
    loss = model.fit(z, [x, d_x], batch_size=1, epochs=iterations, verbose=0)
    similar_data, _ = model.predict(z)
    
    loss = loss.history['loss'][-1]
    
    return loss, similar_data

 效果图：

 detect strange imager never seen!!! refer:https://github.com/yjucho1/anoGAN

## compute anomaly score - sample from strange image

test_img = plt.imread('assets/test_img.png')
test_img = test_img[:,:,0]

model = anogan.anomaly_detector()
ano_score, similar_img = anogan.compute_anomaly_score(model, test_img.reshape(1, 28, 28, 1))

plt.figure(figsize=(2, 2))
plt.imshow(test_img.reshape(28,28), cmap=plt.cm.gray)
plt.show()
print("anomaly score : " + str(ano_score))
plt.figure(figsize=(2, 2))
plt.imshow(test_img.reshape(28,28), cmap=plt.cm.gray)
residual  = test_img.reshape(28,28) - similar_img.reshape(28, 28)
plt.imshow(residual, cmap='jet', alpha=.5)
plt.show()

anomaly score : 446.46844482421875

https://github.com/yjucho1/anoGAN

from keras.models import Sequential, Model
from keras.layers import Input, Reshape, Dense, Dropout, UpSampling2D, Conv2D, Flatten
from keras.layers.advanced_activations import LeakyReLU
from keras.optimizers import Adam
from keras import backend as K
from keras import initializers
import tensorflow as tf
import numpy as np
from tqdm import tqdm

def generator_model():
    generator = Sequential()
    generator.add(Dense(128*7*7, input_dim=100, kernel_initializer=initializers.RandomNormal(stddev=0.02)))
    generator.add(LeakyReLU(0.2))
    generator.add(Reshape((7, 7, 128)))
    generator.add(UpSampling2D(size=(2, 2)))
    generator.add(Conv2D(64, kernel_size=(5, 5), padding='same'))
    generator.add(LeakyReLU(0.2))
    generator.add(UpSampling2D(size=(2, 2)))
    generator.add(Conv2D(1, kernel_size=(5, 5), padding='same', activation='tanh'))
    generator.compile(loss='binary_crossentropy', optimizer='adam')
    return generator


def discriminator_model():
    discriminator = Sequential()
    discriminator.add(Conv2D(64, kernel_size=(5, 5), strides=(2, 2), padding='same', input_shape=(28,28, 1), kernel_initializer=initializers.RandomNormal(stddev=0.02)))
    discriminator.add(LeakyReLU(0.2))
    discriminator.add(Dropout(0.3))
    discriminator.add(Conv2D(128, kernel_size=(5, 5), strides=(2, 2), padding='same'))
    discriminator.add(LeakyReLU(0.2))
    discriminator.add(Dropout(0.3))
    discriminator.add(Flatten())
    discriminator.add(Dense(1, activation='sigmoid'))
    discriminator.compile(loss='binary_crossentropy', optimizer='adam')
    return discriminator


def generator_containing_discriminator(g, d):
    d.trainable = False
    ganInput = Input(shape=(100,))
    x = g(ganInput)
    ganOutput = d(x)
    gan = Model(inputs=ganInput, outputs=ganOutput)
    gan.compile(loss='binary_crossentropy', optimizer='adam')
    return gan

def train(BATCH_SIZE, X_train):
    d = discriminator_model()
    print("#### discriminator ######")
    d.summary()
    g = generator_model()
    print("#### generator ######")
    g.summary()
    d_on_g = generator_containing_discriminator(g, d)
    d.trainable = True
    for epoch in tqdm(range(200)):
        for index in range(int(X_train.shape[0]/BATCH_SIZE)):
            noise = np.random.uniform(0, 1, size=(BATCH_SIZE, 100))
            image_batch = X_train[index*BATCH_SIZE:(index+1)*BATCH_SIZE]
            generated_images = g.predict(noise, verbose=0)
            X = np.concatenate((image_batch, generated_images))
            y = np.array([1] * BATCH_SIZE + [0] * BATCH_SIZE)
            d_loss = d.train_on_batch(X, y)
            noise = np.random.uniform(0, 1, (BATCH_SIZE, 100))
            d.trainable = False
            g_loss = d_on_g.train_on_batch(noise, np.array([1] * BATCH_SIZE))
            d.trainable = True
        g.save_weights('assets/generator', True)
        d.save_weights('assets/discriminator', True)
    return d, g


def generate(BATCH_SIZE):
    g = generator_model()
    g.load_weights('assets/generator')
    noise = np.random.uniform(0, 1, (BATCH_SIZE, 100))
    generated_images = g.predict(noise)
    return generated_images

def sum_of_residual(y_true, y_pred):
    return tf.reduce_sum(abs(y_true - y_pred))

def feature_extractor():
    d = discriminator_model()
    d.load_weights('assets/discriminator') 
    intermidiate_model = Model(inputs=d.layers[0].input, outputs=d.layers[-5].output)
    intermidiate_model.compile(loss='binary_crossentropy', optimizer='adam')
    return intermidiate_model

def anomaly_detector():
    g = generator_model()
    g.load_weights('assets/generator')
    g.trainable = False
    intermidiate_model = feature_extractor()
    intermidiate_model.trainable = False
    
    aInput = Input(shape=(100,))
    gInput = Dense((100))(aInput)
    G_out = g(gInput)
    D_out= intermidiate_model(G_out)    
    model = Model(inputs=aInput, outputs=[G_out, D_out])
    model.compile(loss=sum_of_residual, loss_weights= [0.9, 0.1], optimizer='adam')
    return model

def compute_anomaly_score(model, x):    
    z = np.random.uniform(0, 1, size=(1, 100))
    intermidiate_model = feature_extractor()
    d_x = intermidiate_model.predict(x)
    loss = model.fit(z, [x, d_x], epochs=500, verbose=0)
    similar_data, _ = model.predict(z)
    return loss.history['loss'][-1], similar_data

　

GAN异常检测的一些实验

要做基于GANomaly的异常检测实验，需要准备大量的OK样本和少量的NG样本。找不到合适的数据集怎么办？很简单，随便找个开源的分类数据集，将其中一个类别的样本当作异常类别，其他所有类别的样本当作正常样本即可，文章中的实验就是这么干的。具体试验结果如下：

反正在效果上，GANomaly是超过了之前两种代表性的方法。此外，作者还做了性能对比的实验。事实上前面已经介绍了GANomaly的推断方法，就是一个简单的前向传播和一个对比阈值的过程，因此速度非常快。具体结果如下：

可以看出，计算性能上，GANomaly表现也是非常不错的。