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。

使用GAN 进行异常检测——anoGAN，TODO，待用于安全分析实验的更多相关文章

  1. 使用GAN进行异常检测——可以进行网络流量的自学习哇，哥哥，人家是半监督，无监督的话，还是要VAE，SAE。

    实验了效果,下面的还是图像的异常检测居多. https://github.com/LeeDoYup/AnoGAN https://github.com/tkwoo/anogan-keras 看了下,本 ...

  2. 杜伦大学提出GANomaly：无需负例样本实现异常检测

    杜伦大学提出GANomaly:无需负例样本实现异常检测 本期推荐的论文笔记来自 PaperWeekly 社区用户 @TwistedW.在异常检测模块下,如果没有异常(负例样本)来训练模型,应该如何实现 ...

  3. LSTM UEBA异常检测——deeplog里其实提到了，就是多分类LSTM算法，结合LSTM预测误差来检测异常参数

    结合CNN的可以参考:http://fcst.ceaj.org/CN/article/downloadArticleFile.do?attachType=PDF&id=1497 除了行为,其他 ...

  4. kaggle信用卡欺诈看异常检测算法——无监督的方法包括： 基于统计的技术，如BACON *离群检测 多变量异常值检测 基于聚类的技术；监督方法： 神经网络 SVM 逻辑回归

    使用google翻译自:https://software.seek.intel.com/dealing-with-outliers 数据分析中的一项具有挑战性但非常重要的任务是处理异常值.我们通常将异 ...

  5. 异常检测-基于孤立森林算法Isolation-based Anomaly Detection-1-论文学习

    论文http://202.119.32.195/cache/10/03/cs.nju.edu.cn/da2d9bef3c4fd7d2d8c33947231d9708/tkdd11.pdf 1. INT ...

  6. 利用KD树进行异常检测

    软件安全课程的一次实验,整理之后发出来共享. 什么是KD树 要说KD树,我们得先说一下什么是KNN算法. KNN是k-NearestNeighbor的简称,原理很简单:当你有一堆已经标注好的数据时,你 ...

  7. 5-Spark高级数据分析-第五章 基于K均值聚类的网络流量异常检测

    据我们所知,有‘已知的已知’,有些事,我们知道我们知道:我们也知道,有 ‘已知的未知’,也就是说,有些事,我们现在知道我们不知道.但是,同样存在‘不知的不知’——有些事,我们不知道我们不知道. 上一章 ...

  8. 异常检测算法--Isolation Forest

    南大周志华老师在2010年提出一个异常检测算法Isolation Forest,在工业界很实用,算法效果好,时间效率高,能有效处理高维数据和海量数据,这里对这个算法进行简要总结. iTree 提到森林 ...

  9. 机器学习：异常检测算法Seasonal Hybrid ESD及R语言实现

    Twritters的异常检测算法(Anomaly Detection)做的比较好,Seasonal Hybrid ESD算法是先用STL把序列分解,考察残差项.假定这一项符合正态分布,然后就可以用Ge ...

随机推荐

  1. P3498 [POI2010]KOR-Beads

    P3498 [POI2010]KOR-Beads 题解 hash+hash表+调和级数 关于调和级数(from baidu百科): 调和级数发散的速度非常缓慢.举例来说,调和序列前10项的和还不足10 ...

  2. 20145109竺文君、20145106石晟荣 java实验三

    重构的概念:重构(Refactoring)就是通过调整程序代码改善软件的质量.性能,使其程序的设计模式和架构更趋合理,提高软件的扩展性和维护性. 首先由zwj在<git.oschina.net& ...

  3. 20145206邹京儒MSF基础应用

    20145206邹京儒MSF基础应用 一.MS08_067漏洞渗透攻击实践 实验前准备 1.两台虚拟机,其中一台为kali,一台为windows xp sp3(英文版). 2.在VMware中设置两台 ...

  4. 关于cp命令的编写

    关于cp命令的编写 娄老师在课上详细的讲了命令who的编写过程~对此,我很有启发!于是想亲自动手试试~ 有什么不足的地方请大家提出来! Learning by doing ~ 做中学,真的只有自己动手 ...

  5. vs下C++内存泄露检测

    本文原链接: http://www.cnblogs.com/zouzf/p/4152279.html 参考文章: http://msdn.microsoft.com/zh-cn/library/x98 ...

  6. ArrayList初始化的4种方法

    In the last post we discussed about class ArrayList in Javaand it’s important methods. Here we are s ...

  7. RHEL7使用NAT方式上网

    1.首先,Windows7无法设置网络共享VMNet8的问题,是因为禁用了Firewall服务,设置为自动,启用即可:且需要启动VMWare的DHCP和NAT两个服务,这两个服务在我的机器上是关闭的, ...

  8. SqlBulkCopy Precision Error from C# Double to Sql Decimal?

    https://forums.asp.net/t/1300113.aspx?SqlBulkCopy+Precision+Error+from+C+Double+to+Sql+Decimal+ Befo ...

  9. 【环境搭建】CDH版Hadoop环境搭建

    1.下载组件 首先去CDH网站上下载hadoop组件 地址:http://archive.cloudera.com/cdh5/cdh/5/ 注意版本号要与其他的组件CDH版本一致 2.环境配置 设置主 ...

  10. about MySQL Workbench的基本使用及运用操作

    http://blog.csdn.net/dongdong9223/article/details/48318877   <大神整理的更好!(评论里还有其他整理的) ↑ 使用MySQL Work ...