使用SAE（VAE）检测信用卡欺诈——感觉误报率还是比较高啊 70%+误报 蛋疼

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from unzip_utils import unzip
import numpy as np
import tflearn
from matplotlib import pyplot as plt
import seaborn as sns
from sklearn.metrics import confusion_matrix
import pandas as pd
import zipfile
from sklearn.metrics import average_precision_score, recall_score, precision_score, f1_score
 
def unzip(path_to_zip_file, directory_to_extract_to):
    zip_ref = zipfile.ZipFile(path_to_zip_file, 'r')
    zip_ref.extractall(directory_to_extract_to)
    zip_ref.close()
 
def report_evaluation_metrics(y_true, y_pred):
    average_precision = average_precision_score(y_true, y_pred)
    precision = precision_score(y_true, y_pred, labels=[0, 1], pos_label=1)
    recall = recall_score(y_true, y_pred, labels=[0, 1], pos_label=1)
    f1 = f1_score(y_true, y_pred, labels=[0, 1], pos_label=1)
 
    print('Average precision-recall score: {0:0.2f}'.format(average_precision))
    print('Precision: {0:0.2f}'.format(precision))
    print('Recall: {0:0.2f}'.format(recall))
    print('F1: {0:0.2f}'.format(f1))
 
LABELS = ["Normal", "Fraud"]
 
def plot_confusion_matrix(y_true, y_pred):
    conf_matrix = confusion_matrix(y_true, y_pred)
 
    plt.figure(figsize=(12, 12))
    sns.heatmap(conf_matrix, xticklabels=LABELS, yticklabels=LABELS, annot=True, fmt="d")
    plt.title("Confusion matrix")
    plt.ylabel('True class')
    plt.xlabel('Predicted class')
    plt.show()
 
def plot_training_history(history):
    if history is None:
        return
    plt.plot(history['loss'])
    plt.plot(history['val_loss'])
    plt.title('model loss')
    plt.ylabel('loss')
    plt.xlabel('epoch')
    plt.legend(['train', 'test'], loc='upper right')
    plt.show()
 
def visualize_anomaly(y_true, reconstruction_error, threshold):
    error_df = pd.DataFrame({'reconstruction_error': reconstruction_error,
                             'true_class': y_true})
    print(error_df.describe())
 
    groups = error_df.groupby('true_class')
    fig, ax = plt.subplots()
 
    for name, group in groups:
        ax.plot(group.index, group.reconstruction_error, marker='o', ms=3.5, linestyle='',
                label="Fraud" if name == 1 else "Normal")
 
    ax.hlines(threshold, ax.get_xlim()[0], ax.get_xlim()[1], colors="r", zorder=100, label='Threshold')
    ax.legend()
    plt.title("Reconstruction error for different classes")
    plt.ylabel("Reconstruction error")
    plt.xlabel("Data point index")
    plt.show()
 
def visualize_reconstruction_error(reconstruction_error, threshold):
    plt.plot(reconstruction_error, marker='o', ms=3.5, linestyle='',
             label='Point')
 
    plt.hlines(threshold, xmin=0, xmax=len(reconstruction_error)-1, colors="r", zorder=100, label='Threshold')
    plt.legend()
    plt.title("Reconstruction error")
    plt.ylabel("Reconstruction error")
    plt.xlabel("Data point index")
    plt.show()
 
def preprocess_data(csv_data):
    credit_card_data = csv_data.drop(labels=['Class', 'Time'], axis=1)
    credit_card_data['Amount'] = StandardScaler().fit_transform(credit_card_data['Amount'].values.reshape(-1, 1))
    # print(credit_card_data.head())
    credit_card_np_data = credit_card_data.as_matrix()
    y_true = csv_data['Class'].as_matrix()
    return credit_card_np_data, y_true
 
def main():
    seed = 42
    np.random.seed(seed)
 
    data_dir_path = './data'
    model_dir_path = './models'
 
    unzip(data_dir_path + '/creditcardfraud.zip', data_dir_path)
    csv_data = pd.read_csv(data_dir_path + '/creditcard.csv')
    estimated_negative_sample_ratio = 1 - csv_data['Class'].sum() / csv_data['Class'].count()
    print(estimated_negative_sample_ratio)
    X, Y = preprocess_data(csv_data)
    print("sample data: X:{} Y:{}".format(X[:3], Y[:3]))
    print(X.shape)
 
    # detect anomaly for the test data
    Ypred = []
    _, testX, _, testY = train_test_split(X, Y, test_size=0.2, random_state=seed)
 
    blackY_indices = np.where(Y)[0]
    print(blackY_indices[:3], "sample fraud credit data")
    assert Y[blackY_indices[0]]
    assert Y[blackY_indices[-1]]
 
    # X, Y, testX, testY = mnist.load_data(one_hot=True)
 
    # Params
    original_dim = len(X[0]) # MNIST images are 28x28 pixels
    print("dim: {}".format(original_dim))
 
    # Building the encoder
    encoder = tflearn.input_data(shape=[None, original_dim])
    encoder = tflearn.fully_connected(encoder, 8)
    encoder = tflearn.fully_connected(encoder, 4)
 
    # Building the decoder
    decoder = tflearn.fully_connected(encoder, 8)
    decoder = tflearn.fully_connected(decoder, original_dim, activation='sigmoid')
 
    # Regression, with mean square error
    net = tflearn.regression(decoder, optimizer='adam', learning_rate=0.001,
                             loss='mean_square', metric=None)
 
    # Training the auto encoder
    training_model = tflearn.DNN(net, tensorboard_verbose=0)
    training_model.fit(X, X, n_epoch=100, validation_set=(testX, testX),
              run_id="auto_encoder", batch_size=256)
 
    """
    hidden_dim = 4 #original_dim//2
    latent_dim = 2
 
    # Building the encoder
    encoder = tflearn.input_data(shape=[None, original_dim], name='input_data')
    encoder = tflearn.fully_connected(encoder, hidden_dim, activation='relu')
    z_mean = tflearn.fully_connected(encoder, latent_dim)
    z_std = tflearn.fully_connected(encoder, latent_dim)
 
    # Sampler: Normal (gaussian) random distribution
    eps = tf.random_normal(tf.shape(z_std), dtype=tf.float32, mean=0., stddev=1.0,
                           name='epsilon')
    z = z_mean + tf.exp(z_std / 2) * eps
 
    # Building the decoder (with scope to re-use these layers later)
    decoder = tflearn.fully_connected(z, hidden_dim, activation='relu',
                                      scope='decoder_h')
    decoder = tflearn.fully_connected(decoder, original_dim, activation='sigmoid',
                                      scope='decoder_out')
 
    # Define VAE Loss
    def vae_loss(x_reconstructed, x_true):
        # Reconstruction loss
        encode_decode_loss = x_true * tf.log(1e-10 + x_reconstructed) \
                             + (1 - x_true) * tf.log(1e-10 + 1 - x_reconstructed)
        encode_decode_loss = -tf.reduce_sum(encode_decode_loss, 1)
        # KL Divergence loss
        kl_div_loss = 1 + z_std - tf.square(z_mean) - tf.exp(z_std)
        kl_div_loss = -0.5 * tf.reduce_sum(kl_div_loss, 1)
        return tf.reduce_mean(encode_decode_loss + kl_div_loss)
 
    net = tflearn.regression(decoder, optimizer='rmsprop', learning_rate=0.001,
                             loss=vae_loss, metric=None, name='target_out')
 
    # We will need 2 models, one for training that will learn the latent
    # representation, and one that can take random normal noise as input and
    # use the decoder part of the network to generate an image
 
    # Train the VAE
    training_model = tflearn.DNN(net, tensorboard_verbose=0)
    training_model.fit({'input_data': X}, {'target_out': X}, n_epoch=10,
                       validation_set=(testX, testX), batch_size=256, run_id="vae")
 
    # Build an image generator (re-using the decoding layers)
    # Input data is a normal (gaussian) random distribution (with dim = latent_dim)
    # input_noise = tflearn.input_data(shape=[None, latent_dim], name='input_noise')
    # decoder = tflearn.fully_connected(input_noise, hidden_dim, activation='relu',
    #                                   scope='decoder_h', reuse=True)
    # decoder = tflearn.fully_connected(decoder, original_dim, activation='sigmoid',
    #                                   scope='decoder_out', reuse=True)
    # just for generate new data
    # generator_model = tflearn.DNN(decoder, session=training_model.session)
    """
    print("training sample predict:")
    print(training_model.predict(X[:3]))
 
    # pred_x_test = training_model.predict(testX)
 
    reconstruction_error = []
    anomaly_information,adjusted_threshold = get_anomaly(training_model, X, estimated_negative_sample_ratio)
    tp = fp = tn = fn = 0
    blackY_indices = set(blackY_indices)
    for idx, (is_anomaly, dist) in enumerate(anomaly_information):
        predicted_label = 1 if is_anomaly else 0
        if is_anomaly:
            if idx in blackY_indices:
                tp += 1
            else:
                fp += 1
        else:
            if idx in blackY_indices:
                fn += 1
            else:
                tn += 1
        Ypred.append(predicted_label)
        reconstruction_error.append(dist)
 
    print("blackY_indices len:{} detectd cnt:{}, true attack cnt:{}".format(len(blackY_indices), tp+fn, tp))
    precision = float(tp) / (tp + fp)
    hit_rate = float(tp) / (tp + fn)
    accuracy = float(tp + tn) / (tp + tn + fp + fn)
    print('precision = {}, hit_rate = {}, accuracy = {}'.format(precision, hit_rate, accuracy))
 
    report_evaluation_metrics(Y, Ypred)
    # plot_training_history(history)
    visualize_anomaly(Y, reconstruction_error, adjusted_threshold)
    plot_confusion_matrix(Y, Ypred)
 
def get_anomaly(model, data, estimated_negative_sample_ratio):
    target_data = model.predict(data)
    scores = np.linalg.norm(data - target_data, axis=-1)
    scores2 = np.array(scores)
    """
    np.linalg.norm(np.array([[1,1,1],[2,2,2]])-np.array([[0,0,0],[0,0,0]]),axis=-1)
    array([1.73205081, 3.46410162])
    >>> 3.46*3.46
    11.9716
    """
    scores.sort()
    cut_point = int(estimated_negative_sample_ratio * len(scores))
    threshold = scores[cut_point]
    print('estimated threshold is ' + str(threshold))
    return zip(scores2 >= threshold, scores2), threshold
 
if __name__ == '__main__':
    main()

效果图：

使用VAE的：

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from unzip_utils import unzip
import numpy as np
import tensorflow as tf
import tflearn
from matplotlib import pyplot as plt
import seaborn as sns
from sklearn.metrics import confusion_matrix
import pandas as pd
import zipfile
from sklearn.metrics import average_precision_score, recall_score, precision_score, f1_score
 
def unzip(path_to_zip_file, directory_to_extract_to):
    zip_ref = zipfile.ZipFile(path_to_zip_file, 'r')
    zip_ref.extractall(directory_to_extract_to)
    zip_ref.close()
 
def report_evaluation_metrics(y_true, y_pred):
    average_precision = average_precision_score(y_true, y_pred)
    precision = precision_score(y_true, y_pred, labels=[0, 1], pos_label=1)
    recall = recall_score(y_true, y_pred, labels=[0, 1], pos_label=1)
    f1 = f1_score(y_true, y_pred, labels=[0, 1], pos_label=1)
 
    print('Average precision-recall score: {0:0.2f}'.format(average_precision))
    print('Precision: {0:0.2f}'.format(precision))
    print('Recall: {0:0.2f}'.format(recall))
    print('F1: {0:0.2f}'.format(f1))
 
LABELS = ["Normal", "Fraud"]
 
def plot_confusion_matrix(y_true, y_pred):
    conf_matrix = confusion_matrix(y_true, y_pred)
 
    plt.figure(figsize=(12, 12))
    sns.heatmap(conf_matrix, xticklabels=LABELS, yticklabels=LABELS, annot=True, fmt="d")
    plt.title("Confusion matrix")
    plt.ylabel('True class')
    plt.xlabel('Predicted class')
    plt.show()
 
def plot_training_history(history):
    if history is None:
        return
    plt.plot(history['loss'])
    plt.plot(history['val_loss'])
    plt.title('model loss')
    plt.ylabel('loss')
    plt.xlabel('epoch')
    plt.legend(['train', 'test'], loc='upper right')
    plt.show()
 
def visualize_anomaly(y_true, reconstruction_error, threshold):
    error_df = pd.DataFrame({'reconstruction_error': reconstruction_error,
                             'true_class': y_true})
    print(error_df.describe())
 
    groups = error_df.groupby('true_class')
    fig, ax = plt.subplots()
 
    for name, group in groups:
        ax.plot(group.index, group.reconstruction_error, marker='o', ms=3.5, linestyle='',
                label="Fraud" if name == 1 else "Normal")
 
    ax.hlines(threshold, ax.get_xlim()[0], ax.get_xlim()[1], colors="r", zorder=100, label='Threshold')
    ax.legend()
    plt.title("Reconstruction error for different classes")
    plt.ylabel("Reconstruction error")
    plt.xlabel("Data point index")
    plt.show()
 
def visualize_reconstruction_error(reconstruction_error, threshold):
    plt.plot(reconstruction_error, marker='o', ms=3.5, linestyle='',
             label='Point')
 
    plt.hlines(threshold, xmin=0, xmax=len(reconstruction_error)-1, colors="r", zorder=100, label='Threshold')
    plt.legend()
    plt.title("Reconstruction error")
    plt.ylabel("Reconstruction error")
    plt.xlabel("Data point index")
    plt.show()
 
def preprocess_data(csv_data):
    credit_card_data = csv_data.drop(labels=['Class', 'Time'], axis=1)
    credit_card_data['Amount'] = StandardScaler().fit_transform(credit_card_data['Amount'].values.reshape(-1, 1))
    # print(credit_card_data.head())
    credit_card_np_data = credit_card_data.as_matrix()
    y_true = csv_data['Class'].as_matrix()
    return credit_card_np_data, y_true
 
# encoder
def encode(input_x, encoder_hidden_dim, latent_dim):
    """
    # keras
# build encoder model
inputs = Input(shape=input_shape, name='encoder_input')
x = Dense(intermediate_dim, activation='relu')(inputs)
z_mean = Dense(latent_dim, name='z_mean')(x)
z_log_var = Dense(latent_dim, name='z_log_var')(x)
    """
    encoder = tflearn.fully_connected(input_x, encoder_hidden_dim, activation='relu')
    mu_encoder = tflearn.fully_connected(encoder, latent_dim, activation='linear')
    logvar_encoder = tflearn.fully_connected(encoder, latent_dim, activation='linear')
    return mu_encoder, logvar_encoder
 
# decoder
def decode(z, decoder_hidden_dim, input_dim):
    """
# build decoder model
latent_inputs = Input(shape=(latent_dim,), name='z_sampling')
x = Dense(intermediate_dim, activation='relu')(latent_inputs)
outputs = Dense(original_dim, activation='sigmoid')(x)
    """
    decoder = tflearn.fully_connected(z, decoder_hidden_dim, activation='relu')
    x_hat = tflearn.fully_connected(decoder, input_dim, activation='linear')
    return x_hat
 
# sampler
def sample(mu, logvar):
    """
    keras
    z = Lambda(sampling, output_shape=(latent_dim,), name='z')([z_mean, z_log_var])
    # reparameterization trick
# instead of sampling from Q(z|X), sample eps = N(0,I)
# z = z_mean + sqrt(var)*eps
def sampling(args):
    z_mean, z_log_var = args
    batch = K.shape(z_mean)[0]
    dim = K.int_shape(z_mean)[1]
    # by default, random_normal has mean=0 and std=1.0
    epsilon = K.random_normal(shape=(batch, dim))
    return z_mean + K.exp(0.5 * z_log_var) * epsilon
    """
    epsilon = tf.random_normal(tf.shape(logvar), dtype=tf.float32, name='epsilon')
    # std_encoder = tf.exp(tf.mul(0.5, logvar))
    # z = tf.add(mu, tf.mul(std_encoder, epsilon))
    z = mu + tf.exp(logvar/2) * epsilon
    return z
 
# loss function(regularization)
def calculate_regularization_loss(mu, logvar):
    kl_divergence = -0.5 * tf.reduce_sum(1 + logvar - tf.square(mu) - tf.exp(logvar), reduction_indices=1)
    return kl_divergence
 
# loss function(reconstruction)
def calculate_reconstruction_loss(x_hat, input_x):
    mse = tflearn.objectives.mean_square(x_hat, input_x)
    return mse
 
def main():
    seed = 42
    np.random.seed(seed)
 
    data_dir_path = './data'
    model_dir_path = './models'
 
    unzip(data_dir_path + '/creditcardfraud.zip', data_dir_path)
    csv_data = pd.read_csv(data_dir_path + '/creditcard.csv')
    estimated_negative_sample_ratio = 1 - csv_data['Class'].sum() / csv_data['Class'].count()
    print(estimated_negative_sample_ratio)
    X, Y = preprocess_data(csv_data)
    print("sample data: X:{} Y:{}".format(X[:3], Y[:3]))
    print(X.shape)
 
    # detect anomaly for the test data
    Ypred = []
    _, testX, _, testY = train_test_split(X, Y, test_size=0.2, random_state=seed)
 
    blackY_indices = np.where(Y)[0]
    print(blackY_indices[:3], "sample fraud credit data")
    assert Y[blackY_indices[0]]
    assert Y[blackY_indices[-1]]
 
    # X, Y, testX, testY = mnist.load_data(one_hot=True)
 
    # Params
    original_dim = len(X[0]) # MNIST images are 28x28 pixels
    print("dim: {}".format(original_dim))
 
    """
    # Building the encoder
    encoder = tflearn.input_data(shape=[None, original_dim])
    encoder = tflearn.fully_connected(encoder, 8)
    encoder = tflearn.fully_connected(encoder, 4)
 
    # Building the decoder
    decoder = tflearn.fully_connected(encoder, 8)
    decoder = tflearn.fully_connected(decoder, original_dim, activation='sigmoid')
 
    # Regression, with mean square error
    net = tflearn.regression(decoder, optimizer='adam', learning_rate=0.001,
                             loss='mean_square', metric=None)
 
    # Training the auto encoder
    training_model = tflearn.DNN(net, tensorboard_verbose=0)
    training_model.fit(X, X, n_epoch=100, validation_set=(testX, testX),
              run_id="auto_encoder", batch_size=256)
 
    """
    hidden_dim = 8 #original_dim//2
    latent_dim = 4
    input_x = tflearn.input_data(shape=(None, original_dim), name='input_x')
    mu, logvar = encode(input_x, hidden_dim, latent_dim)
    z = sample(mu, logvar)
    x_hat = decode(z, hidden_dim, original_dim)
 
    regularization_loss = calculate_regularization_loss(mu, logvar)
    reconstruction_loss = calculate_reconstruction_loss(x_hat, input_x)
    target = tf.reduce_mean(tf.add(regularization_loss, reconstruction_loss))
 
    net = tflearn.regression(x_hat, optimizer='rmsprop', learning_rate=0.001,
                             loss=target, metric=None, name='target_out')
 
    # We will need 2 models, one for training that will learn the latent
    # representation, and one that can take random normal noise as input and
    # use the decoder part of the network to generate an image
 
    # Train the VAE
    training_model = tflearn.DNN(net, tensorboard_verbose=0)
    training_model.fit({'input_x': X}, {'target_out': X}, n_epoch=30,
                       validation_set=(testX, testX), batch_size=256, run_id="vae")
 
    """
    # Build an image generator (re-using the decoding layers)
    # Input data is a normal (gaussian) random distribution (with dim = latent_dim)
    # input_noise = tflearn.input_data(shape=[None, latent_dim], name='input_noise')
    # decoder = tflearn.fully_connected(input_noise, hidden_dim, activation='relu',
    #                                   scope='decoder_h', reuse=True)
    # decoder = tflearn.fully_connected(decoder, original_dim, activation='sigmoid',
    #                                   scope='decoder_out', reuse=True)
    # just for generate new data
    # generator_model = tflearn.DNN(decoder, session=training_model.session)
    """
 
    print("training sample predict:")
    print(training_model.predict(X[:3]))
 
    # pred_x_test = training_model.predict(testX)
 
    reconstruction_error = []
    anomaly_information,adjusted_threshold = get_anomaly(training_model, X, estimated_negative_sample_ratio)
    tp = fp = tn = fn = 0
    blackY_indices = set(blackY_indices)
    for idx, (is_anomaly, dist) in enumerate(anomaly_information):
        predicted_label = 1 if is_anomaly else 0
        if is_anomaly:
            if idx in blackY_indices:
                tp += 1
            else:
                fp += 1
        else:
            if idx in blackY_indices:
                fn += 1
            else:
                tn += 1
        Ypred.append(predicted_label)
        reconstruction_error.append(dist)
 
    print("blackY_indices len:{} detectd cnt:{}, true attack cnt:{}".format(len(blackY_indices), tp+fn, tp))
    precision = float(tp) / (tp + fp)
    hit_rate = float(tp) / (tp + fn)
    accuracy = float(tp + tn) / (tp + tn + fp + fn)
    print('precision = {}, hit_rate = {}, accuracy = {}'.format(precision, hit_rate, accuracy))
 
    report_evaluation_metrics(Y, Ypred)
    # plot_training_history(history)
    visualize_anomaly(Y, reconstruction_error, adjusted_threshold)
    plot_confusion_matrix(Y, Ypred)
 
def get_anomaly(model, data, estimated_negative_sample_ratio):
    target_data = model.predict(data)
    scores = np.linalg.norm(data - target_data, axis=-1)
    scores2 = np.array(scores)
    """
    np.linalg.norm(np.array([[1,1,1],[2,2,2]])-np.array([[0,0,0],[0,0,0]]),axis=-1)
    array([1.73205081, 3.46410162])
    >>> 3.46*3.46
    11.9716
    """
    scores.sort()
    cut_point = int(estimated_negative_sample_ratio * len(scores))
    threshold = scores[cut_point]
    print('estimated threshold is ' + str(threshold))
    return zip(scores2 >= threshold, scores2), threshold
 
if __name__ == '__main__':
    main()