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

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

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

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：

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

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

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()