《Web安全之深度学习实战》笔记：第五章验证码识别

本小节实际上就是用KNN、SVM、MLP、CNN识别mnist数据集&＃xff0c;实际上这些在《web安全之机器学习入门》中都有讲过&＃xff0c;而这些也是非常基础的应用。

一、KNN

整体设计如下

 源码如下所示

def do_knn_1d(x_train, y_train,x_test, y_test):print ("KNN and 1d")clf &＃61; neighbors.KNeighborsClassifier(n_neighbors&＃61;15)print (clf)clf.fit(x_train, y_train)y_pred &＃61; clf.predict(x_test)print (metrics.accuracy_score(y_test, y_pred))

二、SVM

整体设计如下

源码如下

def do_svm_1d(x_train, y_train,x_test, y_test):print ("SVM and 1d")clf &＃61; svm.SVC(decision_function_shape&＃61;&＃39;ovo&＃39;)print (clf)clf.fit(x_train, y_train)y_pred &＃61; clf.predict(x_test)print (metrics.accuracy_score(y_test, y_pred))

三、MLP

整体设计如下

源码如下

def do_dnn_1d(X, Y, testX, testY ):print ("DNN and 1d")# Building deep neural networkinput_layer &＃61; tflearn.input_data(shape&＃61;[None, 784])dense1 &＃61; tflearn.fully_connected(input_layer, 64, activation&＃61;&＃39;tanh&＃39;,regularizer&＃61;&＃39;L2&＃39;, weight_decay&＃61;0.001)dropout1 &＃61; tflearn.dropout(dense1, 0.8)dense2 &＃61; tflearn.fully_connected(dropout1, 64, activation&＃61;&＃39;tanh&＃39;,regularizer&＃61;&＃39;L2&＃39;, weight_decay&＃61;0.001)dropout2 &＃61; tflearn.dropout(dense2, 0.8)softmax &＃61; tflearn.fully_connected(dropout2, 10, activation&＃61;&＃39;softmax&＃39;)# Regression using SGD with learning rate decay and Top-3 accuracysgd &＃61; tflearn.SGD(learning_rate&＃61;0.1, lr_decay&＃61;0.96, decay_step&＃61;1000)top_k &＃61; tflearn.metrics.Top_k(3)net &＃61; tflearn.regression(softmax, optimizer&＃61;sgd, metric&＃61;top_k,loss&＃61;&＃39;categorical_crossentropy&＃39;)# Trainingmodel &＃61; tflearn.DNN(net, tensorboard_verbose&＃61;0)model.fit(X, Y, n_epoch&＃61;10, validation_set&＃61;(testX, testY),show_metric&＃61;True, run_id&＃61;"mnist")


四、CNN

设计如下

源码如下

def do_cnn_2d(X, Y, testX, testY ):# Building convolutional networknetwork &＃61; input_data(shape&＃61;[None, 28, 28, 1], name&＃61;&＃39;input&＃39;)network &＃61; conv_2d(network, 32, 3, activation&＃61;&＃39;relu&＃39;, regularizer&＃61;"L2")network &＃61; max_pool_2d(network, 2)network &＃61; local_response_normalization(network)network &＃61; conv_2d(network, 64, 3, activation&＃61;&＃39;relu&＃39;, regularizer&＃61;"L2")network &＃61; max_pool_2d(network, 2)network &＃61; local_response_normalization(network)network &＃61; fully_connected(network, 128, activation&＃61;&＃39;tanh&＃39;)network &＃61; dropout(network, 0.8)network &＃61; fully_connected(network, 256, activation&＃61;&＃39;tanh&＃39;)network &＃61; dropout(network, 0.8)network &＃61; fully_connected(network, 10, activation&＃61;&＃39;softmax&＃39;)network &＃61; regression(network, optimizer&＃61;&＃39;adam&＃39;, learning_rate&＃61;0.01,loss&＃61;&＃39;categorical_crossentropy&＃39;, name&＃61;&＃39;target&＃39;)# Trainingmodel &＃61; tflearn.DNN(network, tensorboard_verbose&＃61;0)model.fit({&＃39;input&＃39;: X}, {&＃39;target&＃39;: Y}, n_epoch&＃61;5,validation_set&＃61;({&＃39;input&＃39;: testX}, {&＃39;target&＃39;: testY}),snapshot_step&＃61;100, show_metric&＃61;True, run_id&＃61;&＃39;mnist&＃39;)

五、总结

识别mnist数据集较基础&＃xff0c;深度学习的识别结果稍好。