This notebook implements a Convolutional Neural Network (CNN) based on the MesoNet architecture to classify images as real or deepfake. The workflow includes data loading, model definition, training, and prediction visualization. TensorFlow and Keras are used to build and train the model, while Matplotlib is used for result visualization.
# Import packages
import numpy as np
import matplotlib.pyplot as plt
import os
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Conv2D, BatchNormalization, MaxPooling2D, Flatten, Dense, Dropout
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.preprocessing.image import load_img, img_to_array
# Image dimensions
im_dim = {"height": 256, "width": 256, "channels": 3}
# Create a classifier class
class Classifier:
def __init__(self):
self.model = None
def predict(self, x):
return self.model.predict(x)
def fit(self, x, y, epochs=10, batch_size=32):
return self.model.fit(x, y, epochs=epochs, batch_size=batch_size)
def get_accuracy(self, x, y):
return self.model.evaluate(x, y)
def load(self, path):
self.model.load_weights(path)
# Create MesoNet class using Classifier
class Meso(Classifier):
def __init__(self, learning_rate=0.001):
super().__init__()
self.model = self.init_model()
optimizer = Adam(learning_rate=learning_rate)
self.model.compile(optimizer=optimizer, loss="binary_crossentropy", metrics=['accuracy'])
def init_model(self):
x = Input(shape=(im_dim['height'], im_dim['width'], im_dim['channels']))
x1 = Conv2D(8, (3, 3), padding='same', activation='relu')(x)
x1 = BatchNormalization()(x1)
x1 = MaxPooling2D(pool_size=(2, 2), padding='same')(x1)
x2 = Conv2D(8, (5, 5), padding='same', activation='relu')(x1)
x2 = BatchNormalization()(x2)
x2 = MaxPooling2D(pool_size=(2, 2), padding='same')(x2)
x3 = Conv2D(16, (5, 5), padding='same', activation='relu')(x2)
x3 = BatchNormalization()(x3)
x3 = MaxPooling2D(pool_size=(2, 2), padding='same')(x3)
x4 = Conv2D(8, (3, 3), padding='same', activation='relu')(x3)
x4 = BatchNormalization()(x4)
x4 = MaxPooling2D(pool_size=(2, 2), padding='same')(x4)
flat = Flatten()(x4)
dense1 = Dense(16, activation='relu')(flat)
drop = Dropout(0.5)(dense1)
output = Dense(1, activation='sigmoid')(drop)
return Model(inputs=x, outputs=output)
# Data Loading Function
def load_data(directory):
images, labels = [], []
for label, folder in enumerate(['real', 'fake']):
folder_path = os.path.join(directory, folder)
for filename in os.listdir(folder_path):
img_path = os.path.join(folder_path, filename)
img = load_img(img_path, target_size=(im_dim['height'], im_dim['width']))
img_array = img_to_array(img) / 255.0
images.append(img_array)
labels.append(label) # 0 for real, 1 for fake
return np.array(images), np.array(labels)
# Prediction Function
def predict_image(model, image_path):
img = load_img(image_path, target_size=(im_dim['height'], im_dim['width']))
img_array = img_to_array(img) / 255.0
img_array = np.expand_dims(img_array, axis=0)
prediction = model.predict(img_array)[0][0]
label = "Real" if prediction < 0.5 else "Deepfake"
plt.imshow(img)
plt.title(f'Prediction: {label}')
plt.axis('off')
plt.show()
# Training the Model
train_images, train_labels = load_data('train_data')
model = Meso()
print("Starting model training...")
model.fit(train_images, train_labels, epochs=10, batch_size=32)
print("Model training complete.")Starting model training...
Epoch 1/10
626/626 [==============================] - 44s 70ms/sample - loss: 0.7518 - acc: 0.7093
Epoch 2/10
626/626 [==============================] - 51s 82ms/sample - loss: 0.4196 - acc: 0.8195
Epoch 3/10
626/626 [==============================] - 49s 78ms/sample - loss: 0.3201 - acc: 0.8674
Epoch 4/10
626/626 [==============================] - 44s 71ms/sample - loss: 0.2645 - acc: 0.8978
Epoch 5/10
626/626 [==============================] - 40s 65ms/sample - loss: 0.2038 - acc: 0.9105
Epoch 6/10
626/626 [==============================] - 43s 69ms/sample - loss: 0.1459 - acc: 0.9393
Epoch 7/10
626/626 [==============================] - 42s 67ms/sample - loss: 0.1154 - acc: 0.9601
Epoch 8/10
626/626 [==============================] - 38s 61ms/sample - loss: 0.1360 - acc: 0.9505
Epoch 9/10
626/626 [==============================] - 41s 66ms/sample - loss: 0.1142 - acc: 0.9537
Epoch 10/10
626/626 [==============================] - 44s 71ms/sample - loss: 0.0923 - acc: 0.9569
Model training complete.
In this section, we test the trained model using two sample images of political figures found online. One image is real, and the other is a deepfake. The model predicts the authenticity of each image, and the results are visualized with corresponding labels.
# Testing the Model
print("\nTesting on Real Image:")
predict_image(model.model, 'test_data/real/sample.jpg')
print("Testing on Fake Image:")
predict_image(model.model, 'test_data/fake/sample.jpg')Testing on Real Image:
Testing on Fake Image:
We developed a Chrome extension that could integrate with our AI system via an API to provide confidence levels for various tasks. The extension collects data from the browser and passes it to the TensorFlow model, which processes the information and returns a confidence score indicating how likely the AI is to make an accurate prediction or decision. This setup allows for real-time feedback and helps users assess the reliability of the AI’s responses, making the system more transparent and user-friendly.
