Real vs. Fake Face Classification Using Convolutional Neural Networks and TF-IDF
Abstract
This proposes a hybrid approach in real and fake face classification with the Convolutional Neural Networks (CNN) and TF-IDF feature weighting is introduced. The training data is 6,799 and the validation data is 2,278. The CNN model derives hierarchical features of space via several convolutional layers with ReLU activation functions and max-pooling functions. They use data augmentation methods to improve generalization. Besides the deep feature extraction, there is the use of TF-IDF (Term Frequency-Inverse Document Frequency) that is used to statistically weight the discriminative feature descriptors based on the image patterns, which focus on minor artifacts that are normally found in manipulated faces. The model is trained on batch size 64 and learning rate 0.001 taking 10 epochs. The experiment in question yields an accuracy of validation equal to 98.5% and a final training loss of 0.05, which is indicative of a successful convergence and a high discriminative power. The CNN and TF-IDF hybrid method enhances the resilience of the real and fake face detection.