REAL VS. FAKE FACE CLASSIFICATION USING DEEP LEARNING-BASED CONVOLUTIONAL NEURAL NETWORKS

Abstract

The widespread use of artificial intelligence (AI) has accelerated the creation of hyperrealistic synthetic images, commonly referred to as deepfakes. These AI-generated faces present growing challenges in domains such as security, media verification, and digital identity. This study presents a primary research effort that investigates the classification of real and AI-generated facial images using a deep learning-based convolutional neural network (CNN). Unlike previous works that solely rely on existing public datasets, this study integrates a curated set of AI-generated and real faces from Kaggle with a custom-collected image set to enhance dataset diversity and realism. The experimental CNN architecture was developed using the Keras Sequential API and consists of four convolutional layers with ReLU activation, four max-pooling layers, one batch normalization layer, dropout regularization, and two fully connected dense layers. The training process incorporated data augmentation techniques such as rescaling, shearing, zooming, and horizontal flipping to improve generalization and mitigate overfitting. The model was compiled using the Adam optimizer and binary cross-entropy loss, with accuracy used as the primary metric. The model was trained for 25 epochs, achieving a test accuracy of approximately 82% on the final evaluation. Performance metrics indicate high precision (83.08%) and recall (82.44%), as reflected in the classification report and confusion matrix. These results confirm the model's effectiveness in accurately differentiating between real and AI-generated faces. Model performance was monitored through training and validation curves, revealing consistent accuracy gains and manageable loss fluctuations. In addition to model training, feature map visualizations were generated to illustrate how the CNN processes facial structures across layers. The findings suggest that intermediate convolutional outputs successfully capture critical features like facial symmetry, edge clarity, and texture anomalies, which are typically absent or uniform in synthetic faces. The inclusion of a custom dataset and extended experimental controls, such as validation monitoring and architectural tuning, elevate this work from a secondary to a primary research contribution. This study highlights the practical application of CNNs in deepfake detection and proposes future extensions to real-time video classification and architectural comparisons involving ensemble or transfer learning models.