The proliferation of AI-generated synthetic faces has raised significant concerns about media authenticity, identity theft, and misinformation. Detecting such fake faces reliably is critical for securing biometric systems and restoring public trust in digital content. This study investigates the effectiveness of transfer learning using the Xception model—a deep convolutional neural network originally trained on ImageNet—for binary classification of real versus AI-generated face images.

We combined and preprocessed two publicly available datasets, resulting in a balanced corpus of authentic and synthetic face images. The data was resized to 299×299 pixels, normalized, and split into training (70%), validation (20%), and test (10%) sets. A fully unfrozen Xception model was fine-tuned using an optimized architecture and trained over 30 epochs with the Adam optimizer and binary cross-entropy loss. Performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrix, along with qualitative analysis through prediction visualizations.

The fine-tuned model achieved nearly 100% training accuracy and 90% validation accuracy. On the unseen test set of 1,205 images, it attained 89.88% accuracy and an F1-score of 0.90, indicating high reliability across both real and fake face classes. The model performed slightly better at identifying synthetic faces, highlighting detectable artifacts introduced during generation.

Our findings confirm that transfer learning with the Xception model is a practical, reproducible solution for fake face detection, even in resource-limited academic settings. This study contributes a streamlined pipeline and benchmarks for future work in visual deepfake detection and media forensics.