Synthetic Facial Image Classifier

About This Project

Part of the group research project "Detection of Synthetic Voice, Video, Images and Text in Social Media Content" — a multi-module system tackling deepfake detection across all major media types. My individual contribution was the Synthetic Facial Image Detection module. Developed a neural network-based system that detects and classifies synthetic facial images across three major types of facial synthesis: entire face synthesis, face swapping, and face morphing. Created SynFace — a balanced dataset of 12,000 images combining real and synthesized facial data from multiple established sources to address the lack of diverse training data in existing approaches. Designed SFIC (Synthetic Facial Image Classifier) — a Siamese Neural Network architecture that leverages pairwise comparison of facial landmarks (e.g., ocular symmetry, corneal highlight shape) to detect physiological inconsistencies introduced by synthesis. The model not only identifies whether a facial image is synthetic but also determines the specific synthesis method used. SFIC outperformed established architectures including Xception, ResNet50, and EfficientNet-B0, achieving 74.13% accuracy and 73.83% F1-score on multi-class classification — demonstrating improved generalizability over single-type detection approaches.