Decoding Synthetic Face Detectors: Enhancing Interpretability with 3D Morphable Models

The rise of synthetic face generation and deepfakes has introduced significant challenges in multimedia forensics, particularly in ensuring the authenticity of visual content. While many detectors exist to identify these forgeries, they often operate as black boxes, providing little insight into the decision-making process. In this paper, we propose a novel method for improving the interpretability of synthetic face detectors. Our approach leverages 3D Morphable Models (3DMMs) to analyze and interpret the output of a heatmap-based detector, identifying the most important facial regions for detection. Specifically, we utilize 3DMMs to reverse-engineer the detector by averaging heatmaps from multiple images warped onto a common face geometry, revealing which areas the detector focuses on most. We validate our proposed approach by testing a state-of-the-art synthetic image detector on a dataset of real and synthetic faces generated using various Stable Diffusion (SD) models, offering insights into its behavior. Our experiments provide a valuable understanding of the internal workings of synthetic face detection, contributing to the growing need for interpretable and trustworthy forensic tools in the fight against synthetic media. Our experimental code is available at https://github.com/polimiispl/synthetic_image_interpretability.