[Paper Review] Eyes Tell All: Irregular Pupil Shapes Reveal GAN-generated Faces

Hui Guo, Shu Hu+3 more|arXiv (Cornell University)|Sep 1, 2021

Generative Adversarial Networks and Image Synthesis4 citations

TL;DR

This paper proposes a novel, interpretable method to detect GAN-generated faces by exploiting irregular pupil shapes—caused by the lack of physiological constraints in GAN training—using automatic pupil segmentation and ellipse fitting to compute a Boundary IoU (BIoU) score. The method achieves 91% AUC on real vs. GAN-generated face detection, demonstrating high effectiveness and interpretability.

ABSTRACT

Generative adversary network (GAN) generated high-realistic human faces have been used as profile images for fake social media accounts and are visually challenging to discern from real ones. In this work, we show that GAN-generated faces can be exposed via irregular pupil shapes. This phenomenon is caused by the lack of physiological constraints in the GAN models. We demonstrate that such artifacts exist widely in high-quality GAN-generated faces and further describe an automatic method to extract the pupils from two eyes and analysis their shapes for exposing the GAN-generated faces. Qualitative and quantitative evaluations of our method suggest its simplicity and effectiveness in distinguishing GAN-generated faces.

Motivation & Objective

• Address the growing problem of GAN-generated fake faces used in malicious social media profiles.
• Overcome limitations of existing deep learning-based detection methods, which lack interpretability and generalization across GAN architectures.
• Exploit inherent physiological inconsistencies in GAN-generated faces, specifically irregular pupil shapes, as a forensic cue.
• Develop a method that is both effective and interpretable by leveraging the natural geometric regularity of human pupils.

Proposed method

• Automatically segment pupil regions in face images using a deep learning-based segmentation model.
• Fit an ellipse parametric model to each segmented pupil boundary to approximate ideal physiological shape.
• Compute the Boundary Intersection-over-Union (BIoU) score between the segmented pupil mask and the fitted ellipse as a measure of shape regularity.
• Use the average BIoU score across both eyes as a forensic feature to classify faces as real or GAN-generated.
• Apply a distance parameter $d$ to control sensitivity to boundary deviations in the BIoU calculation.
• Train and evaluate the method on real faces from the FFHQ dataset and GAN-generated faces from StyleGAN2.

Experimental results

Research questions

• RQ1 Can irregular pupil shapes in GAN-generated faces serve as a reliable forensic cue for detection?
• RQ2 How does the shape regularity of pupils differ between real human faces and GAN-generated faces?
• RQ3 Can a simple, interpretable method based on pupil geometry outperform black-box deep learning detectors in generalization and explainability?
• RQ4 To what extent does the BIoU score correlate with the authenticity of a face image?

Key findings

• The BIoU score distribution for real faces is significantly higher and clearly separated from that of GAN-generated faces, indicating strong discriminative power.
• The method achieves an Area Under the ROC Curve (AUC) of 0.91 on the combined FFHQ and StyleGAN2 dataset, demonstrating high detection accuracy.
• The optimal sensitivity to boundary irregularities is achieved at a distance parameter $d=4$ in the BIoU calculation, balancing detection performance and robustness.
• False positives can occur due to pathological pupil shapes (e.g., infections) or imaging artifacts (e.g., occlusions, lighting), though these are rare in real-world contexts.
• The method is computationally efficient and interpretable, as the BIoU score directly reflects the geometric regularity of the pupil.
• Visual inspection alone can often distinguish GAN-generated faces by observing irregular pupil shapes, confirming the method’s intuitive basis.