IRIS-GAN: Staged Specialist Detection of Deepfake Faces

View on arXiv ← Back to list

Authors: Jaume M. Trenchs, Veronica Sanz

Published: 2026-06-03 13:29:24+00:00

Comment: 20 pages, 10 figures

AI Summary

The paper introduces IRIS-GAN, a specialist forensic detector for synthetic face images generated by GANs, designed to handle cross-generator shift. It employs a staged training protocol that progressively exposes the detector to increasingly complex GAN families while retaining earlier ones. This strategy achieves high fake-detection rates (above 99%) across considered GAN families and demonstrates improved generalization to unseen GANs compared to one-shot training.

Abstract

We introduce IRIS-GAN, a specialist forensic detector for synthetic face images under cross-generator shift. Rather than addressing universal synthetic-image detection, we focus on faces generated by generative adversarial networks (GANs), which are state-of-the-art in deepfake content, and train the detector through staged exposure to increasingly demanding GAN families while retaining earlier generators. The final model reaches fake-detection rates above 99% across the GAN families considered and classifies an external real-face dataset with 98.9% accuracy. Grad-CAM analysis further reveals measurable generator-dependent spatial response patterns, which remain informative for a secondary heatmap-only classifier. Out-of-family tests on diffusion-generated faces confirm that IRIS-GAN is a specialist detector, with some capability to reach non-GAN deepfakes. These results establish staged training as an effective strategy for robust GAN-face forensics.

Key findings
The IRIS-GAN detector achieved fake-detection rates above 99% across the considered GAN families and classified an external real-face dataset with 98.9% accuracy. Staged training significantly improved generalization to unseen GAN families compared to a one-shot approach. Grad-CAM analysis revealed generator-dependent spatial response patterns, which remained informative for a secondary heatmap-only classifier, though the detector showed limited capability to generalize to non-GAN deepfakes like diffusion-generated faces.
Approach
The proposed approach is a staged specialist detector based on ConvNeXt-Large. It is trained by progressively introducing new, more challenging GAN families (ProGAN, StyleGAN2, StyleGAN3, StyleGANXL, EG3D) while retaining earlier ones in the training set. This curriculum is designed to reduce overspecialization and improve transferability to unseen GAN generators. A secondary ResNet18 classifier is also trained on Grad-CAM heatmaps for diagnostic analysis.
Datasets
FFHQ, CelebA, ProGAN, StyleGAN2, StyleGAN3, StyleGANXL, EG3D, Stable Diffusion 1.5, FLUX.1-dev, FLUX.1-schnell, FLUX.1-pro, Stable Diffusion XL, DALL-E 3 (from SFHQ-T2I collection).
Model(s)
ConvNeXt-Large, ResNet18, Stable Diffusion VAE (for reconstruction-based experiments), custom real-only VAE (for reconstruction-based experiments).
Author countries
Spain
← Previous
Next →