📄 Abstract
Generative adversarial networks (GANs) can synthesize highly realistic fictional faces, creating new challenges for digital image authentication and source attribution. Forensic examination requires more than a binary real-or-fake decision: relevant traces should be reproducible, mechanistically interpretable, and evaluated under clearly defined conditions. This structured narrative review updates the literature through October 2026 and groups forensic traces in GAN-generated face images into three mechanism-based categories: differences between physical camera imaging and neural synthesis pipelines; traces introduced by generator architecture; and anomalies in physiological, geometric, and physical consistency. The review focuses on color and noise statistics, texture, upsampling and spatial-frequency structure, generator fingerprints, StyleGAN-specific artifacts, and ocular geometry, with particular attention to whether these traces persist as generator architectures evolve. Evidence shows that StyleGAN's blob-like and progressive-growing-related phase artifacts were specifically addressed in StyleGAN2, while texture sticking was substantially reduced by the alias-free design of StyleGAN3. High-frequency spectral-falloff discrepancies can also disappear after changes to the upsampling architecture and therefore are not intrinsic properties of CNN-based generators. Frequency-domain, autocorrelation, and ocular geometric differences continue to be reported in newer models, but their expression depends on the generator, training data, image quality, and post-processing history. The forensic significance of these findings is discussed in terms of trace persistence, alternative explanations, processing history, and independent corroboration. No single anomaly should be used in isolation to support a conclusion of AI generation or a specific source.
🏷️ Keywords
📚 How to Cite:
Yi Wang , FORENSIC TRACES IN GAN-GENERATED FACE IMAGES: FORMATION MECHANISMS, EVOLUTION, AND FORENSIC INTERPRETATION , Volume 12 , Issue 10, October 2026, EPRA International Journal of Multidisciplinary Research (IJMR) , Pages: 160 - 167 , DOI: https://doi.org/10.36713/epra31768