Logo image
Back
The Safe Image Authenticity Challenge: Detecting and Localizing Partial and Fully Synthetic Manipulations
Conference paper

The Safe Image Authenticity Challenge: Detecting and Localizing Partial and Fully Synthetic Manipulations

Tai Duc Nguyen, Matthew C Stamm, Jill Crisman, Jesse Hostetler, Laura Cassano, Michael Davinroy and Peter Bautista
2026 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)
04 Aug 2026
DOI: https://doi.org/10.1109/WACVW68408.2026.00103
Featured in Collection :   Drexel's Newest Publications

Abstract

Advances in generative AI and AI-assisted image editing tools have made it increasingly easy to produce not only fully AI-generated images, but also more subtle, locally manipulated partially synthetic content that is difficult for both humans and automated systems to detect. This emerging class of mixed-authenticity imagery challenges existing multimedia forensic benchmarks, which typically evaluate only a single manipulation type. Consequently, state-ofthe-art forensic methods remain limited in their ability to operate reliably across a diverse range of fake image types and to provide insights beyond binary real-versus-fake decisions. To address this gap, we introduce the SAFE: Image Editing Detection and Localization Challenge, a new benchmark and competition designed to evaluate detection, classification, and localization across a broad spectrum of fake and manipulated images, including fully AI-generated and AI-edited content. Participating systems must output a global authenticity decision, a manipulation type classification, and a pixel-level localization mask. We support SAFE with SAFE-FORGE, a large-scale dataset generated by a scalable forensic data engine that produces diverse fake images with comprehensive ground-truth annotations and traceable manipulation histories. Preliminary evaluations using state-of-the-art forensic systems reveal substantial vulnerabilities in both detection and localization, underscoring the need for unified forensic approaches that remain robust to the rapidly evolving capabilities of modern generative and image editing models.

Metrics

1 Record Views

Details

Logo image