ML Security PapersNTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild
Aleksandr Gushchin 1, Khaled Abud 1, Ekaterina Shumitskaya 1, Artem Filippov 1, Georgii Bychkov 1, Sergey Lavrushkin 1, Mikhail Erofeev 1, Anastasia Antsiferova 1, Changsheng Chen 2, Shunquan Tan 2, Radu Timofte 3, Dmitry Vatolin 1, Chuanbiao Song 4, Zijian Yu 4, Hao Tan 5, Jun Lan 5, Zhiqiang Yang 5, Yongwei Tang 5, Zhiqiang Wu 6, Jia Wen Seow , Hong Vin Koay , Haodong Ren 7, Feng Xu 7, Shuai Chen 7, Ruiyang Xia 8, Qi Zhang 8, Yaowen Xu 8, Zhaofan Zou 9, Hao Sun 9, Dagong Lu 9, Mufeng Yao 9, Xinlei Xu 9, Fei Wu 10, Fengjun Guo 10, Cong Luo 10, Hardik Sharma 11, Aashish Negi 11, Prateek Shaily 11, Jayant Kumar 11, Sachin Chaudhary 12, Akshay Dudhane 12, Praful Hambarde 12, Amit Shukla 12, Zhilin Tu 13, Fengpeng Li 13, Jiamin Zhang 13, Jianwei Fei 13, Kemou Li 14, Haiwei Wu 14, Bilel Benjdira 15, Anas M. Ali 15, Wadii Boulila 15, Chenfan Qu 16, Junchi Li 16
1 Lomonosov Moscow State University
7 Shanghai Jiao Tong University
8 University of Science and Technology of China
9 Harbin Institute of Technology
11 Indian Institute of Technology Jodhpur
12 Visvesvaraya National Institute of Technology
Published on arXiv
2604.11487
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
20 teams submitted robust detection solutions evaluated across 42 generators and 36 image transformations representing real-world deployment scenarios
This paper presents an overview of the NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild, held in conjunction with the NTIRE workshop at CVPR 2026. The goal of this challenge was to develop detection models capable of distinguishing real images from generated ones in realistic scenarios: the images are often transformed (cropped, resized, compressed, blurred) for practical usage, and therefore, the detection models should be robust to such transformations. The challenge is based on a novel dataset consisting of 108,750 real and 185,750 AI-generated images from 42 generators comprising a large variety of open-source and closed-source models of various architectures, augmented with 36 image transformations. Methods were evaluated using ROC AUC on the full test set, including both transformed and untransformed images. A total of 511 participants registered, with 20 teams submitting valid final solutions. This report provides a comprehensive overview of the challenge, describes the proposed solutions, and can be used as a valuable reference for researchers and practitioners in increasing the robustness of the detection models to real-world transformations.
Key Contributions
- Novel dataset of 294,500 images (108,750 real, 185,750 AI-generated) from 42 diverse generators with 36 realistic transformations
- Benchmark challenge with 511 participants and 20 validated detection solutions evaluated on robustness to real-world image transformations
- Comprehensive evaluation framework using ROC AUC on mixed transformed/untransformed test set for realistic detection scenarios
🛡️ Threat Analysis
Challenge focuses on AI-generated image detection (distinguishing real from synthetic images) and robustness to common transformations (compression, resizing, cropping) that could evade detection — core ML09 output integrity and content provenance task.