稳健深度伪造检测，NTIRE 2026 挑战报告
计算机视觉与模式识别
2026-04-28 v1
摘要
深度伪造检测的鲁棒性长期未被充分关注。然而，如果检测性能在面对轻微图像退化时受到影响，其实际价值几乎为零。在除常见退化外，还有一种风险，即恶意深度伪造会刻意引入退化，釣然利用检测器在此方面的弱点。本文概述了NTIRE 2026稳健深度伪造检测挑战，特别针对该问题进行了设计算。参赛者被要求构建一个检测器，其后将在包含各种强度的常见和不常见退化的未知测试集上进行测试。本次挑战共有337名参赛者提交57个最终排行榜解答，首届挑战获得了良好反响。为确保结果可靠性，参赛者仅在24小时内完成测试运行，且未提供标签，从而限制了对测试数据的训练可能性。此外，前三名解答在私人测试集上进行评分，以检测任何可能的过拟合现象。本报告介绍了竞赛设置、数据集准备以及方法的细节与性能。顶级方法依赖于大型基础模型、集成和退化训练，以实现通用性和鲁棒性的结合。
引用
@article{arxiv.2604.24163,
title = {Robust Deepfake Detection, NTIRE 2026 Challenge: Report},
author = {Benedikt Hopf and Radu Timofte and Chenfan Qu and Junchi Li and Fei Wu and Dagong Lu and Mufeng Yao and Xinlei Xu and Fengjun Guo and Yongwei Tang and Zhiqiang Yang and Zhiqiang Wu and Jia Wen Seow and Hong Vin Koay and Haodong Ren and Feng Xu and Shuai Chen and Minh-Khoa Le-Phan and Minh-Hoang Le and Trong-Le Do and Minh-Triet Tran and Chih-Yu Jian and Yi-Fan Wang and Bang-Kang Chen and You-Chen Chao and Chia-Ming Lee and Fu-En Yang and Yu-Chiang Frank Wang and Chih-Chung Hsu and Aashish Negi and Hardik Sharma and Prateek Shaily and Jayant Kumar and Sachin Chaudhary and Akshay Dudhane and Praful Hambarde and Amit Shukla and Jielun Peng and Yabin Wang and Yaqi Li and Jincheng Liu and Xiaopeng Hong and Krish Wadhwani and Liam Fitzpatrick and Utkarsh Tiwari and Bilel Benjdira and Anas M. Ali and Wadii Boulila and Cristian Lazo Quispe and Aishwarya A and Akshara S and Ashwathi N and Jiachen Tu and Guoyi Xu and Yaoxin Jiang and Jiajia Liu and Yaokun Shi},
journal= {arXiv preprint arXiv:2604.24163},
year = {2026}
}