DeepfakeGenome: Toward Next-Generation Deepfake Attribution

Authors: Shunli Wang, Xinyu Zhou, Yandan Zhao, Wenbin Liang, Taiping Yao, Ke-Yue Zhang, Zhiyuan Yan, Shouhong Ding, Lizhuang Ma

Published: 2026-11-04 06:48:44+00:00

Comment: 28 Pages

AI Summary

This paper introduces DeepfakeGenome (DFG), a comprehensive benchmark for deepfake detection and attribution, addressing limitations in existing datasets and evaluation protocols. DFG comprises 100 facial forgery algorithms and 2 million images, significantly larger than previous benchmarks. It also proposes novel retrieval-based attribution protocols to better align with real-world active defense scenarios, leading to extensive experimental evaluations and key insights into deepfake attribution research.

Abstract

In recent years, AIGC technologies are capable of generating hyper-realistic forged facial images, posing severe threats to facial security safety. To tackle these challenges, early Deepfake research primarily centered on real-fake classification. Recently, growing efforts have been devoted to the Deepfake Attribution (DFA) of generated content. Nevertheless, existing research on Deepfake detection and attribution faces saturated performance in binary classification, limited diversity in datasets and algorithms, and imperfect evaluation protocols, which severely impede practical application. To address these limitations, we propose a comprehensive deepfake detection and attribution benchmark named DeepfakeGenome (DFG). It contains 100 facial forgery algorithms and 2M images in total, achieving 4x to 100x larger than prior DFA benchmarks. We further design 4 protocols for practical evaluation, including a novel retrieval-based attribution paradigm. Unlike previous open-set evaluation metrics, the proposed retrieval metrics are more aligned with the real-world active defense situation of blacklist registration mechanisms. Based on these elaborate designs, we investigate the performance ceiling of deepfake attribution task. Over 2k experimental evaluations are conducted, and 10 insightful findings are derived. We believe this work can provide new insights into the DFA research field. The Code and Benchmark of DFG are avaliable on https://huggingface.co/spaces/DeepfakeGenome/DFG_Page.


Key findings
The study found that traditional deepfake training settings may not be suitable for evaluating emerging algorithms, as clip-style models show significant performance deviation in sub-class algorithms like Entire Face Synthesis. Latent space augmentation and representation learning strategies can significantly improve generalization on new attacks. DINOv2 performed best in attribution, while CLIP excelled in Out-of-Distribution (OOD) retrieval, indicating that CLIP generalizes well in OOD settings while DINOv2 collapses, and that low-level and high-level features are complementary for OOD generalization.
Approach
The authors developed a large-scale benchmark (DeepfakeGenome) by integrating and deduplicating existing deepfake datasets and incorporating recently released open-source and closed-source image forgery models. They designed four evaluation protocols, including a novel retrieval-based attribution paradigm, to comprehensively assess model performance across various tasks like real-fake classification and deepfake attribution under different data partitioning strategies.
Datasets
DeepfakeGenome (DFG), FF++, DF40, ForgeryNet, DFFD, DNA-Det, OSMA, WILD-20, DiFF, DiffusionFace, CelebA
Model(s)
Xception, Eff-B4, CLIP-ViT-FFT, Effort, RECCE, SBI, F3Net, SRM, SPSL, ProDet, LSDA, UCF, IID, CORE, sladd, ResNet-34, LoRAX, OCC-CLIP, NPR, POSE, HIFI-IFD, Repmix, GAN Attribution, DCT, DNA-DET, Universal, DINOv2-ViT, DINOv3-ViT
Author countries
UNKNOWN