ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization

Authors: Bo Du, Xuekang Zhu, Xiaochen Ma, Chenfan Qu, Kaiwen Feng, Zhe Yang, Chi-Man Pun, Jian Liu, Jizhe Zhou

Published: 2025-05-16 08:49:59+00:00

Comment: NeurIPS 2025 DB Track Paper. Code available at: https://github.com/scu-zjz/ForensicHub

AI Summary

ForensicHub introduces the first unified benchmark and codebase to address the fragmentation in Fake Image Detection and Localization (FIDL) across four domains: deepfake, image manipulation, AI-generated image, and document manipulation. It proposes a modular, configuration-driven architecture that integrates existing benchmarks while establishing new ones for AIGC and Document domains. The platform facilitates extensive cross-domain evaluations and provides key actionable insights to advance research in the FIDL field.

Abstract

The field of Fake Image Detection and Localization (FIDL) is highly fragmented, encompassing four domains: deepfake detection (Deepfake), image manipulation detection and localization (IMDL), artificial intelligence-generated image detection (AIGC), and document image manipulation localization (Doc). Although individual benchmarks exist in some domains, a unified benchmark for all domains in FIDL remains blank. The absence of a unified benchmark results in significant domain silos, where each domain independently constructs its datasets, models, and evaluation protocols without interoperability, preventing cross-domain comparisons and hindering the development of the entire FIDL field. To close the domain silo barrier, we propose ForensicHub, the first unified benchmark & codebase for all-domain fake image detection and localization. Considering drastic variations on dataset, model, and evaluation configurations across all domains, as well as the scarcity of open-sourced baseline models and the lack of individual benchmarks in some domains, ForensicHub: i) proposes a modular and configuration-driven architecture that decomposes forensic pipelines into interchangeable components across datasets, transforms, models, and evaluators, allowing flexible composition across all domains; ii) fully implements 10 baseline models, 6 backbones, 2 new benchmarks for AIGC and Doc, and integrates 2 existing benchmarks of DeepfakeBench and IMDLBenCo through an adapter-based design; iii) conducts indepth analysis based on the ForensicHub, offering 8 key actionable insights into FIDL model architecture, dataset characteristics, and evaluation standards. ForensicHub represents a significant leap forward in breaking the domain silos in the FIDL field and inspiring future breakthroughs.

Key findings

Less-explored visual backbones like ConvNeXt and Swin Transformer surprisingly outperform most domain-specific SoTA methods when trained on unified fake images under the IFF-Protocol, indicating their strong generalization potential. Shallow feature extractors generally negatively impact performance on large, diverse datasets, though lightweight models like EfficientNet can still benefit. Additionally, current AIGC and Deepfake evaluations often neglect generalization, highlighting the need for more complex, realistic datasets and generalization-aware model designs.

Approach

The authors propose ForensicHub, a modular and configuration-driven architecture that decomposes forensic pipelines into interchangeable components (datasets, transforms, models, evaluators) to support all four FIDL domains. It implements 10 baseline models and 6 backbones, introduces 2 new benchmarks for AIGC and Document, and integrates 2 existing benchmarks (DeepfakeBench, IMDLBenCo) using an adapter-based design. This unified framework enables flexible composition and comprehensive cross-domain evaluation.

Datasets

FaceForensics++, Celeb-DF, DFD, FaceShifter, UADFV, DeepFakeDetection, DFDCP (for Deepfake); CASIA, COVERAGE, Columbia, IMD2020, NIST16, CocoGlide, Autosplice (for IMDL); DiffusionForensics, GenImage (for AIGC); Doctamper, T-SROIE, OSTF, TPIC-13, RTM, Tampered-IC13 (for Document).

Model(s)

Capsule-Net, RECCE, SPSL, UCF, SBI (Deepfake); MVSS-Net, CAT-Net, PSCC-Net, Trufor, IML-ViT, Mesorch (IMDL); Dire, DualNet, HiFiNet, Synthbuster, UnivFD (AIGC); DTD, FFDN, CAFTB, TIFDM (Document); Resnet, Xception, EfficientNet, Segformer, Swin Transformer, ConvNext (Backbones).

Author countries

China, Macao

← Previous