Research map / Detection and segmentation / Open-set and few-shot detection
ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents ForensicHub, a unified modular benchmark and codebase covering deepfake, IMDL, AIGC, and document image forgery detection/localization, with adapters for DeepfakeBench and IMDLBenCo, two new AIGC and Doc protocols, and a cross-domain IFF-Protocol.
Problem
Fake image detection and localization is fragmented into four domains (deepfake, IMDL, AIGC, document) with incompatible datasets, models, and evaluation protocols, so no unified benchmark enables cross-domain comparison.
Approach
A modular, configuration-driven codebase decomposing forensic pipelines into interchangeable datasets, transforms, models, and evaluators, with adapters that integrate existing benchmarks and a joint multi-domain training protocol (IFF-Protocol) for cross-domain evaluation.
Key results
- Under the IFF-Protocol, ConvNeXt (average AUC 0.822) and Swin (0.824) outperform nearly all domain-specific SoTA methods, while Effort reaches 0.892 average AUC among the listed models.
- On the Doc protocol (trained on DocTamper), PSCC-Net achieves the best cross-domain average F1 of 0.4083 on external document sets.
- Adding shallow low-level feature extractors (Sobel, Bayar, DCT, FPH) lowers mean AUC for most backbones, with EfficientNet as the exception that benefits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with IFF-Protocol's combined cross-domain data rather than task-specific training | visual backbones and domain-specific detectors · four-domain benchmark | AUC | Helped |
| add shallow feature extractors to backbones | backbones under IFF-Protocol · four forensic domains | AUC | Hurt |
| add shallow feature extractors (ablation) | EfficientNet · AIGC | mean AUC difference: +0.1425 | Helped |
| add FPH shallow feature extractor (ablation) | ConvNext · AIGC | mean AUC difference: -0.5071 | Hurt |
| add Sobel shallow feature extractor (ablation) | EfficientNet · Deepfake | mean AUC difference: +0.0273 | Helped |
| add Sobel shallow feature extractor (ablation) | SwinTransformer · Deepfake | mean AUC difference: -0.1319 | Hurt |
| add Bayar shallow feature extractor (ablation) | EfficientNet · Document | mean AUC difference: +0.0025 | Helped |
| add DCT shallow feature extractor (ablation) | EfficientNet · IMDL | mean AUC difference: +0.0087 | Helped |
| add FPH shallow feature extractor (ablation) | SegFormer · IMDL | mean AUC difference: -0.0933 | Hurt |
| train PSCC-Net on Doctamper, then evaluate on external document datasets | PSCC-Net · Document cross-domain Average_C | Binary-F1 | Helped |
| use document detector CAFTB on IMDL datasets | CAFTB · IMDL cross-task evaluation | average pixel-level binary F1 | Helped |
Surprising findings
- General visual backbones such as ConvNeXt and Swin Transformer outperform almost all domain-specific forensic SoTA methods under the IFF-Protocol.
- Shallow low-level feature extractors degrade performance for most backbones when training data is large and diverse, contrary to their use in prior work.
Limitations
- Feature extractors are only fused shallowly; more advanced fusion strategies are unexplored.
- All experiments were run once with no error bars or statistical significance reporting.
- Some reproduced models (PS-Net, CAFTB-Net, TIFDM) are reimplementations rather than official code.
Experimental setup
Datasets: FaceForensics++ Celeb-DF DFD FaceShifter UADFV CASIA COVERAGE Columbia IMD2020 NIST16 CocoGlide Autosplice
Benchmarks: DiffusionForensics GenImage DocTamper T-SROIE OSTF RTM TPIC-13 CASIAv1 COVERAGE Columbia IMD2020 NIST16
Metrics: AUC F1 AP MCC TNR TPR ACC IOU
Baselines: Capsule-Net RECCE SPSL UCF SBI MVSS-Net CAT-Net PSCC-Net Trufor IML-ViT Mesorch Dire
Builds on: DeepfakeBench IMDLBenCo CAT-Net training data strategy MVSS-Net training protocol
Closest papers by meaning
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.89)
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.88)
- Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.87)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
- Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation (NeurIPS 2025 · similarity 0.86)
- Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images (CVPR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.