Research map / Detection and segmentation / CNN features
TruFor: Leveraging All-Round Clues for Trustworthy Image Forgery Detection and Localization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Noiseprint with Noiseprint++ (learned fingerprint trained with editing-history contrastive loss) (ablation) | Noiseprint baseline · ablation dataset of 1000 locally edited images (original/Res/Res&Cmp) | pixel-level F1 (best threshold): Original .706 -> .877; Res .375 -> .666 | Helped |
| Add transformer segmentation (SegFormer) on top of Noiseprint++ features (ablation) | Noiseprint++ · ablation dataset (original images) | pixel-level F1 (best threshold): .877 -> .974 | Helped |
| Fuse RGB and Noiseprint++ branches (full TruFor) instead of single-input SegFormer (ablation) | SegFormer (NP++) / SegFormer (RGB) · ablation dataset, original and Res&Cmp | pixel-level F1 (best threshold): Original .974 -> .982 (vs NP++); Res&Cmp .649 -> .765 (vs NP++) but .756 (RGB-only) vs .765 | Mixed |
| Add dice loss and online augmentation with resized/compressed images to cross-entropy loss (ablation) | Localization training · Res&Cmp images | pixel-level F1 (best threshold): .583 -> .765 | Helped |
| Add dice loss without augmentation (ablation) | Localization training · Original images | pixel-level F1 (best threshold): .980 (CE) vs .973 (CE+DL) | No clear effect |
| Use confidence-weighted pooling of anomaly map for detection instead of max pooling (ablation) | Detection score · ablation on original images and Res&Cmp | image-level AUC / accuracy: Res&Cmp accuracy .500 -> .675; Res&Cmp AUC .730 -> .740; original AUC .974 -> .996 | Helped |
| Add confidence map to detection pipeline (ablation) | TruFor detection · average over 7 datasets | image-level AUC and accuracy: AUC .877 -> .996 (original); Acc .785 -> .905 | Helped |
| Use mean of anomaly map as detection statistic instead of max (ablation) | Detection score · ablation on original images | image-level AUC: .974 (max) -> .544 (mean) | Hurt |
| TruFor full framework vs prior SOTA on localization | TruFor vs CAT-Net v2 and others · 8 datasets averaged (pixel F1, best threshold) | pixel-level F1: .785 vs .709 (CAT-Net v2) | Helped |
| TruFor full framework vs prior SOTA on detection | TruFor vs CAT-Net v2 and others · 7 datasets averaged (image-level) | balanced accuracy at fixed 0.5: .781 vs .649 (CAT-Net v2) | Helped |
| Joint training with confidence decoder and detection classifier | TruFor detection · robustness to social network processing (Facebook/WhatsApp) | pixel-level F1 (fixed): TruFor outperforms IF-OSN, CAT-Net v2, MVSS-Net on all Fb/Wa datasets | Helped |
Closest papers by meaning
- Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (AAAI 2024 · similarity 0.89)
- A Lottery Ticket Hypothesis Approach with Sparse Fine-tuning and MAE for Image Forgery Detection and Localization (AAAI 2025 · similarity 0.87)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.87)
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.86)
- Forgery-aware Adaptive Transformer for Generalizable Synthetic Image Detection (CVPR 2024 · similarity 0.84)
- SUMI-IFL: An Information-Theoretic Framework for Image Forgery Localization with Sufficiency and Minimality Constraints (AAAI 2025 · similarity 0.84)
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.