Research map / Detection and segmentation / CNN features
Learning Discriminative Noise Guidance for Image Forgery 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.
Proposes NGNet, a two-step noise-guided image forgery detection and localization method that first learns discriminative noise with a CBDNet-based denoiser constrained by a JS-divergence loss, then uses cross-attention-based guided filters to inject noise guidance into a ResNet-50 RGB branch.
Problem
Image forgery detection and localization (IFDL) is hard because modern generative and post-processing techniques make manipulation traces visually invisible and reduce noise inconsistencies between forged and authentic regions.
Approach
Explicitly enlarge the noise-distribution gap between authentic and forged regions by training a denoiser plus Bayar convolution with a Gaussian JS-divergence constraint, then use a learnable cross-attention-based guided filter (CAGF) that merges guided filtering with attention to transfer noise structure into the RGB branch rather than using simple feature fusion.
Key results
- Pre-trained NGNet reaches 94.1% pixel AUC on Coverage and 87.2% on CASIA, best among compared methods on Coverage, CASIA, NIST16 and IMD20 (85.2%).
- On CASIA-D detection, NGNet achieves 99.81% AUC and 98.72% F1, versus 99.70% and 97.34% for ObjectFormer.
- Ablation on CASIA shows removing noise representation learning drops AUC from 91.3 to 82.6 (w/o NRL) and removing CMA drops it to 84.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full NGNet instead of TANet | Pre-trained image forgery localization · Columbia, Coverage, CASIA, NIST16, IMD20 | pixel-level AUC (%): Columbia: 98.7 -> 98.5; Coverage: 93.2 -> 94.1; CASIA: 85.8 -> 87.2; NIST16: 89.8 -> 90.0; IMD20: 84.9 -> 85.2 | Mixed |
| Use the full NGNet instead of the strongest baseline | Fine-tuned image forgery localization · Coverage, CASIA, NIST16 | pixel-level AUC (%): Coverage: 97.8 -> 98.1; CASIA: 89.3 -> 91.3; NIST16: 99.7 -> 99.8 | Helped |
| Use the full NGNet instead of the strongest baseline | Fine-tuned image forgery localization · Coverage, CASIA, NIST16 | F1 (%): Coverage: 80.1 -> 81.2; CASIA: 61.6 -> 62.1; NIST16: 86.5 -> 86.8 | Helped |
| Use the full NGNet instead of the strongest baseline | Image forgery detection · CASIA-D | image-level AUC (%): 99.70 -> 99.81 | Helped |
| Add noise representation learning (NRL) (ablation) | NGNet · CASIA and NIST16 | AUC: CASIA: 82.6 -> 91.3; NIST16: 83.2 -> 99.8 | Helped |
| Add noise representation learning (NRL) (ablation) | NGNet · CASIA | F1: 49.1 -> 62.1 | Helped |
| Add cross-modal attention (CMA) (ablation) | NGNet · CASIA and NIST16 | AUC: CASIA: 84.4 -> 91.3; NIST16: 88.7 -> 99.8 | Helped |
| Add guided filtering (GF) (ablation) | NGNet · CASIA and NIST16 | AUC: CASIA: 87.2 -> 91.3; NIST16: 93.9 -> 99.8 | Helped |
| Replace CBDNet with another denoising architecture (ablation) | Noise representation learning · Columbia, Coverage, CASIA, NIST16, IMD20 | pixel-level AUC (%): CBDNet*: 98.5, 94.1, 87.2, 90.0, 85.2; FFDNet: 88.4, 87.3, 85.2, 82.1, 75.6; DnCNN: 87.0, 89.4, 86.1, 83.9, 78.4; RIDNet | Mixed |
Surprising findings
- Using a denoiser at sigma=25 (CBDNet) amplifies the noise distribution gap more than directly extracting noise, and the choice of denoiser matters (DnCNN and FFDNet perform worse on several datasets).
Limitations
- Underperforms TANet on Columbia by 0.2% AUC pre-trained, attributed to synthetic training distribution.
- Noise-based guidance relies on the denoiser; the paper states denoiser choice is not the focus and no adaptive selection is provided.
- Pre-training dataset is self-constructed and not released, limiting reproducibility.
Experimental setup
Datasets: CASIA Coverage Columbia NIST16 IMD20 CASIA-D
Benchmarks: CASIA Coverage Columbia NIST16 IMD20 CASIA-D
Metrics: pixel-level AUC pixel-level F1 image-level AUC image-level F1
Baselines: ManTra-Net SPAN PSCC-Net ObjectFormer TANet HiFi-Net RGB-N DnCNN FFDNet RIDNet DRUNet
Builds on: CBDNet MVSS-Net ManTra-Net guided filter ObjectFormer
Closest papers by meaning
- A Lottery Ticket Hypothesis Approach with Sparse Fine-tuning and MAE for Image Forgery Detection and Localization (AAAI 2025 · similarity 0.89)
- MUN: Image Forgery Localization Based on M³ Encoder and UN Decoder (AAAI 2025 · similarity 0.88)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.88)
- DiffForensics: Leveraging Diffusion Prior to Image Forgery Detection and Localization (CVPR 2024 · similarity 0.88)
- SUMI-IFL: An Information-Theoretic Framework for Image Forgery Localization with Sufficiency and Minimality Constraints (AAAI 2025 · similarity 0.87)
- Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.87)
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.