Research map / Image generation / Editing and style transfer
A New Benchmark and Model for Challenging Image Manipulation Detection
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.
Introduces the CIMD benchmark (raw-image editing subset and same-QF double-compressed subset) and a two-branch RGB + DCT frequency network built on HRNet for detecting small and identically-quantized tampered regions.
Problem
Image manipulation detection struggles with very small tampered regions and with double JPEG compression using identical quality factors, and no public benchmark targets these challenging cases.
Approach
A two-branch HRNet-based network: an RGB stream for editing artifacts and a frequency stream fed with quantized DCT histograms, de-quantized features, and a residual DCT map from repeated recompression, plus ASPP and channel/spatial attention and soft-selection heatmap fusion.
Key results
- On CIMD-R, the proposed model reaches best pixel F1 0.444 and image AUC 0.677, versus 0.301 and 0.589 for the strongest baseline CAT-Net.
- On CIMD-C (same-QF double compression), the model reaches best pixel F1 0.542 and AUC 0.727, versus 0.395 and 0.534 for CAT-Net.
- Ablation on CIMD-R shows the frequency stream alone gives F1 0.130, while the RGB stream alone gives 0.330 and the combined model 0.335.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed two-branch method instead of CAT-Net | CIMD-R · CIMD-R | pixel-level F1 (best): 0.301 -> 0.444 | Helped |
| use the proposed two-branch method instead of CAT-Net | CIMD-R · CIMD-R | pixel-level F1 (fixed): 0.194 -> 0.335 | Helped |
| use the proposed two-branch method instead of CAT-Net | CIMD-R · CIMD-R | AUC: 0.589 -> 0.677 | Helped |
| use the proposed two-branch method instead of CAT-Net | CIMD-R · CIMD-R | image-level accuracy: 0.537 -> 0.545 | Helped |
| use the proposed two-branch method instead of CAT-Net | CIMD-C · CIMD-C | pixel-level F1 (best): 0.395 -> 0.542 | Helped |
| use the proposed two-branch method instead of CAT-Net | CIMD-C · CIMD-C | pixel-level F1 (fixed): 0.259 -> 0.442 | Helped |
| use the proposed two-branch method instead of CAT-Net | CIMD-C · CIMD-C | AUC: 0.534 -> 0.727 | Helped |
| use the proposed two-branch method instead of CAT-Net | CIMD-C · CIMD-C | image-level accuracy: 0.490 -> 0.525 | Helped |
| add the frequency stream to the RGB stream (ablation) | two-branch HRNet · CIMD-R | F1: 0.330 -> 0.335 | Helped |
| add the frequency stream to the RGB stream (ablation) | two-branch HRNet · CIMD-R | AUC: 0.593 -> 0.677 | Helped |
| add the frequency stream to the RGB stream (ablation) | two-branch HRNet · CIMD-C | F1: 0.409 -> 0.442 | Helped |
| add the frequency stream to the RGB stream (ablation) | two-branch HRNet · CIMD-C | AUC: 0.525 -> 0.727 | Helped |
Surprising findings
- The frequency stream alone performs poorly on CIMD-R (uncompressed images), as expected since no compression artifacts exist there.
- Several prior IMD methods score near chance (AUC about 0.5) on the challenging CIMD subsets.
Limitations
- Training datasets are not specified in the text beyond being adopted from prior work.
- The CIMD benchmark is small (600 raw TIFF images and 200 JPEG images) and created with one camera and one editing tool.
Experimental setup
Datasets: training datasets adopted from prior work (unspecified)
Benchmarks: CIMD-R CIMD-C
Metrics: pixel-level F1 (best and fixed 0.5 threshold) image-level AUC image-level accuracy
Baselines: RRU-Net CR-CNN MantraNet SPAN HiFi_IFDL PSCC-Net CAT-Net MVSS-Net IF-OSN DJPEG Comprint
Builds on: HRNet ASPP CAT-Net
Closest papers by meaning
- IDseq: Decoupled and Sequentially Detecting and Grounding Multi-Modal Media Manipulation (AAAI 2025 · similarity 0.85)
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.85)
- Training-Free Image Manipulation Localization Using Diffusion Models (AAAI 2025 · similarity 0.84)
- Towards Modern Image Manipulation Localization: A Large-Scale Dataset and Novel Methods (CVPR 2024 · similarity 0.84)
- IMDPrompter: Adapting SAM to Image Manipulation Detection by Cross-View Automated Prompt Learning (ICLR 2025 · similarity 0.84)
- GIM: A Million-scale Benchmark for Generative Image Manipulation Detection and Localization (AAAI 2025 · similarity 0.83)
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.