Research map / Label-efficient and robust learning / Recognition and long-tail
DNA: Uncovering Universal Latent Forgery Knowledge
Jingtong Dou, Chuancheng Shi, Anqi Yi, Shiming Guo, Wenhua Wu, Yemin Wang, Li Zhang, Fei Shen, Tat-Seng Chua
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use critical layers 12-18 instead of shallow layers 1-6 (ablation) | DNA · SDv1.4 | accuracy: 62.1 -> 98.4 | Improved |
| use critical layers 12-18 instead of deep layers 19-24 (ablation) | DNA · SDv1.4 | accuracy: 86.2 -> 98.4 | Improved |
| use critical layers 12-18 instead of transition layers 7-11 (ablation) | DNA · SDv1.4 | accuracy: 90.7 -> 98.4 | Improved |
| use FDUs instead of excluding them (Random-Ex) (ablation) | DNA · SDv1.4 | accuracy: 89.25 -> 98.4 | Improved |
| keep FDUs unmasked rather than masking them (ablation) | DNA · base model test | accuracy: 93.1 -> 65.1 | Improved |
| keep FDUs unmasked rather than masking them (ablation) | DNA · Midjourney | accuracy: 81.6 -> 59.3 | Improved |
| mask random neurons (ablation) | DNA · base model test | accuracy: 93.1 -> 91.1 | Hurt |
| mask magnitude-matched non-FDU neurons (ablation) | DNA · base model test | accuracy: 93.1 -> 92.0 | Hurt |
| increase training set from 5,000 to 30,000 images (ablation) | DNA · SDv1.4 | accuracy: 96.7 -> 95.0 | Hurt |
| apply JPEG compression quality 40 | DNA · perturbation robustness | AP: AP remains 92.4 | No clear effect |
| use DNA instead of MoLD | GenImage Midjourney subset · GenImage Midjourney subset | accuracy: 76.3 -> 93.7 | Improved |
| use DNA instead of MoLD | HIFI-Gen · HIFI-Gen | accuracy: 81.0 -> 96.4 | Improved |
Closest papers by meaning
- Asymmetric Anchoring: Opening the Black Box of MLLMs for Forgery Detection (ECCV 2026 · similarity 0.86)
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.85)
- OmniVL-Guard: Towards Unified Vision-Language Forgery Detection and Grounding via Balanced RL (ICML 2026 · similarity 0.84)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.84)
- DNA-DetectLLM: Unveiling AI-Generated Text via a DNA-Inspired Mutation-Repair Paradigm (NeurIPS 2025 · similarity 0.83)
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 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.
