Research map / Detection and segmentation / Open-set and few-shot detection
C2P-CLIP: Injecting Category Common Prompt in CLIP to Enhance Generalization in Deepfake 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.
Decodes CLIP detection features into text to show CLIP matches similar concepts, then injects category common prompts (e.g., Deepfake/Camera) via caption-based contrastive learning with LoRA on the CLIP image encoder for generalizable AIGC detection.
Problem
Generalizable deepfake/AIGC image detection with CLIP features and linear classifiers: why CLIP features work and how to improve CLIP's detection performance on unseen generators.
Approach
Analyze CLIP detection features by decoding them to text (ClipCap) and word frequency analysis, then inject manually specified category common prompts into captions and train CLIP's image encoder (LoRA) with contrastive plus classification loss while freezing the text encoder, so no extra parameters are needed at test time.
Key results
- On UniversalFakeDetect, the Deepfake/Camera prompt variant reaches 93.79% mAcc and 98.66% mAP, improving mAcc by 12.41% and mAP by 8.52% over UniFD.
- On UniversalFakeDetect accuracy, the method improves about 2.93% over FatFormer (state-of-the-art per the paper).
- On GenImage (trained on SDv1.4), the Deepfake/Camera variant reaches 95.8% mAcc, 7.0% above UniFD and 6.9% above the best prior method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Inject category common prompts into CLIP | CLIP ViT-L/14 · UniversalFakeDetect | mAcc: 12.41% | Helped |
| Inject category common prompts into CLIP | CLIP ViT-L/14 · UniversalFakeDetect | mAP: 8.52% | Helped |
| Inject category common prompts into CLIP | CLIP ViT-L/14 · UniversalFakeDetect | accuracy: 2.93% | Helped |
| Inject category common prompts into CLIP | CLIP ViT-L/14 · GenImage | accuracy: 7.0% | Helped |
| Use Deepfake and Camera category prompts instead of Trump and Biden (ablation) | C2P-CLIP · UniversalFakeDetect | mAcc: 93.00 -> 93.79 | Helped |
| Use Deepfake and Camera category prompts instead of Trump and Biden (ablation) | C2P-CLIP · GenImage | mAcc: 94.7 -> 95.8 | Helped |
Surprising findings
- Decoding CLIP's linear-classifier detection features into text yields captions unrelated to image content (e.g., a cat image decodes to a street scene), suggesting CLIP matches similar concepts rather than true/fake semantics.
Limitations
- Word frequency analysis of decoded detection features considers only individual words, lacking analysis of the full caption.
- Category common prompts are manually chosen, and results vary across prompt pairs (93.79% vs 93.00% accuracy).
- Lower accuracy on some subsets such as SAN (76.00% and 84.56% AP) compared to other generators.
Experimental setup
Datasets: UniversalFakeDetect GenImage
Benchmarks: UniversalFakeDetect GenImage
Metrics: mAP mAcc accuracy average precision
Baselines: UniFD FatFormer NPR FreqNet LGrad F3Net CNN-Spot Patchfor Co-occurence Freq-spec ResNet-50 DeiT-S
Builds on: UniFD ClipCap LoRA CLIP
Closest papers by meaning
- Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection (ICML 2025 · similarity 0.85)
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.85)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.85)
- Standing on the Shoulders of Giants: Reprogramming Visual-Language Model for General Deepfake Detection (AAAI 2025 · similarity 0.84)
- Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning (CVPR 2025 · similarity 0.83)
- Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · 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.