Research map / Detection and segmentation / CNN features
Beyond [CLS] Token: Query-Driven Token-Level Forgery Purification for Generalizable Deepfake Detection
Changshuo Wang, Jiangming Wang, Ke-Yue Zhang, Taiping Yao, Shouhong Ding, Shunli Wang, Ran Yi, Lizhuang Ma
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 |
|---|---|---|---|
| replace [CLS] token with learnable query tokens (ablation) | Effort r=1 · CDF-v1/FSh/DFDCP avg | average AUC: 0.914 -> 0.925 | Improved |
| add fake-likelihood contrastive learning (FLCL) to query tokens (ablation) | QTFP · CDF-v1/FSh/DFDCP avg | average AUC: 0.925 -> 0.931 | Improved |
| add real-attention alignment loss (ablation) | QTFP · CDF-v1/FSh/DFDCP avg | average AUC: 0.931 -> 0.939 | Improved |
| apply FLCL to [CLS] token instead of query tokens (ablation) | Effort r=1 · CDF-v1/FSh/DFDCP avg | average AUC: 0.914 -> 0.929 (vs 0.931 with queries) | Mixed |
| add QTFP to full fine-tuning (ablation) | CLIP FFT · DFDC/DFDCP/WDF avg | average AUC: 0.868 -> 0.895 | Improved |
| add QTFP to LoRA r=4 (ablation) | CLIP LoRA · DFDC/DFDCP/WDF avg | average AUC: 0.877 -> 0.898 | Improved |
| add QTFP to Effort r=1 (ablation) | CLIP Effort · DFDC/DFDCP/WDF avg | average AUC: 0.867 -> 0.903 | Improved |
| add QTFP to BEiTv2 backbone (ablation) | BEiTv2 LoRA r=4 · CDF-v2/FSh avg | average AUC: 0.723 -> 0.781 | Improved |
| add QTFP to DINOv3 backbone (ablation) | DINOv3 LoRA r=4 · CDF-v2/FSh avg | average AUC: 0.854 -> 0.881 | Improved |
| add QTFP to Siglip2 backbone (ablation) | Siglip2 LoRA r=4 · CDF-v2/FSh avg | average AUC: 0.867 -> 0.879 | Improved |
| add QTFP to PE backbone (ablation) | PE LoRA r=4 · CDF-v2/FSh avg | average AUC: 0.908 -> 0.926 | Improved |
| add QTFP to CLIP backbone (ablation) | CLIP LoRA r=4 · CDF-v2/FSh avg | average AUC: 0.909 -> 0.932 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Forensics Adapter — Forensics Adapter: Adapting CLIP for Generalizable Face Forgery Detection (CVPR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.90)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.90)
- Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (CVPR 2024 · similarity 0.89)
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.89)
- Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection (ICML 2025 · similarity 0.88)
- FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing (CVPR 2025 · similarity 0.88)
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.
