Research map / Architectures and efficiency / Pruning and efficient tuning
SRRA: Stable-Rank-Based Residual Adaptation for Generalizable Deepfake Detection
Huakun Liu, Wenjie Li, Changsheng Xu
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 |
|---|---|---|---|
| add Stable Rank Decomposition (adaptive residual rank) to CLIP baseline (ablation) | SRRA · Cross-dataset avg (train FF++ c23) | AUC: 0.832 -> 0.880 | Improved |
| add Residual Energy Constraint on top of SRD (ablation) | SRRA · Cross-dataset avg (train FF++ c23) | AUC: 0.880 -> 0.894 | Improved |
| add Residual Subspace Orthogonalization alone on top of SRD (ablation) | SRRA · Cross-dataset avg (train FF++ c23) | AUC: 0.880 -> 0.881 | No clear effect |
| add RSO on top of SRD + REC (ablation) | SRRA · Cross-dataset avg (train FF++ c23) | AUC: 0.894 -> 0.901 | Improved |
| replace fixed-rank Effort (best r_res=16) with adaptive stable-rank allocation (ablation) | SRRA · CDF/DFDCP/DFD avg | AUC: 0.897 -> 0.916 | Improved |
| replace fixed-rank LoRA (best r_res=16) with adaptive stable-rank allocation (ablation) | SRRA · CDF/DFDCP/DFD avg | AUC: 0.887 -> 0.916 | Improved |
| increase fixed residual rank from 16 to 64 in Effort (ablation) | Effort · CDF/DFDCP/DFD avg | AUC: 0.897 -> 0.891 | Hurt |
| scale backbone from CLIP ViT-B/16 to ViT-L/14 (ablation) | SRRA · Cross-dataset avg (train FF++ c23) | AUC: 0.849 -> 0.901 | Improved |
| use SRRA instead of Effort on ViT-B/16 (ablation) | SRRA · Cross-dataset avg (ViT-B/16) | AUC: 0.829 -> 0.849 | Improved |
| scale stable rank by tau=1/10 instead of using it directly (tau=1) (ablation) | SRRA · Cross-dataset avg (train FF++ c23) | AUC: 0.869 -> 0.901 | Improved |
| lower tau from 1/10 to 1/16 (ablation) | SRRA · Cross-dataset avg (train FF++ c23) | AUC: 0.901 -> 0.898 | Hurt |
| use loss weights lambda1=0.1, lambda2=0.1 instead of 0.05/0.02 (ablation) | SRRA · Cross-dataset avg (train FF++ c23) | AUC: 0.881 -> 0.901 | Improved |
Closest papers by meaning
- Towards More General Video-based Deepfake Detection through Facial Component Guided Adaptation for Foundation Model (CVPR 2025 · similarity 0.88)
- Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning (CVPR 2025 · similarity 0.87)
- Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (CVPR 2024 · similarity 0.87)
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.87)
- FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing (CVPR 2025 · similarity 0.87)
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · 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.
