Research map / Label-efficient and robust learning / Recognition and long-tail
Orthogonal Subspace Decomposition for Generalizable AI-Generated Image 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.
Proposes Effort, which applies SVD to a CLIP ViT-L/14 linear layer to split weights into frozen principal and trainable residual subspaces with orthogonality and singular-value constraints, improving generalization for deepfake and synthetic image detection.
Problem
AI-generated image detectors trained naively overfit to limited fake patterns, yielding low-rank feature spaces that generalize poorly to unseen forgery methods.
Approach
Decompose a pre-trained VFM weight matrix via SVD into a frozen top-r principal part and a learnable residual part, enforcing orthogonal singular vectors and a preserved Frobenius norm so fake-specific adaptation does not distort pre-trained knowledge.
Key results
- On FF++-trained cross-dataset evaluation, Effort reaches 0.917 average AUC with only 0.19M trainable parameters, versus 0.878 for CDFA (87M).
- On UniversalFakeDetect with CLIP ViT-L/14, Effort achieves 99.41% mAP and 95.19% mAcc, improving over UniFD by 13.81 mAP and 9.27 mAcc.
- On GenImage (SDv1.4 training), Effort reports 91.1% mAcc, above DRCT (89.5%) and FatFormer (88.9%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SVD orthogonal subspace decomposition to the detector (ablation) | CLIP-based Effort on deepfake detection · CDF-v2 and SimSwap (Avg.) | AUC: 0.859 -> 0.925 | Helped |
| Add the singular value constraint to SVD adaptation (ablation) | CLIP-based Effort on deepfake detection · CDF-v2 and SimSwap (Avg.) | AUC: 0.925 -> 0.936 | Helped |
| Add the orthogonal constraint to SVD adaptation (ablation) | CLIP-based Effort on deepfake detection · CDF-v2 and SimSwap (Avg.) | AUC: 0.925 -> 0.930 | Helped |
| Use SVD with both singular value and orthogonal constraints (ablation) | CLIP-based Effort on deepfake detection · CDF-v2 and SimSwap (Avg.) | AUC: 0.859 -> 0.941 | Helped |
| Set SVD residual rank n-r to 1 instead of 256 (ablation) | Effort on UniversalFakeDetect · UniversalFakeDetect (19 subsets) | mAcc: 92.13 -> 95.19 | Helped |
| Use Effort with residual rank n-r=1 rather than LoRA rank 1 (ablation) | CLIP-based detector on UniversalFakeDetect · UniversalFakeDetect (19 subsets) | mAcc: 93.03 -> 95.19 | Helped |
| Use Effort rather than full fine-tuning (ablation) | CLIP-based detector on UniversalFakeDetect · UniversalFakeDetect (19 subsets) | mAcc: 86.22 -> 95.19 | Helped |
| Add Effort adaptation to BEIT-v2 initialization (ablation) | Deepfake detector · CDF-v2 | AUC: 0.855 -> 0.894 | Helped |
| Add Effort adaptation to SigLIP initialization (ablation) | Deepfake detector · SimSwap | AUC: 0.713 -> 0.778 | Helped |
| Add Effort adaptation to CLIP ViT-L/14 initialization (ablation) | Synthetic image detector · UniversalFakeDetect (19 subsets) | mAcc: 86.22 -> 95.19 | Helped |
| Use CLIP-Large/14 rather than CLIP-Base/16 with Effort (ablation) | Effort on UniversalFakeDetect · UniversalFakeDetect (19 subsets) | mAcc: 88.46 -> 95.19 | Helped |
| Use Effort instead of UniFD | Synthetic image detector on UniversalFakeDetect · UniversalFakeDetect (19 subsets) | mAcc: +9.27% | Helped |
Surprising findings
- Full fine-tuning of CLIP reduces the effective rank of the feature space from 160 to 57 on the face domain, while SVD-based tuning keeps it at 159.
Limitations
- Treats all fake methods as a single class during training, ignoring specificity across generators.
- Rank choice (n-r=1) was motivated by binary task simplicity and may not transfer to more diverse fake data.
- Results depend on the choice of pre-trained VFM; CLIP performed best among tested backbones.
Experimental setup
Datasets: FaceForensics++ (c23) UniversalFakeDetect (ProGAN training) GenImage (SDv1.4 training) DF40
Benchmarks: CDF-v2 DFD DFDC DFDCP DFo WDF FFIW DF40 UniversalFakeDetect GenImage GID cross-manipulation FF++
Metrics: video-level AUC AP accuracy mAP mAcc effective rank
Baselines: F3Net SPSL SRM CORE RECCE SBI UCF IID LSDA ProDet CDFA UniFD
Builds on: LoRA UniFD DeepfakeBench
Closest papers by meaning
- Detecting Generated Images by Fitting Natural Image Distributions (NeurIPS 2025 · similarity 0.85)
- PiD: Generalized AI-Generated Images Detection with Pixelwise Decomposition Residuals (ICML 2025 · similarity 0.85)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.85)
- Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection (NeurIPS 2025 · similarity 0.84)
- Training-free Detection of AI-generated images via Cropping Robustness (NeurIPS 2025 · similarity 0.84)
- MLEP: Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection (NeurIPS 2025 · similarity 0.84)
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.