Paired Geometric Supervision for Generalizable Deepfake Detection
Organizations: School of Information Technology Deakin University Waurn Ponds, Australia · Department of Computer Science University of Warwick Coventry, UK · Universitat Autònoma de Barcelona, Spain, as a Distinguished Researcher
Abstract
We present OrthoPairGTL, a geometric training framework for cross-domain deepfake detection, built on a partially frozen DINOv3 backbone (300M parameters), and make three contributions: (1) We re-evaluate six recent transfer-learning detectors (Forensics Adapter, Effort, three GenD variants and FakeSTormer) using the authors' released checkpoints under a single preprocessing pipeline across 17 cross-domain datasets (2018-2025), so that the reported gaps reflect method rather than per-method pipeline tuning. (2) We model the transformation between paired real and fake samples, training Parameter-Efficient Fine-Tuning (PEFT) adapters with an orthogonal pair decomposition, hyperspherical regularisation, and (in the GSD-LoRA variant) a set-level coherence objective. (3) Under LoRA, this regularisation gives a more stable optimisation profile: GSD-LoRA's post-peak validation degradation is shallower than plain LoRA at the same parameter footprint, and the time-averaged validation AUC across training is 1.82 pp higher. On the 17 datasets, GSD-LoRA matches or exceeds every re-evaluated baseline (92.82% vs. 91.58% for the strongest; paired Wilcoxon signed-rank , 13/17 per-dataset wins), with the largest gains on modern reenactment-heavy benchmarks (DSv1.1, DSv2). Under strict parameter budgets, LayerNorm-only tuning (0.03% of the backbone) is the most stable configuration and matches the strongest baseline with a mean of 91.97%.
Figures & tables
| Configuration | Backbone params | Head | Description |
|---|---|---|---|
| LN | 100K (0.03%) | 2.1K | LayerNorm affine only |
| LoRA | 28.3M (9.34%) | 2.1M | Mid-rank PEFT adapter |
| GSD-LoRA | 28.3M (9.34%) | 2.1M | LoRA + set-geometric regularisation |
| Method | FF-DF | FF-F2F | FF-FS | FF-NT | MEAN |
| Forensics Adapter [ Cui et al.(2025)Cui, Li, Luo, Zhou, and Dong ] | 99.30 | 95.64 | 97.70 | 89.46 | 95.53 |
| Effort [ \citeyear yan2025orthogonal] | 99.42 | 93.79 | 98.59 | 84.20 | 94.00 |
| GenD CLIP [ \citeyear yermakov2026deepfake] | 99.17 | 97.17 | 98.63 | 93.97 | 97.24 |
| GenD PE [ \citeyear yermakov2026deepfake] | 99.70 | 98.84 | 99.14 | 96.79 | 98.62 |
| GenD DINO [ \citeyear yermakov2026deepfake] | 99.67 | 99.16 | 99.38 | 96.64 | 98.71 |
| FakeSTormer [ \citeyear nguyen2025vulnerability] | 98.93 | 78.21 | 84.36 | 79.56 | 85.27 |
| Methods | UADFV | DFD | CDFv1 | DFDCP | DFDC | CDFv2 | FF-FSH | DeFo | FFIW | KoDF | FAVC | DFDM | DSv1.1 | PGF | IDFv1 | DSv2 | CDFv3 | MEAN |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Forensics Adapter [ Cui et al.(2025)Cui, Li, Luo, Zhou, and Dong ] | 98.63 | 96.57 | 93.55 | 89.20 | 85.41 | 93.74 | 76.30 | 99.00 | 83.06 | 83.60 | 89.06 | 96.02 | 83.89 | 82.88 | 85.85 | 72.04 | 73.98 | 87.22 |
| Effort [ Yan et al.(2025)Yan, Wang, Jin, Zhang, Liu, Chen, Yao, Ding, Wu, and Yuan ] | 97.92 | 95.55 | 9 5.08 | 84.13 | 84.02 | 93.03 | 90.70 | 98.96 | 85.66 | 88.19 | 88.79 | 97.97 | 81.99 | 84.10 | 89.68 | 63.43 | 77.43 | 88.04 |
| GenD CLIP [ Yermakov et al.(2026)Yermakov, Cech, Matas, and Fritz ] | 99.46 | 97.58 | 93.29 | 89.92 | 85.60 | 9 4.79 | 85.07 | 95.70 | 87.47 | 86.71 | 94.81 | 99.77 | 90.24 | 89.50 | 94.01 | 73.73 | 79.19 | 90.40 |
| GenD PE [ Yermakov et al.(2026)Yermakov, Cech, Matas, and Fritz ] | 98.13 | 97.71 | 95.42 | 86.40 | 81.60 | 95.66 | 87.11 | 93.76 | 86.47 | 85.98 | 97.03 | 98.54 | 89.51 | 92.47 | 95.05 | 79.19 | 87.96 | 91.06 |
| GenD DINO [ Yermakov et al.(2026)Yermakov, Cech, Matas, and Fritz ] | 9 8.88 | 9 7.90 | 92.40 | 88.82 | 84.66 | 91.30 | 90.02 | 98.57 | 8 9.31 | 9 0.96 | 97.82 | 99.81 | 87.77 | 92.28 | 96.66 | 80.17 | 79.51 | 91.58 |
| FakeSTormer [ Nguyen et al.(2025)Nguyen, Astrid, Kacem, Ghorbel, and Aouada ] | 98.21 | 87.38 | 90.15 | 90.10 | 74.64 | 89.83 | 79.19 | 53.26 | 71.82 | 81.48 | 88.24 | 99.67 | 73.95 | 73.86 | 88.97 | 61.91 | 68.62 | 80.66 |
| Variant | MEAN-17 (%) | Wins | ||
|---|---|---|---|---|
| LN | 91.97 | 0.43 | 11/17 | |
| LoRA | 92.62 | 0.020 | 13/17 | |
| GSD-LoRA | 92.82 | 0.015 | 13/17 |
| Variant | Peak epoch | Peak val-AUC | Post-peak slope | AUC-under-curve |
|---|---|---|---|---|
| (%) | (pp/epoch) | (%) | ||
| LN | ||||
| LoRA | ||||
| GSD-LoRA |
| Configuration | DSv2 | CDFv3 | MEAN |
|---|---|---|---|
| Frozen + linear probe | 68.48 | 65.41 | 70.75 ( ) |
| LN tuning only (no geometric) | 79.32 | 79.63 | 91.32 ( ) |
| + OPA | 81.67 | 83.88 | 92.36 ( ) |
| + OPA + uniformity | 80.82 | 74.38 | 90.90 ( ) |
| + OPA + uniformity + GTL ( Full method ) | 81.33 | 83.37 | 91.97 ( ) |
| LoRA + MLP (no geometric) | 82.81 | 84.30 | 92.24 ( ) |
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Year | Mod. | #Real | #Fake |
|---|---|---|---|---|
| UADFV | 2018 | RGB | 49 | 49 |
| FF++ (c23) | 2019 | RGB | 140 | 560 |
| DFD | 2019 | RGB | 363 | 3,068 |
| CDFv1 | 2019 | RGB | 38 | 62 |
| DFDCP ‡ | 2019 | AV | 275 | 495 |
| DFDC ‡ | 2020 | AV | 2,490 | 2,489 |
| Method | UADFV | DFD | DFDC | CDFv2 | FFIW | KoDF | FAVC | DFDM | DSv1.1 | PGF | IDFv1 | DSv2 | CDFv3 | MEAN |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FA (reprod.) | 99.42 | 97.19 | 87.26 | 95.69 | 89.41 | 89.44 | 92.29 | 97.11 | 82.09 | 86.44 | 85.91 | 73.40 | 75.21 | 88.53 |
| FA (ours) | 98.63 \textcolor red!70!black | 96.57 \textcolor red!70!black | 85.41 \textcolor red!70!black | 93.74 \textcolor red!70!black | 83.06 \textcolor red!70!black | 83.60 \textcolor red!70!black | 89.06 \textcolor red!70!black | 96.02 \textcolor red!70!black | 83.89 \textcolor darkgreen | 82.88 \textcolor red!70!black | 85.85 \textcolor red!70!black | 72.04 \textcolor red!70!black | 73.98 \textcolor red!70!black | 86.52 \textcolor red!70!black |
| Effort (reprod.) | 97.63 | 95.32 | 85.29 | 93.00 | 91.10 | 90.57 | 91.45 | 98.20 | 81.54 | 84.53 | 88.93 | 63.90 | 79.63 | 87.78 |
| Effort (ours) | 97.92 \textcolor darkgreen | 95.55 \textcolor darkgreen | 84.02 \textcolor red!70!black | 93.03 \textcolor darkgreen | 85.66 \textcolor red!70!black | 88.19 \textcolor red!70!black | 88.79 \textcolor red!70!black | 97.97 \textcolor red!70!black | 81.99 \textcolor darkgreen | 84.10 \textcolor red!70!black | 89.68 \textcolor darkgreen | 63.43 \textcolor red!70!black | 77.43 \textcolor red!70!black | 86.75 \textcolor red!70!black |
| GenD CLIP (reprod.) | 99.00 | 97.08 | 87.05 | 96.25 | 91.93 | 88.98 | 96.19 | 99.82 | 92.04 | 90.48 | 94.75 | 73.28 | 86.86 | 91.82 |
| GenD CLIP (ours) | 99.46 \textcolor darkgreen | 97.58 \textcolor darkgreen | 85.60 \textcolor red!70!black | 94.79 \textcolor red!70!black | 87.47 \textcolor red!70!black | 86.71 \textcolor red!70!black | 94.81 \textcolor red!70!black | 99.77 \textcolor red!70!black | 90.24 \textcolor red!70!black | 89.50 \textcolor red!70!black | 94.01 \textcolor red!70!black | 73.73 \textcolor darkgreen | 79.19 \textcolor red!70!black | 90.22 \textcolor red!70!black |
| Method | FF-DF | FF-F2F | FF-FS | FF-NT | MEAN |
|---|---|---|---|---|---|
| FA (reprod.) | 99.71 | 96.95 | 98.65 | 91.89 | 96.80 |
| FA (ours) | 99.30 \textcolor red!70!black | 95.64 \textcolor red!70!black | 97.70 \textcolor red!70!black | 89.46 \textcolor red!70!black | 95.53 \textcolor red!70!black |
| Effort (reprod.) | 99.42 | 93.43 | 98.49 | 84.76 | 94.03 |
| Effort (ours) | 99.42 \textcolor darkgreen | 93.79 \textcolor darkgreen | 98.59 \textcolor darkgreen | 84.20 \textcolor red!70!black | 94.00 \textcolor red!70!black |
| GenD CLIP (reprod.) | 99.54 | 98.20 | 98.77 | 95.28 | 97.95 |
| GenD CLIP (ours) | 99.17 \textcolor red!70!black | 97.17 \textcolor red!70!black | 98.63 \textcolor red!70!black | 93.97 \textcolor red!70!black | 97.24 \textcolor red!70!black |
| Method | UADFV | DFD | CDFv1 | DFDCP | DFDC | CDFv2 | FF-FSH | DeFo | FFIW | KoDF | FAVC | DFDM | DSv1.1 | PGF | IDFv1 | DSv2 | CDFv3 | MEAN |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FA (orig.) | — | 97.2 | — | 92.9 | 87.2 | 95.7 | — | — | — | — | — | — | — | — | — | — | — | 93.25 ( ) |
| FA (ours) | 98.63 | 96.57 \textcolor red!70!black | 93.55 | 89.20 \textcolor red!70!black | 85.41 \textcolor red!70!black | 93.74 \textcolor red!70!black | 76.30 | 99.00 | 83.06 | 83.60 | 89.06 | 96.02 | 83.89 | 82.88 | 85.85 | 72.04 | 73.98 | 91.23 \textcolor red!70!black ( ) |
| Effort (orig.) | — | 96.5 | — | 90.9 | 84.3 | 95.6 | — | 97.7 | 92.1 | — | — | — | — | — | — | — | — | 92.85 ( ) |
| Effort (ours) | 97.92 | 95.55 \textcolor red!70!black | 95.08 | 84.13 \textcolor red!70!black | 84.02 \textcolor red!70!black | 93.03 \textcolor red!70!black | 90.70 | 98.96 \textcolor darkgreen | 85.66 \textcolor red!70!black | 88.19 | 88.79 | 97.97 | 81.99 | 84.10 | 89.68 | 63.43 | 77.43 | 90.23 \textcolor red!70!black ( ) |
| FakeSTormer (orig., T=4) | — | 98.5 | — | 90.0 | 74.6 | 92.4 | — | — | — | — | — | — | — | — | — | — | — | 88.88 ( ) |
| FakeSTormer (ours) | 98.21 | 87.38 \textcolor red!70!black | 90.15 | 90.10 \textcolor darkgreen | 74.64 \textcolor darkgreen | 89.83 \textcolor red!70!black | 79.19 | 53.26 | 71.82 | 81.48 | 88.24 | 99.67 | 73.95 | 73.86 | 88.97 | 61.91 | 68.62 | 85.49 \textcolor red!70!black ( ) |
| Method | UADFV | DFD | DFDC | CDFv2 | FFIW | KoDF | FAVC | DFDM | DSv1.1 | PGF | IDFv1 | DSv2 | CDFv3 | MEAN-13 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FA | 98.63 | 96.57 | 85.41 | 93.74 | 83.06 | 83.60 | 89.06 | 96.02 | 83.89 | 82.88 | 85.85 | 72.04 | 73.98 | 86.52 |
| Effort | 97.92 | 95.55 | 84.02 | 93.03 | 85.66 | 88.19 | 88.79 | 97.97 | 81.99 | 84.10 | 89.68 | 63.43 | 77.43 | 86.75 |
| GenD CLIP | 99.46 | 97.58 | 85.60 | 9 4.79 | 87.47 | 86.71 | 94.81 | 99.77 | 90.24 | 89.50 | 94.01 | 73.73 | 79.19 | 90.22 |
| GenD PE | 98.13 | 97.71 | 81.60 | 95.66 | 86.47 | 85.98 | 97.03 | 98.54 | 89.51 | 92.47 | 95.05 | 79.19 | 87.96 | 91.18 |
| GenD DINO | 9 8.88 | 9 7.90 | 84.66 | 91.30 | 8 9.31 | 9 0.96 | 97.82 | 99.81 | 87.77 | 92.28 | 96.66 | 80.17 | 79.51 | 91.31 |
| FST | 98.21 | 87.38 | 74.64 | 89.83 | 71.82 | 81.48 | 88.24 | 99.67 | 73.95 | 73.86 | 88.97 | 61.91 | 68.62 | 81.43 |
| Checkpoint policy | LN | LoRA | GSD-LoRA |
|---|---|---|---|
| macro-best (default) | 91.97 0.10 | 92.62 0.33 | 92.82 0.35 |
| DSv2-VAL only | 90.42 0.43 | 92.48 0.24 | 92.80 0.54 |
| CDFv3-VAL only | 91.99 0.07 | 92.35 0.45 | 92.69 0.56 |
| FF-FSH-VAL only | 92.01 0.05 | 92.30 0.87 | 92.74 0.32 |
| Reference: GenD DINO MEAN-17 = 91.58 (Table 3 ). | |||
| Explained Var. ( ) | Raw- attrib. (%) | Residual attrib. (%) | |
|---|---|---|---|
| 64 | 0.963 | 39.5 | 65.7 |
| 128 | 0.985 | 38.6 | 69.4 |
| 256 | 0.995 | 38.7 | 70.8 |
| 512 | 0.999 | 38.3 | 71.9 |
| 768 | 1.000 | 39.2 | 66.6 |
| Type combo | mean cos(arc) | mean cos( ) | Pearson | |
|---|---|---|---|---|
| swap swap | 514,605 | |||
| swap reenact | 817,075 | |||
| reenact reenact | 323,610 | |||
| all combinations | 1,655,290 |
| Partition | mean cos(arc) | mean cos( ) | |
|---|---|---|---|
| same target identity | 11,445 | ||
| different target identity | 1,643,845 | ||
| separation (same different) | — |
| Component | Symbol | Value | Introduced |
|---|---|---|---|
| OPA outer weight | § 3.6 , Eq. 11 | ||
| Uniformity outer weight | § 3.6 , Eq. 11 | ||
| GTL outer weight | § 3.6 , Eq. 11 | ||
| OPA axis-margin | § 3.2 , Eq. 4 | ||
| GTL small-transform penalty | § 3.4 , Eq. 8 | ||
| GTL angular/Euclidean mixing | § 3.4 , Eq. 8 |
| Configuration | DSv2 | CDFv3 | MEAN |
|---|---|---|---|
| LN tier | |||
| + OPA + uniformity + GTL ( Full method , anchor) | 81.33 | 83.37 | 91.97 ( ) |
| Full method, no hard RR negatives | 81.80 | 81.26 | 91.76 ( ) |
| Full method, no small-transform penalty | 82.03 | 81.25 | 91.69 ( ) |
| GSD-LoRA tier (seed 42 only) | |||
| GSD-LoRA, per-pair (anchor) | 83.08 | 83.77 | 92.48 ( ) |
| Value | Epoch | FF-FSH | MEAN-17 |
|---|---|---|---|
| Hinge margin (tier: LN; default ) | |||
| 0.5 | ep 09 | 88.55 | 90.72 |
| 1.0 | ep 09 | 88.35 | 91.01 |
| 2.0 | ep 09 | 87.65 | 90.71 |
| 3.0 | ep 09 | 87.80 | 90.58 |
| Small-transform penalty (tier: LN; default ) | |||
| Run | FFSH | CDFv1 | CDFv2 | CDFv3 | DeFo | DFD | DSv1 | DSv2 | DFDCP | DFDC | DFDM | FAVC | FFIW | IDFv1 | KoDF | PGF | UADFV | Unseen |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LN (s42, epoch18) | 88.86 | 94.23 | 93.96 | 83.49 | 97.23 | 98.24 | 86.30 | 81.15 | 87.28 | 86.30 | 99.88 | 97.97 | 90.52 | 97.12 | 92.13 | 91.24 | 97.83 | 91.98 |
| LN (s137, epoch24) | 88.92 | 94.23 | 94.27 | 84.28 | 97.39 | 98.09 | 86.30 | 81.63 | 87.60 | 85.89 | 99.89 | 97.88 | 90.27 | 96.99 | 91.96 | 91.45 | 97.96 | 92.06 |
| LN (s1729, epoch09) | 88.83 | 94.52 | 93.81 | 82.32 | 97.72 | 98.12 | 85.83 | 81.20 | 87.55 | 86.39 | 99.83 | 97.95 | 90.40 | 96.99 | 91.37 | 90.39 | 98.29 | 91.85 |
| LN mean ( ) | 88.87 0.04 | 94.33 0.17 | 94.01 0.24 | 83.37 0.99 | 97.45 0.25 | 98.15 0.08 | 86.14 0.27 | 81.33 0.26 | 87.47 0.17 | 86.19 0.27 | 99.87 0.03 | 97.93 0.05 | 90.39 0.12 | 97.04 0.08 | 91.82 0.40 | 91.03 0.56 | 98.03 0.24 | 91.97 0.10 |
| LoRA (s42, epoch04) | 92.58 | 94.27 | 93.57 | 84.02 | 99.83 | 97.05 | 88.29 | 82.75 | 91.68 | 87.02 | 99.87 | 98.30 | 87.88 | 97.06 | 89.07 | 95.24 | 97.75 | 92.72 |
| LoRA (s137, epoch05) | 90.40 | 92.49 | 92.62 | 85.31 | 99.32 | 96.65 | 89.48 | 83.07 | 91.16 | 84.98 | 99.61 | 97.79 | 87.44 | 97.09 | 89.17 | 94.50 | 97.04 | 92.24 |
| Run | FFSH | CDFv1 | CDFv2 | CDFv3 | DeFo | DFD | DSv1 | DSv2 | DFDCP | DFDC | DFDM | FAVC | FFIW | IDFv1 | KoDF | PGF | UADFV | Unseen |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LN (s42, epoch04) | 88.07 | 92.87 | 92.31 | 77.63 | 96.83 | 97.00 | 83.77 | 80.23 | 88.50 | 85.79 | 99.63 | 97.33 | 88.41 | 95.38 | 89.75 | 85.69 | 98.67 | 90.46 |
| LN (s137, epoch05) | 87.68 | 93.80 | 92.45 | 78.70 | 97.25 | 97.59 | 83.77 | 80.06 | 88.69 | 86.11 | 99.77 | 97.52 | 89.05 | 95.70 | 90.02 | 87.22 | 98.67 | 90.83 |
| LN (s1729, epoch03) | 88.62 | 92.49 | 91.64 | 75.64 | 97.52 | 96.42 | 82.42 | 80.28 | 88.92 | 85.61 | 99.41 | 96.69 | 87.96 | 94.58 | 88.75 | 83.71 | 98.83 | 89.97 |
| LN mean ( ) | 88.12 0.47 | 93.05 0.68 | 92.13 0.43 | 77.32 1.55 | 97.20 0.34 | 97.00 0.59 | 83.32 0.78 | 80.19 0.11 | 88.70 0.21 | 85.84 0.25 | 99.61 0.18 | 97.18 0.43 | 88.47 0.55 | 95.22 0.58 | 89.50 0.67 | 85.54 1.76 | 98.72 0.10 | 90.42 0.43 |
| LoRA (s42, epoch04) | 92.58 | 94.27 | 93.57 | 84.02 | 99.83 | 97.05 | 88.29 | 82.75 | 91.68 | 87.02 | 99.87 | 98.30 | 87.88 | 97.06 | 89.07 | 95.24 | 97.75 | 92.72 |
| LoRA (s137, epoch05) | 90.40 | 92.49 | 92.62 | 85.31 | 99.32 | 96.65 | 89.48 | 83.07 | 91.16 | 84.98 | 99.61 | 97.79 | 87.44 | 97.09 | 89.17 | 94.50 | 97.04 | 92.24 |
| Run | FFSH | CDFv1 | CDFv2 | CDFv3 | DeFo | DFD | DSv1 | DSv2 | DFDCP | DFDC | DFDM | FAVC | FFIW | IDFv1 | KoDF | PGF | UADFV | Unseen |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LN (s42, epoch15) | 88.66 | 94.10 | 93.84 | 83.44 | 97.15 | 98.23 | 86.31 | 81.02 | 87.31 | 86.18 | 99.88 | 97.95 | 90.41 | 97.12 | 92.02 | 91.03 | 97.88 | 91.91 |
| LN (s137, epoch23) | 88.92 | 94.23 | 94.26 | 84.26 | 97.40 | 98.09 | 86.29 | 81.63 | 87.64 | 85.90 | 99.89 | 97.88 | 90.26 | 96.98 | 91.95 | 91.44 | 97.96 | 92.06 |
| LN (s1729, epoch23) | 88.80 | 94.74 | 94.25 | 83.43 | 97.69 | 98.17 | 86.11 | 80.97 | 87.05 | 85.74 | 99.89 | 97.98 | 90.34 | 97.24 | 91.92 | 91.55 | 98.21 | 92.01 |
| LN mean ( ) | 88.79 0.13 | 94.35 0.34 | 94.12 0.24 | 83.71 0.48 | 97.42 0.27 | 98.16 0.07 | 86.24 0.11 | 81.20 0.37 | 87.33 0.30 | 85.94 0.22 | 99.89 0.01 | 97.94 0.05 | 90.34 0.07 | 97.11 0.13 | 91.97 0.05 | 91.34 0.28 | 98.01 0.17 | 91.99 0.07 |
| LoRA (s42, epoch04) | 92.58 | 94.27 | 93.57 | 84.02 | 99.83 | 97.05 | 88.29 | 82.75 | 91.68 | 87.02 | 99.87 | 98.30 | 87.88 | 97.06 | 89.07 | 95.24 | 97.75 | 92.72 |
| LoRA (s137, epoch04) | 88.38 | 92.02 | 92.07 | 83.88 | 99.12 | 96.80 | 89.43 | 81.88 | 91.69 | 85.79 | 99.76 | 98.35 | 85.20 | 97.13 | 88.43 | 94.62 | 96.83 | 91.85 |
| Run | FFSH | CDFv1 | CDFv2 | CDFv3 | DeFo | DFD | DSv1 | DSv2 | DFDCP | DFDC | DFDM | FAVC | FFIW | IDFv1 | KoDF | PGF | UADFV | Unseen |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LN (s42, epoch19) | 88.92 | 94.27 | 93.97 | 83.55 | 97.23 | 98.16 | 86.28 | 80.93 | 87.35 | 86.02 | 99.89 | 97.98 | 90.38 | 97.10 | 92.08 | 91.33 | 98.04 | 91.97 |
| LN (s137, epoch22) | 88.93 | 94.27 | 94.26 | 84.24 | 97.42 | 98.08 | 86.26 | 81.57 | 87.65 | 85.92 | 99.89 | 97.87 | 90.27 | 96.97 | 91.94 | 91.41 | 98.04 | 92.06 |
| LN (s1729, epoch13) | 88.79 | 94.65 | 94.11 | 82.85 | 97.85 | 98.19 | 86.20 | 81.27 | 87.16 | 86.08 | 99.84 | 98.03 | 90.54 | 97.22 | 91.75 | 91.41 | 98.25 | 92.01 |
| LN mean ( ) | 88.88 0.08 | 94.40 0.22 | 94.12 0.15 | 83.55 0.70 | 97.50 0.32 | 98.14 0.05 | 86.25 0.04 | 81.26 0.32 | 87.39 0.25 | 86.01 0.08 | 99.87 0.03 | 97.96 0.08 | 90.40 0.14 | 97.10 0.13 | 91.92 0.17 | 91.38 0.05 | 98.11 0.12 | 92.01 0.05 |
| LoRA (s42, epoch04) | 92.58 | 94.27 | 93.57 | 84.02 | 99.83 | 97.05 | 88.29 | 82.75 | 91.68 | 87.02 | 99.87 | 98.30 | 87.88 | 97.06 | 89.07 | 95.24 | 97.75 | 92.72 |
| LoRA (s137, epoch07) | 89.17 | 89.69 | 89.46 | 82.98 | 99.61 | 95.33 | 90.00 | 82.25 | 89.63 | 84.64 | 99.75 | 98.04 | 85.19 | 96.25 | 87.63 | 94.93 | 97.67 | 91.31 |
Deepfake Detection
Deepfake detection identifies manipulated or AI-generated media to support trustworthy communication, identity verification, and forensic analysis.
See all 259 papers