Deepfake detectors routinely achieve high aggregate accuracy while exhibiting substantial demographic disparities in their errors: false-positive rates differ by double digits across gender and skin tone. Most existing remedies require either full retraining with a fairness-aware objective or demographic-annotated training data, both impractical when only a pretrained checkpoint is available, while label-free post-hoc calibrators leave the detector's weights unchanged. We introduce FYC (FPR-disparity Weight Perturbation and Youden Calibration), a two-stage post-hoc method that perturbs a detector's weights under an FPR-disparity criterion, then calibrates a separate decision threshold for each demographic group - requiring no architectural changes and no full retraining. Evaluated against every detector benchmarked on AI-Face v2 across six intersectional gender x skin-tone groups, FYC outperforms the field on 18 of 21 reported fairness and utility metrics, including purpose-built fairness architectures, at a fraction of the training cost. We further identify why the choice of fairness criterion is critical: the original Demographic Parity objective used in prior weight-perturbation work is architecture-dependent and unpredictable, improving two of four evaluated architectures, leaving a third largely unchanged, and substantially degrading the fourth (PG-FDD), with no way to anticipate the outcome for a given checkpoint in advance. Conditioning the same objective on real images only - our FPR-disparity criterion - eliminates this unpredictability, improving fairness and accuracy jointly across all four architectures with no exceptions.
Figures & tables
Method
Category
Label-free
Weight-free
DAG-FDD [ 12 ]
In-processing
✓
×
DAW-FDD [ 12 ]
In-processing
×
×
PG-FDD [ 13 ]
In-processing
×
×
DAID [ 2 ]
In-processing
×
×
Weight Perturbation [ 11 ]
Post-hoc weight adapt.
×
×
BMFT [ 31 ]
Post-hoc weight adapt.
×
×
Table 1: Landscape of fair deepfake detection methods. Label-free: no demographic labels required at any stage. Weight-free: no model weight updates.
Figure 1: Method overview. Stage 1 fine-tunes a pretrained detector with a weight-perturbation ascent/descent loop, using FPR-disparity (real images only) as the disparity criterion driving the ascent—the criterion we adopt after diagnosing it as reliable across all four architectures we test, unlike the Demographic Parity alternative (§ 4.4 ). Stage 2 applies zero-cost per-group Youden-J calibration (§ 3.3 ) on top of the resulting checkpoint, producing the final calibrated model.
Xcep [ 3 ]
EffB4 [ 25 ]
ViT [ 5 ]
F3N [ 23 ]
SPSL [ 16 ]
SRM [ 18 ]
UCF [ 32 ]
UnivFD [ 22 ]
CORE [ 21 ]
DAW [ 12 ]
DAG [ 12 ]
PG [ 13 ]
X+FYC
D+FYC
Skin
FMEO ↓
8.84
8.30
6.26
19.94
8.06
10.00
17.33
2.58
10.78
14.12
6.55
6.47
3.34
1.24
FDP ↓
9.75
6.18
7.73
12.88
9.38
10.90
12.58
8.56
10.32
10.71
8.62
9.75
9.48
8.14
FOAE ↓
1.27
4.38
2.17
2.82
1.14
0.92
1.88
2.75
1.33
1.67
1.39
0.88
0.73
1.19
FEO ↓
12.13
11.06
8.81
23.71
9.79
14.24
21.92
5.54
13.07
16.60
7.38
9.12
4.48
2.24
Gender
FMEO ↓
3.98
5.39
5.10
4.72
4.41
6.27
5.07
4.50
5.80
5.51
5.91
3.19
0.45
0.35
FDP ↓
1.69
1.73
1.34
1.86
1.83
1.96
1.74
1.19
2.15
2.02
2.15
1.25
0.96
0.80
Table 2: Full comparison to all 12 detectors benchmarked on AI-Face v2 [ 14 ] (their Table 4, reproduced in full, with per-detector citations) plus our two best FYC results, each a single run (§ 4.2 ). FMEO/FDP/FOAE/FEO per demographic axis and Intersectional; Utility rows below, including per-epoch training time from [ 14 ] . Bold = best value per row across all 14 detectors. FYC wins 18 of 21 fairness/utility rows (Time/Epoch excluded from this tally); all three exceptions are FDP (§ 4.2 ). X+FYC = Xception+FYC; D+FYC = DAG-FDD+FYC.
Metric
Xception
EfficientNet-B4
DAG-FDD
PG-FDD
Base
+ FYC
Base
+ FYC
Base
+ FYC
Base
+ FYC
Intersectional ( ↓ )
FMEO
10.51
6.52 ± 1.21
17.59
8.83 ± 0.33
12.18
3.65 ± 0.65
9.58
5.76 ± 1.23
EO
24.32
8.89 ± 1.04
25.11
11.03 ± 0.48
26.30
6.17 ± 0.61
18.35
10.39 ± 2.93
FOAE
2.54
1.90 ± 0.11
8.46
6.57 ± 0.13
4.04
1.78 ± 0.05
3.06
3.88 ± 0.63
FPR
12.96
1.48 ± 0.07
20.07
3.39 ± 0.14
14.84
1.26 ± 0.06
10.97
2.39 ± 0.20
Table 3: Fairness metrics: Baseline vs. FYC ( + FYC), N=329,309 . + FYC columns are mean ± std over 3 seeds for all four architectures. Red : worse. Teal : better. Ul : best baseline per row. Per-axis FDP/FOAE/FEO breakdown for the two winning architectures is in Table 2 .
Model / Setting
AUC ↑
Acc ↑
FPR ↓
EER ↓
ERM baselines
Xception
98.58
96.31
12.96
5.15
Xception + FYC
99.87 ± 0.01
98.77 ± 0.11
1.48 ± 0.07
1.27 ± 0.08
EfficientNet-B4
98.61
94.20
20.07
6.69
EfficientNet + FYC
99.46 ± 0.01
96.44 ± 0.11
3.39 ± 0.14
3.50 ± 0.04
Fairness-aware baselines
Table 4: Utility metrics: Baseline vs. FYC. + FYC rows are mean ± std over 3 seeds for all four architectures. Red : degradation. Teal : improvement. Underline : best baseline per column.
Model
Criterion
Baseline
Raw
FMEO ↓
AUC ↑
FMEO ↓
AUC ↑
Xception
DP
10.51
98.58
5.63
99.89
FPR-disp. (ours)
3.01
99.85
EffB4
DP
17.59
98.61
13.75
99.52
FPR-disp. (ours)
11.88
99.47
DAG-FDD
DP
12.18
98.77
3.10
99.87
Table 5: DP diagnostic: baseline → raw (pre-calibration) weight perturbation, DP [ 11 ] vs. FPR-disparity (ours), same single seed (seed 1) both rows for a like-for-like comparison. This motivated adopting FPR-disparity as FYC’s default criterion. Red : worse than baseline. Teal : better.
Figure 2: Raw (pre-calibration) intersectional FMEO at every Stage 1 epoch, DP vs. FPR-disparity, for all four architectures (50k-image random subsample of the test set, global-Youden threshold recomputed per epoch). DP is stable for Xception and DAG-FDD, spikes once for EfficientNet-B4, and oscillates without converging for PG-FDD; FPR-disparity is stable for all four.
Model / Setting
AUC ↑
Acc ↑
FMEO ↓
FPR ↓
ERM baselines
Xception
98.58
96.31
10.51
12.96
Xception + Cal-only
98.33
96.12
9.56
7.39
Xception + FYC
99.87 ± 0.01
98.77 ± 0.11
6.52 ± 1.21
1.48 ± 0.07
EfficientNet-B4
98.61
94.20
17.59
20.07
EfficientNet + Cal-only
98.69
93.69
9.87
6.76
Table 6: Calibration-only vs. FYC. “Cal-only” = Youden-J applied to the pretrained checkpoint (no weight perturbation). + FYC rows are mean ± std over 3 seeds; Cal-only rows are single runs. Bold = best value between Cal-only and FYC in each block.
Figure 3: Fairness–utility scatter (AUC vs. Intersectional FMEO). Filled circles = baseline; filled squares = FPR-disparity FYC; hollow diamonds = calibration-only (no weight perturbation). Arrows connect baseline to FYC; ideal direction is lower-right (higher AUC, lower FMEO).
Figure 4: Per-group FPR across the six intersectional groups (Female/Male × Light/Medium/Dark skin tone) for Xception, train-seed 1. Gray = pretrained baseline; red = raw Stage 1 (FPR-disparity, pre-calibration); blue = FYC (Stage 1 + Stage 2). Calibration further reduces FPR for five of six groups; M-Light is the one group where per-group Youden-J raises FPR relative to the raw checkpoint in this particular seed, illustrating the per-seed variance behind Table 3 ’s 3-seed means.
Deepfake detectors show large performance gaps across demographic groups. Existing fairness approaches require demographic labels, retraining, or sacrifice accuracy. We introduce Face-Fairness (FF), a plug-and-play framework for bias mitigation. Our primary contribution, Face-Feature Tuning (FFT), is the first demographic label-free fairness method demonstrated for deepfake detection: a lightweight calibrator that performs a logit remapping conditioned on frozen face embeddings. We complement FFT with two variants: FF-Max, which maximizes worst-group accuracy when demographics are available, and FF-Discover, which does the same with embedding-discovered groups. Across in-domain and cross-dataset test settings, FF consistently reduces FPR/TPR gaps and improves minimum group accuracy while maintaining (often improving) overall accuracy. The approach is detector-agnostic, adds negligible runtime overhead, and requires no access to identity attributes.
Although recent deepfake detectors achieve high overall accuracy, their errors remain unevenly distributed across demographic subgroups, with real faces from certain groups more often misclassified as fake. Existing fairness-aware detectors typically regularise the entire feature representation, without identifying or controlling the specific components that drive unfair predictions. Such coarse intervention can over-suppress useful forgery cues while leaving demographic structure in component-specific subspaces. To address this, we identify two subgroup-sensitive components: multi-scale spatial features, which encode local facial and forgery patterns, and fine-tuning-induced residual features, which adapt the backbone to the unfair training distribution. We propose FairReL, a fairness-aware representation-learning framework that targets both components with dedicated demographic supervision. FairReL uses an SVD-decomposed foundation-model backbone to isolate the fine-tuning-induced residual representation, and introduces two complementary losses. Group-Conditional Wavelet Decorrelation (GCWD) suppresses subgroup-imbalanced structure across spatial wavelet sub-bands, while Subspace-Localised Mean Alignment (SLMA) aligns subgroup means within each real/fake class in the residual representation. Experiments on FF++, Celeb-DF, DFD and DFDC show that, against the state-of-the-art fairness-aware detector, FairReL improves unseen-dataset AUC by 3.9% while reducing subgroup FPR disparity by 10.2%. Code is available at https://github.com/xiaoman89/FairReL .
Xiaoman Lu, Jiaqi Li, Shuntian Zheng +2
Department of Computer Science University of Warwick Coventry, United Kingdom · School of Computer Science University of Birmingham Birmingham, United Kingdom
Deepfake detection models achieve high overall accuracy but exhibit large performance disparities across demographic groups, causing some groups to experience substantially higher false-positive rates than others. Existing fairness-aware methods address this by constraining model outputs to equalize group predictions, but do not tackle the underlying feature-level bias. As a result, models may still rely on demographic characteristics, such as skin texture, as indirect cues for detecting manipulated content. To address this problem, we propose Intersectional Invariant Learning (IIL), a framework that improves feature learning rather than only adjusting model outputs. IIL alternates between identifying the most challenging demographic partitions and learning features that remain consistent across those partitions. The framework also includes a false-positive-rate fairness objective and group-balanced sampling across six gender x skin-tone intersectional groups. We evaluate IIL on a large-scale fairness benchmark containing over one million images, 10 baseline detectors, and 21 fairness and utility metrics. Results show that IIL achieves Pareto-optimal performance, simultaneously attaining an AUC above 99%, a false-positive rate under 2.5%, and the lowest intersectional fairness disparity (FMEO) among all compared methods. IIL outperforms prior fairness-aware methods on the large majority of the 21 evaluation metrics, while reducing both false-positive rates and intersectional fairness disparities by several-fold compared with the strongest competing method.