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.
Figures & tables
Component
Prior work (INV-REG) [ 22 ]
IIL
Step 1: REx [ 15 ]
✓
✓
Step 2: IRMv1 [ 2 ]
✓
✓
Classification loss Lcls
Supcon. [ 14 ]
BCE
Partition level
Identity
Gender × Skin-tone
K (partitions)
2
6
FPR parity loss
×
✓
Table 1: IIL components compared with prior invariant-learning work [ 22 ] . ✓ = present; × = absent.
Detector
Attr.
Metric
Xception [ 5 ]
EffB4 [ 30 ]
F3Net [ 26 ]
SPSL [ 19 ]
SRM [ 21 ]
UCF [ 36 ]
CORE [ 24 ]
DAW [ 13 ]
DAG [ 13 ]
PG-FDD [ 17 ]
IIL (ours)
Skin Tone
FMEO ↓
8.836
8.300
19.938
8.055
10.002
17.325
10.779
14.118
6.551
6.465
0.730
FDP ↓
9.751
6.184
12.876
9.379
10.897
12.581
10.317
10.706
8.617
9.746
8.468
FOAE ↓
1.271
4.377
2.818
1.135
0.915
1.883
1.332
1.667
1.388
0.882
0.770
FEO ↓
12.132
11.062
23.708
9.789
14.239
21.920
13.069
16.604
7.383
9.115
1.537
Gender
FMEO ↓
3.975
5.385
4.717
4.411
6.271
5.074
5.795
5.510
5.910
3.190
0.566
Table 2: Fairness and utility results on the AI-Face test set. Benchmark values from Table 4 of [ 18 ] . All values in %. ↓ lower is better; ↑ higher is better. Bold = best across all 11 detectors. EffB4 = EfficientNet-B4; DAW = DAW-FDD; DAG = DAG-FDD. Gray column = IIL (ours).
Figure 1: Per-group FPR (left) and TPR (right) across the six gender × skin-tone demographic partitions (Female/Male × Skin-Tone 1–3) for IIL (ours), EfficientNet-B4, DAG-FDD, and PG-FDD. IIL achieves near-uniform FPR across all six gender × skin-tone groups (max 3.4%, range 1.5 pp) compared to EffB4 (max 33.0%, range 24.1 pp), DAG (max 22.8%), and PG-FDD (max 16.9%), while maintaining TPR above 97% for every group. The dominant disparity axis in baselines is skin tone; gender contributes a secondary gap within each skin-tone tier.
Figure 2: Fairness–utility analysis.
Components
Variant
Balanced Sampler
FPR Loss
IRM/ REx
AUC ↑
FPR ↓
inter-FMEO ↓
skin-FMEO ↓
S↓
A
×
×
×
98.23
12.18
17.18
6.58
54.12
B
✓
×
×
97.98
12.74
14.32
5.63
44.43
C †
✓
✓
×
93.94
100.0
—
—
—
D
✓
✓
✓
98.26
9.15
6.49
1.98
24.43
E ( Pfix )
✓
✓
✓
98.22
9.68
6.97
3.39
26.62
Table 3: Component ablation (15-epoch run, 50 000 training samples, paper hyperparameters). ✓ = enabled; × = disabled; P fix = fixed identity partition instead of learned. S = fairness score (Eq. 3 ; ↓ ). Bold = best non-collapsed value. † collapsed: FPR=100% , all images predicted fake.
Figure 3: UMAP [ 23 ] embeddings for EffB4, DAG, PG-FDD, and IIL (ours). Top row : colored by Real/Fake label. EffB4 spreads into multiple elongated thread-like strands; DAG produces two large compact blobs; PG-FDD yields serpentine filamentous curves; IIL achieves the tightest compact clusters with the sharpest Real/Fake boundary. All four models achieve clear Real/Fake separation, consistent with AUC values above 98%. Bottom row : colored by intersectional demographic group.
Figure 4: GradCAM [ 29 ] attention maps across the six gender × skin-tone groups and four models. Left block: real images; right block: fake images.
Celeb-DF
DFD
DFDC
FF++
Model
AUC ↑
FPR ↓
FMEO ↓
AUC ↑
FPR ↓
FMEO ↓
AUC ↑
FPR ↓
FMEO ↓
AUC ↑
FPR ↓
FMEO ↓
Xception [ 5 ]
92.85
28.75
87.50
85.42
31.40
16.80
75.24
13.56
61.35
86.57
17.33
82.18
EffB4 [ 30 ]
71.85
85.34
21.74
72.98
87.11
16.69
64.90
55.06
81.71
71.55
35.04
75.79
F3Net [ 26 ]
89.06
9.26
93.33
81.82
16.30
23.75
73.54
9.03
47.46
81.84
6.03
49.89
SPSL [ 19 ]
93.46
14.47
87.50
84.25
51.62
10.18
76.66
15.16
68.12
90.66
17.12
69.09
SRM [ 21 ]
87.64
5.40
37.50
84.60
19.76
12.91
81.67
16.49
71.43
81.75
26.22
59.49
Table 4: Generalization across datasets. Evaluated on AI-Face demographically annotated subsets [ 18 ] (8K–36K images per dataset). AUC ↑ and FPR ↓ (%); FMEO ↓ (%) is the intersectional Max Equalized Odds gap across the six gender × skin-tone groups, as defined in § 3.1 . Bold = best per metric.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Detector
Attr.
Metric
Xception
EffB4
F3Net
SPSL
SRM
UCF
CORE
DAW
DAG
PG-FDD
IIL
Skin Tone
FMEO ↓
8.851
8.276
19.935
8.353
9.998
17.321
10.775
14.122
6.555
6.468
0.730
FDP ↓
9.759
6.195
12.876
9.479
10.894
12.582
10.318
10.707
8.618
9.747
8.468
FOAE ↓
1.270
4.367
2.814
1.163
0.911
1.878
1.329
1.669
1.388
0.881
0.770
FEO ↓
12.144
11.061
23.703
10.049
14.233
21.921
13.068
16.607
7.389
9.120
1.537
Gender
FMEO ↓
3.978
5.366
4.707
4.488
6.264
5.076
5.805
5.507
5.912
3.188
0.566
Appendix
Table 5: Full-test-split re-evaluation of all ten baselines and IIL, same layout as Table 2. Bold = best per row. Deviation from Table 2’s published baseline values (210 cells: 10 baselines × 21 metrics) is rounding-level throughout: max ∣Δ∣ = 0.772 pp (Intersection FEO, SPSL), mean ∣Δ∣ = 0.026 pp overall. Per-attribute max/mean ∣Δ∣ : Skin Tone 0.298/0.021, Gender 0.077/0.009, Age 0.615/0.053, Intersection 0.772/0.036 (N=40 each), Utility 0.027/0.003 (N=50). No cell deviates by more than 0.8 pp, confirming rounding-level agreement with the published values.
Figure 5: FPR (left) and fairness score S (right) across the full 8×5 grid at K=6 /20k samples/8 epochs. Black box = selected point (λIRM=0.07,λREx=5.0) .
(λIRM,λREx)
S↓
AUC ↑
ACC ↑
FPR ↓
inter
skin
gen
age
(0.07, 5.00)
45.5
95.49
90.58
18.28
10.89
5.09
4.96
8.62
(0.06, 3.00)
50.3
96.24
90.94
18.90
11.71
5.70
3.58
11.86
(0.05, 4.00)
52.3
96.32
91.10
14.18
13.14
6.52
1.11
11.87
(0.01, 2.00)
55.6
95.93
90.48
16.17
13.65
7.92
2.66
9.82
(0.02, 5.00)
58.1
95.44
90.66
19.75
14.27
7.10
3.79
11.55
(0.01, 4.00)
66.6
96.22
91.40
18.23
23.08
4.48
2.58
8.89
Appendix
Table 6: Full 8×5 sweep grid at K=6 /20k samples/8 epochs, all 40 combinations, sorted by fairness score S (Eq. 3; bold = selected point, lowest S ). AUC/ACC/FPR in %; FMEO columns in % (inter/skin/gen/age, left to right).
Figure 6: IIL training curves over 15 epochs ( λIRM=0.07 , λREx=5.0 , K=6 ). Left: FPR. Centre: inter- and skin-FMEO. Right: AUC. Star = best checkpoint (epoch 13, S=6.09 ).
Figure 7: Variant C (FPR+DP loss, no IRM/REx) at small (50k samples) and full deployed (300k samples) scale, per-epoch over 15 epochs. Left: FPR — collapsed (100%) at every epoch, at both scales. Right: AUC — rises with training at both scales despite the FPR collapse, since AUC measures ranking quality of raw scores, not the collapsed classification threshold. Fairness score S and FMEO are omitted: both are trivially 0 under complete FPR collapse (uniform failure, not a real parity result), matching Table 3’s “—” convention for this variant.
Partition
Child
Young
Adult
Middle
Senior
F, skin 1
9.18
35.11
42.61
10.66
2.44
F, skin 2
4.98
25.69
48.78
18.38
2.18
F, skin 3
4.58
20.01
53.06
18.74
3.60
M, skin 1
8.34
20.25
44.94
20.45
6.02
M, skin 2
2.38
10.08
49.08
33.76
4.70
M, skin 3
1.28
8.18
50.92
35.52
4.10
Appendix
Table 7: Row-normalized Age-bracket composition (%) within each of the six gender × skin-tone partitions used by LFPR / LDP /IRM. Each row sums to 100%.
Model
Intersection (6-way)
Skin (3-way)
Gender (2-way)
Majority-class baseline
17.0%
33.9%
50.7%
EffB4
55.1 ± 1.0%
75.6 ± 0.7%
71.7 ± 1.3%
DAG-FDD
37.7 ± 1.1%
62.9 ± 1.2%
61.5 ± 2.0%
PG-FDD
22.1 ± 1.2%
41.9 ± 1.0%
54.4 ± 1.5%
IIL (ours)
41.1 ± 0.6%
60.1 ± 0.8%
66.3 ± 0.9%
Appendix
Table 8: Demographic-group classifiability from learned features (accuracy, mean ± std over 5 folds; lower = more invariant). Majority-class baseline given for reference.
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 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.