FairReL: Deepfake Detection using Fairness-Aware Representation Learning

Xiaoman Lu; Jiaqi Li; Shuntian Zheng; Huiping Chen; Yu Guan

cs.CVAug 28, 2026

FairReL: Deepfake Detection using Fairness-Aware Representation Learning

Authors: Xiaoman Lu, Jiaqi Li, Shuntian Zheng, Huiping Chen, Yu Guan

Organizations: Department of Computer Science University of Warwick Coventry, United Kingdom · School of Computer Science University of Birmingham Birmingham, United Kingdom

Abstract

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 .

Figures & tables

Figure 1: False positive rate (the fraction of real faces misclassified as fake, the lower the better) across six demographic subgroups for a vanilla Xception detector [ 4 ] , evaluated at a single shared threshold. The Female-Black group shows the highest rate.

Figure 2: Training pipeline of FairReL. The backbone is decomposed into a frozen principal part $\mathbf{W}_{p}$ and a trainable residual part $\Delta\mathbf{W}$ . Given an input image, the standard path produces a spatial feature map $\mathbf{F}$ and a pooled prediction feature $\mathbf{z}$ , while the residual-only path produces $\mathbf{z}_{\Delta}$ . (A) GCWD applies a wavelet transform to $\mathbf{F}$ and reduces subgroup differences in each frequency band. (B) SLMA aligns subgroup means in $\mathbf{z}_{\Delta}$ within each real/fake class, so that the residual update is less subgroup-dependent. Only $\mathbf{z}$ is fed to the classifier for real/fake prediction. At inference, only the backbone and the classifier are retained.

Figure 3: GCWD computation pipeline. (1) The spatial feature map $\mathbf{F}$ is decomposed by DWT into frequency sub-bands, each describing a different scale and direction of spatial variation. (2) GCWD computes the average energy of each sub-band within every demographic subgroup and measures the cross-subgroup discrepancy $D_{\ell}$ . (3) A detached softmax converts $\{D_{\ell}\}$ into adaptive weights $\{w_{\ell}\}$ , so the loss focuses on the sub-bands with the strongest subgroup imbalance. $\mathcal{L}_{\mathrm{GCWD}}$ reduces these imbalanced frequency components.

Type	Method	Backbone	Celeb-DF				DFD				DFDC
			Detection (%)		Fairness (%)		Detection (%)		Fairness (%)		Detection (%)		Fairness (%)
			AUC $\uparrow$	AUC ${}_{\mathrm{avg}}\uparrow$	$F_{\mathrm{FPR}}\downarrow$	$F_{\mathrm{MEO}}\downarrow$	AUC $\uparrow$	AUC ${}_{\mathrm{avg}}\uparrow$	$F_{\mathrm{FPR}}\downarrow$	$F_{\mathrm{MEO}}\downarrow$	AUC $\uparrow$	AUC ${}_{\mathrm{avg}}\uparrow$	$F_{\mathrm{FPR}}\downarrow$	$F_{\mathrm{MEO}}\downarrow$
ERM	Xception Chollet (2017)	Xception	70.94	70.75	25.43	16.81	76.73	76.57	29.20	18.11	60.58	60.45	48.89	37.15
ERM	CLIP-ERM Radford et al. (2021)	CLIP	74.37	74.26	23.25	15.71	81.14	81.06	31.54	20.69	62.35	62.30	51.15	39.46
Gen.	ForAda Cui et al. (2025)	CLIP	80.21	80.26	20.35	12.75	85.50	85.53	27.41	23.46	64.82	64.87	49.70	38.61
	Effort Yan et al. (2025)	CLIP	78.52	78.58	19.61	13.44	84.22	84.26	19.37	15.08	65.08	65.14	52.04	40.31

Table 1: Cross-domain comparison on Celeb-DF, DFD and DFDC after training on FF++. “Type” groups methods into ERM baselines, generalisation-oriented detectors (Gen.), and fairness-aware detectors (Fair.).

\uparrow

\downarrow

indicate higher/lower is better. The best results are shown in bold and the second-best are underlined .

		Per-subgroup AUC (%) $\uparrow$						Overall (%)
Type	Method	M-W	M-A	M-B	F-W	F-A	F-B	AUC $\uparrow$	AUC ${}_{\mathrm{avg}}\uparrow$	AUC gap $\downarrow$
ERM	Xception Chollet (2017)	61.06	60.69	60.28	61.56	59.78	59.34	60.58	60.45	2.22
ERM	CLIP-ERM Radford et al. (2021)	62.87	62.53	62.14	63.34	61.67	61.26	62.35	62.30	2.08
Gen.	ForAda Cui et al. (2025)	65.38	65.07	64.73	65.80	64.31	63.94	64.82	64.87	1.86
	Effort Yan et al. (2025)	65.64	65.34	65.00	66.06	64.59	64.22	65.08	65.14	1.84
	GenD Yermakov et al. (2026)	67.93	67.19	66.59	67.86	65.91	65.77	66.85	66.88	2.16

Table 2: Per-subgroup AUC (%) on DFDC (trained on FF++).

Figure 4: Per-subgroup FPR (%) of GenD Yermakov et al. (2026) and FairReL.

Figure 5: Frequency-wise diagnostics of subgroup imbalance and forgery separability. (a) Sub-band diagnostic statistics. (b) Sub-band discrepancy before and after FairReL.

Figure 6: Mechanistic diagnostics on DFDC. (a) Principal-vs-residual comparison of subgroup imbalance across spatial-frequency and pooled-feature statistics. (b) PCA visualisation of residual subgroup centroids after subgroup-balanced/imbalanced fine-tuning.

	FF++		Celeb-DF		DFD		DFDC
Variant	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$
w/o subspace decomposition	95.37	12.83	79.12	11.87	84.91	15.26	64.79	40.47
FairReL (w/ decomposition)	96.14	10.97	80.64	10.12	86.13	13.55	66.34	38.36

Table 3: Effect of subspace decomposition. Without decomposition, CLIP is fully fine-tuned; GCWD acts on

\mathbf{F}

and SLMA is applied to the full pooled feature

\mathbf{z}

Variant	Setting			FF++		Celeb-DF		DFD		DFDC
Variant	GCWD	SLMA	Cond.	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$
SVD baseline Yan et al. (2025)	–	–	–	94.83	14.57	78.52	19.61	84.22	19.37	65.08	52.04
GCWD only	✓	–	–	95.34	12.68	79.43	14.87	85.04	16.37	65.43	44.76
SLMA only	–	$\mathbf{z}_{\Delta}$	✓	95.12	13.04	79.08	15.31	84.77	16.83	65.24	45.58
GCWD + SLMA on $\mathbf{z}$	✓	$\mathbf{z}$	✓	95.62	11.88	79.73	12.39	85.24	15.12	65.72	41.19
GCWD + SLMA on $\mathbf{z}_{\Delta}$	✓	$\mathbf{z}_{\Delta}$	–	93.18	11.43	77.28	11.64	83.03	14.37	62.54	40.08

Table 4: Unified ablation of GCWD, SLMA and class conditioning across four datasets. “Cond.” indicates whether SLMA aligns subgroup means within each real/fake class.

	Celeb-DF			DFD			DFDC
Method on the same backbone	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	$F_{\mathrm{MEO}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	$F_{\mathrm{MEO}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	$F_{\mathrm{MEO}}\downarrow$
Subspace-decomposed baseline Yan et al. (2025)	78.52	19.61	13.44	84.22	19.37	15.08	65.08	52.04	40.31
+ prediction-level fairness loss Madras et al. (2018)	78.43	18.92	13.08	84.11	18.57	14.73	64.91	50.82	39.58
+ adversarial classifier (GRL) Ganin et al. (2016) ; Zhang et al. (2018)	77.88	18.23	12.64	83.62	18.08	14.12	64.23	49.71	38.87
+ full-feature mean alignment	78.82	18.03	12.38	84.43	17.82	13.76	66.02	49.24	38.41
FairReL (Ours)	80.64	10.12	7.03	86.13	13.55	8.98	66.34	38.36	30.02

Table 5: Comparison of alternative fairness interventions on the same subspace-decomposed backbone under cross-domain evaluation.

Figure 7: Grad-CAM attribution across the six subgroups on FF++ (left) and DFDC (right).

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Train	Val.	Test	Dataset role	Retained intersectional attributes
FF++ R”ossler et al. (2019)	45,377	15,126	15,126	Source domain	M-A, M-W, M-B, F-A, F-W, F-B
Celeb-DF Li et al. (2020b)	–	–	87,213	Unseen target	M-W, M-B, F-W, F-B
DFD Dufour and Gully (2019)	–	–	20,032	Unseen target	M-W, M-B, F-W, F-B
DFDC Dolhansky et al. (2020)	–	–	63,225	Unseen target	M-A, M-W, M-B, F-A, F-W, F-B

Appendix

Table S1: Dataset split and demographic coverage after excluding samples annotated as Others. M/F denote male/female, and A/W/B denote Asian/White/Black. FF++ is used as the source domain and is split into training, validation, and test sets with an approximate 60/20/20 ratio. Celeb-DF, DFD, and DFDC are used only as unseen target-domain test sets.

Dataset	M-A	M-W	M-B	F-A	F-W	F-B	Total
FF++ R”ossler et al. (2019)	2,475	31,281	1,468	8,013	31,281	1,111	75,629
Celeb-DF Li et al. (2020b)	–	81,194	600	–	5,389	30	87,213
DFD Dufour and Gully (2019)	–	7,784	6,482	–	4,127	1,639	20,032
DFDC Dolhansky et al. (2020)	2,144	21,755	9,603	1,915	18,502	9,306	63,225

Appendix

Table S2: Number of samples in each retained intersectional demographic group after excluding the Others category. The six groups are defined by gender

\{\mathrm{M},\mathrm{F}\}

and race

\{\mathrm{Asian},\mathrm{White},\mathrm{Black}\}

. “–” means that the subgroup does not exist in the dataset annotation.

Type	Method	Backbone	Req. Demo	FF++
Type	Method	Backbone	Req. Demo	AUC $\uparrow$	AUC ${}_{\mathrm{avg}}\uparrow$	$F_{\mathrm{FPR}}\downarrow$	$F_{\mathrm{MEO}}\downarrow$
ERM	Xception-ERM Chollet (2017)	Xception	No	88.50	88.46	16.84	12.07
ERM	CLIP-ERM Radford et al. (2021)	CLIP ViT-L/14	No	92.40	92.44	15.92	11.38
Gen.	ForAda Cui et al. (2025)	CLIP ViT-L/14	No	94.10	94.21	15.13	10.86
	Effort Yan et al. (2025)	CLIP ViT-L/14	No	94.83	94.92	14.57	10.41
	GenD Yermakov et al. (2026)	CLIP ViT-L/14	No	95.35	95.48	14.06	10.02

Appendix

Table S3: Intra-domain comparison on FF++ R”ossler et al. (2019) (trained and tested on FF++). Detection is measured by AUC and

\mathrm{AUC}_{\mathrm{avg}}

; fairness by

F_{\mathrm{FPR}}

and

F_{\mathrm{MEO}}

(%, lower is better), at the threshold pinned on the FF++ validation split at

\mathrm{TPR}{=}0.90

. “Req. Demo” denotes whether demographic labels are used during training. The best results are in bold and the second-best are underlined .

	Detection		Fairness
Dataset	AUC $\uparrow$	AUC ${}_{\mathrm{avg}}\uparrow$	$F_{\mathrm{FPR}}\downarrow$	$F_{\mathrm{MEO}}\downarrow$
FF++ (intra)	$96.14\pm 0.2$	$96.54\pm 0.2$	$10.97\pm 0.3$	$7.41\pm 0.3$
Celeb-DF	$80.64\pm 0.4$	$81.26\pm 0.3$	$10.12\pm 0.5$	$7.03\pm 0.4$
DFD	$86.13\pm 0.3$	$86.78\pm 0.3$	$13.55\pm 0.6$	$8.98\pm 0.5$
DFDC	$66.34\pm 0.4$	$67.03\pm 0.4$	$38.36\pm 0.6$	$30.02\pm 0.5$

Appendix

Table S4: Mean

\pm

standard deviation of FairReL over three seeds. Variation is small relative to the gaps to competing methods in the main tables. The reported means match the headline FairReL entries in the main text (Tables 1–2 and the ablation tables).

	Celeb-DF		DFD		DFDC
Method	TPR $\uparrow$	BalAcc. $\uparrow$	TPR $\uparrow$	BalAcc. $\uparrow$	TPR $\uparrow$	BalAcc. $\uparrow$
Effort Yan et al. (2025)	71.4	64.7	80.6	71.5	60.5	58.4
FairReL (Ours)	74.9	69.6	83.2	76.7	62.5	62.8

Appendix

Table S5: Target-domain TPR (%) and balanced accuracy (%) under the same

\tau

pinned to source-domain

\mathrm{TPR}{=}0.90

: FairReL’s lower FPR does not come at the cost of cross-domain sensitivity.

			Per-subgroup AUC (%) $\uparrow$						Overall (%)
Data	Type	Method	M-W	M-A	M-B	F-W	F-A	F-B	AUC $\uparrow$	AUC ${}_{\mathrm{avg}}\uparrow$	gap $\downarrow$
FF++	ERM	Xception-ERM Chollet (2017)	89.73	88.73	87.68	89.15	88.26	87.18	88.50	88.46	2.55
	ERM	CLIP-ERM Radford et al. (2021)	93.64	92.70	91.71	93.09	92.25	91.24	92.40	92.44	2.40
	Gen.	ForAda Cui et al. (2025)	95.26	94.43	93.57	94.78	94.04	93.16	94.10	94.21	2.10
		Effort Yan et al. (2025)	95.90	95.13	94.32	95.45	94.76	93.94	94.83	94.92	1.96
		GenD Yermakov et al. (2026)	96.38	95.67	94.93	95.97	95.33	94.57	95.35	95.48	1.81

Appendix

Table S6: Per-subgroup AUC (%) on the source domain FF++ R”ossler et al. (2019) (intra-domain test split) and on the unseen targets Celeb-DF Li et al. (2020b) and DFD Dufour and Gully (2019) (trained on FF++). The Asian subgroups (M-A, F-A) are not annotated in Celeb-DF and DFD and are left blank. “gap” denotes the AUC gap (best

-

worst subgroup AUC). The best results are shown in bold and the second-best are underlined .

	Celeb-DF		DFD		DFDC
$\lambda_{\mathrm{orth}}$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$	AUC $\uparrow$	$F_{\mathrm{FPR}}\downarrow$
$0$	77.85	12.64	84.71	15.83	63.74	41.92
$0.01$	79.48	11.21	85.62	14.58	64.93	39.86
$\mathbf{0.1}$ ( ours )	80.64	10.12	86.13	13.55	66.34	38.36
$1.0$	80.21	10.54	85.79	13.92	65.78	38.94
$10$	77.42	12.03	84.33	15.31	63.12	41.27

Appendix

Table S7: Sensitivity to

\lambda_{\mathrm{orth}}

on cross-domain test performance with GCWD and SLMA both active. The optimal value

\lambda_{\mathrm{orth}}{=}0.1

is selected on the FF++ validation split (Sec. 4.1 of the main text) and matches the FairReL entry in Table 1 of the main text.

Sub-band	Demographic probe	Forgery probe
( $\ell$ )	balanced acc. (%) $\uparrow$	AUC (%) $\uparrow$
$\mathrm{LL}_{2}$	48.6	71.2
$\mathrm{LH}_{2}$	40.2	77.8
$\mathrm{HL}_{2}$	42.1	75.9
$\mathrm{HH}_{2}$	34.8	82.4
$\mathrm{LH}_{1}$	26.9	89.1

Appendix

Table S8: Per-sub-band linear probes on DFDC. Demographic information (probe balanced accuracy) is highest in the coarse low-frequency bands, while forgery-discriminative information (probe AUC) is highest in the fine high-frequency bands.

	Attribute	Spatial spectral (GCWD)		Pooled (SLMA)
Feature view	( $a$ )	low-freq	high-freq	mean shift
Principal-only ( $\mathbf{W}_{p}$ )	gender	0.12	0.08	0.34
Principal-only ( $\mathbf{W}_{p}$ )	ethnicity	0.21	0.09	0.22
Residual-only ( $\Delta\mathbf{W}$ )	gender	0.41	0.13	1.00
Residual-only ( $\Delta\mathbf{W}$ )	ethnicity	1.00	0.16	0.58

Appendix

Table S9: Subgroup-sensitive structure on a trained subspace-decomposed detector (DFDC), normalised per column. Visualised as Fig. 6(a) of the main text.

Feature / model	AUC $\uparrow$	Gender $\downarrow$ (chance 50.0)	Ethnicity $\downarrow$ (chance 33.3)
Principal-only $\mathbf{z}_{p}$ (reference)	–	80.1	66.0
Subspace-decomposed baseline Yan et al. (2025) , $\mathbf{z}_{\Delta}$	65.1	78.4	63.7
+ GCWD only	65.6	71.6	54.9
+ SLMA only	65.4	65.8	58.4
FairReL (GCWD + SLMA)	66.3	58.5	50.9

Appendix

Table S10: Leakage probe on DFDC: detection AUC (%) and linear-probe balanced accuracy (%) for recovering gender/ethnicity from the residual. Lower gender/ethnicity accuracy indicates less recoverable demographic information.

Deepfake Detection

Deepfake detection identifies manipulated or AI-generated media to support trustworthy communication, identity verification, and forensic analysis.

See all 233 papers

Toward Calibrated, Fair, and accurate Deepfake Detection
Jun 3, 2026Ryan Brown, Chris RussellAlgorithmic Fairness Deepfake Detection
University of Oxford
FairForensics: Seeing Expressions and Parsing Demographics via Vision-Language Modeling for Generalizable Fair Deepfake Detection
Aug 3, 2026Yaning Zhang, Jiao Wu, Zan Gao +1Algorithmic Fairness Deepfake Detection
Computer Vision Institute, School of Artificial Intelligence, Shenzhen University, Shenzhen, 518060, China · Shandong Artificial Intelligence Institute, Qilu University of Technology (Shandong Academy of Sciences), Jinan, 250014, China · Key Laboratory of Computer Vision and System, Ministry of Education, Tianjin University of Technology, Tianjin, 300384, China +2
Time to Fake It: Continual Machine Learning Intensifies Unfairness in Deepfake Detection
Oct 15, 2026Kajaree Das, Oliver Böse, Subarnaduti Paul +2Algorithmic Fairness Deepfake Detection
Faculty of Mathematics and Computer Science, University of Bremen, Bremen, Germany

Jun 3, 2026J/K move · Enter open · S save

Toward Calibrated, Fair, and accurate Deepfake Detection

Ryan Brown, Chris Russell

University of Oxford

Algorithmic Fairness Deepfake Detection Model Calibration

Open paper AskPDF arXiv

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.

Abstract

Deepfake Detection

Explore similar work

Toward Calibrated, Fair, and accurate Deepfake Detection

FairForensics: Seeing Expressions and Parsing Demographics via Vision-Language Modeling for Generalizable Fair Deepfake Detection

Time to Fake It: Continual Machine Learning Intensifies Unfairness in Deepfake Detection