Deepfake detectors safeguard against misinformation, rendering their failures as much a societal as a technical concern. As generators constantly evolve, detectors trained on static datasets miss newly arising threats, while progressive training interferes catastrophically. Yet, mitigating forgetfulness in terms of accuracy, the prevalent focus of continual learning (CL), is insufficient to reflect real-world impact. We therefore scrutinize the effects of CL methods with respect to evolving fairness. To this end, we analyze four dimensions of fairness across diverse demographics for detectors trained progressively on a chronological sequence of 11 real-world deepfake generation methods. Tracking fairness metrics over time, we expose non-trivial interplays between performance and fairness that depend on both the choice of CL algorithm and demographic intersections. We find accuracy degradation to typically exacerbate unfairness and, conversely, fairness improvements to remain scarce over time. Dissecting the CL method families (rehearsal, prompts, regularization), we observe that inequitable forgetting is not explained by underrepresentation alone, but likely entangled with representation learning. We thus issue a call to action for CL methods to take social outcomes into account when balancing the stability-plasticity tradeoff.
Figures & tables
Figure 1: Our chronological sequence of 11 Deepfake datasets, on which a single detector is trained continually. Each task shows its index, the generator, size ( k = thousands), date and an example image. Beads mark publication dates; bead area is proportional to dataset size and color denotes the generator family. Fairness is evaluated on demographic test annotations after every training step, but training only ever uses real/fake labels.
Figure 2: Demographic distribution across the 11 test sets. Colors denote disjoint attributes, each summing to 100%; green shows binary gender intersected with three skin-tone groups (of ten). Some groups are underrepresented throughout (e.g. very light skin tones), while others fluctuate heavily over time.
Figure 3Table 4
Figure 5: Subgroup ( g ) trajectories over tasks 1 → 11 (circle → arrow) against cumulative share of training data (log scale). Top (GradMatch, memory-based): memory share , the g ’s share of the replay buffer, i.e. how much memory is spent rehearsing it. Middle (SOYO, prompt-based): routing accuracy , the fraction of the g ’s test samples from all seen tasks routed to their own task’s prompt. Lower means more often matched with another task’s prompt. Bottom (EWC, regularization-based): Fisher cosine similarity between the diagonal Fisher on the g ’s samples and on each task’s full dataset at the final checkpoint. Lower means the g ’s important parameters are less protected. Scores are averaged over five seeds. The shared log x-axis allows comparing each subgroup across method families.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Demographic Attributes
Description
Gender
1 - Male, 0 - Female
Age
0 - Child, 1 - Youth, 3 - Adult,
4 - Middle-aged, 5-Senior
Skin Tone
According to Monk Skin Tone Scale
Intersection
0-(Female,Light), 1-(Female,Medium),
2-(Female,Dark), 3-(Male,Light),
Appendix
Table 1: Summary of demographic attributes provided by AI‑Face‑FairnessBench and used in our fairness analysis.
Datasets
Train
Test
Real
Fake
Total
Real
Fake
Total
FaceForensics (FF++)
17,218
84,346
101,564
4,375
21,233
25,608
DFD
37,231
55,099
92,330
9,461
13,823
23,284
CelebDF
18,883
124,647
143,530
4,762
30,842
35,604
DFDC
30,136
30,038
60,174
7,700
7,494
15,194
Appendix
Table 2: FaceForensics, DFD, CelebDF and DFDC are the four datasets which have real data already included and provided as part of AI-Face-FairnessBench. Here we include a detailed overview of the number of samples provided for each of these datasets. 10% of every respective training set is held out as validation set.
SL. No.
Dataset Publication Date
Dataset or Generator Used
Type of Generation
Samples
Real Dataset Used in Training
Train
Test
1
25.01.19
FaceForensics (FF++)
Hybrid
101,564
25,608
NA
2
24.09.19
DFD
Hybrid
92,330
23,284
NA
3
27.09.19
CelebDF
Hybrid
143,530
35,604
NA
4
11.12.19
DFDC
Hybrid
60,174
15,194
NA
5
17.12.20
VQGAN
GAN
39,943
10,057
CelebA, MetFaces
Appendix
Table 3: Task definitions and chronological ordering used across all our experiments. 10% of every respective training set is held out as validation set. This shows the corresponding date of publication of the dataset sources as provided in ( Lin et al., 2025 ) . Dates refer to the public release of the images, which for some generators (e.g., StyleGAN) may lag the generator’s first publication. The table also shows real data sources used for training and/or evaluation stage. A detailed overview of these real datasets are provided in table 4 .
Table 4: We fetched real data for generators if needed from external datasets or internal ones in AI-Face-FairnessBench. External data are not considered for testing as they do not have demographic annotations available. Here we include a detailed overview of the number of samples provided for each of these datasets. 10% of every respective training set is held out as validation set unless otherwise mentioned.
Train
Validation
Test
#
Dataset
Real
Fake
Real
Fake
Real
Fake
1
FaceForensics++
15,496
75,911
1,722
8,435
4,375
21,233
2
DFD
33,508
49,589
3,723
5,510
9,461
13,823
3
CelebDF
16,995
112,182
1,888
12,465
4,762
30,842
4
DFDC
27,122
27,034
3,014
3,004
7,700
7,494
5
VQGAN
35,948
35,949
3,995
3,994
26,298
10,057
Appendix
Table 5: Number of real and fake images per task and split. For VQGAN, Palette and DCFace, real training and validation images come from external sources (Table 4 ), and real test images are re-used from the test sets of earlier tasks.
Methods
Epochs
Learning Rate
Buffer/ Prompt Size
Specific Parameter
Accumulated
25
1×10−5
NA
NA
Naive
25
1×10−5
NA
NA
EWC
25
1×10−5
NA
λ=5000
DER++
25
1×10−5
Buffer Size = 3000
α=0.8,β=0.5
GradMatch
25
1×10−5
Buffer Size = 3000
β=0.5
SOYO
25
5×10−4
Prompt Length = 10
NA
Appendix
Table 6: Overview of method specific and shared hyperparameter settings used in all our experiments.
Figure 6: Inter-task accuracy matrix (mean ± std over five seeds) for the accumulated upper-bound and naive sequential lower-bound by training step (y-axis) and evaluated task (x-axis). The matrix highlights that each task is solvable in itself (diagonal values), but accuracy on generators not yet seen in training (top-right off-diagonal) remains limited and varies strongly across generators.
Figure 7: Inter-task accuracy matrix (mean ± std over five seeds) for EWC, GradMatch, PINA, SOYO and S-Prompts by training step (y-axis) and evaluated task (x-axis). The matrix highlights that each task is solvable in itself (diagonal values), but accuracy on generators not yet seen in training (top-right off-diagonal) remains limited and varies strongly across generators. CL methods mitigate forgetting for most generators (bottom-left off-diagonal).
Table 7: Comparison of (Un)fairness at the end of continually learning across the 11 generator sequence with accumulated and naive sequential baselines along with EWC and GradMatch for test sets of all previously observed tasks, i.e., test set 1 after training on 11 tasks, test set 2 after 11 tasks, and so on. To focus on whether the CL algorithm has overall improved or deteriorated fairness, the shown values (mean ± std over five seeds) are differences of the final fairness with respect to the fairness obtained right after training on the respective generator, e.g. discrepancy in fairness of task 2 after training on task 2 vs. task 2 after training on all 11 tasks. As before, red (blue) marks the many deteriorating (few improving) fairness dimensions. Cells remain white if the absolute difference is less than 2.5% to ensure the figure remains legible.
Table 8: Comparison of (Un)fairness at the end of continually learning across the 11 generator sequence with PINA, SOYO, and S-Prompts for test sets of all previously observed tasks, i.e. test set 1 after training on 11 tasks, test set 2 after 11 tasks, and so on. To focus on whether the CL algorithm has overall improved or deteriorated fairness, the shown values (mean ± std over five seeds) are differences of the final fairness with respect to the fairness obtained right after training on the respective generator, e.g. discrepancy in fairness of task 2 after training on task 2 vs. task 2 after training on all 11 tasks. As before, red (blue) marks the many deteriorating (few improving) fairness dimensions. Cells remain white if the absolute difference is less than 2.5% to ensure the figure remains legible.
Figure 8: Evolution of fairness of EWC across demographics (x-axes) for FDP (top left), FEO (top right), FEOP (bottom left) and FOAE (bottom right). Values are absolute fairness on each earlier dataset at the end of training (step 11), averaged over five seeds. As in earlier tables, higher means less fair. Color shows improvement (blue) or deterioration (red) relative to the naive sequential baseline, our reference to emphasize that fairness often deteriorates overall, and in parts even more than under naive fine-tuning.
Figure 9: Evolution of fairness of DER++ across demographics (x-axes) for FDP (top left), FEO (top right), FEOP (bottom left) and FOAE (bottom right). Values are absolute fairness on each earlier dataset at the end of training (step 11), averaged over five seeds. As in earlier tables, higher means less fair. Color shows improvement (blue) or deterioration (red) relative to the naive sequential baseline, our reference to emphasize that fairness often deteriorates overall, and in parts even more than under naive fine-tuning.
Figure 10: Evolution of fairness of GradMatch across demographics (x-axes) for FDP (top left), FEO (top right), FEOP (bottom left) and FOAE (bottom right). Values are absolute fairness on each earlier dataset at the end of training (step 11), averaged over five seeds. As in earlier tables, higher means less fair. Color shows improvement (blue) or deterioration (red) relative to the naive sequential baseline, our reference to emphasize that fairness often deteriorates overall, and in parts even more than under naive fine-tuning.
Figure 11: Evolution of fairness of SOYO across demographics (x-axes) for FDP (top left), FEO (top right), FEOP (bottom left) and FOAE (bottom right). Values are absolute fairness on each earlier dataset at the end of training (step 11), averaged over five seeds. As in earlier tables, higher means less fair. Color shows improvement (blue) or deterioration (red) relative to the naive sequential baseline, our reference to emphasize that fairness often deteriorates overall, and in parts even more than under naive fine-tuning.
Figure 12: Evolution of fairness of PINA across demographics (x-axes) for FDP (top left), FEO (top right), FEOP (bottom left) and FOAE (bottom right). Values are absolute fairness on each earlier dataset at the end of training (step 11), averaged over five seeds. As in earlier tables, higher means less fair. Color shows improvement (blue) or deterioration (red) relative to the naive sequential baseline, our reference to emphasize that fairness often deteriorates overall, and in parts even more than under naive fine-tuning.
Figure 13: Evolution of fairness of S-Prompts across demographics (x-axes) for FDP (top left), FEO (top right), FEOP (bottom left) and FOAE (bottom right). Values are absolute fairness on each earlier dataset at the end of training (step 11), averaged over five seeds. As in earlier tables, higher means less fair. Color shows improvement (blue) or deterioration (red) relative to the naive sequential baseline, our reference to emphasize that fairness often deteriorates overall, and in parts even more than under naive fine-tuning.
Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.
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
Facial forgery techniques are advancing rapidly, posing severe threats to public trust and information security while imposing higher demands on the continual adaptation of DeepFake detection models. Although continual learning enables models to adapt to emerging forgery methods, existing approaches still face two key bottlenecks. On the one hand, they lack sufficient feature representation capacity to handle increasingly diverse and complex forgery traces. On the other hand, continual adaptation to new forgery distributions leads to severe catastrophic forgetting of prior knowledge, which substantially degrades detection performance. To address these issues, we propose Face-D(^2)CL, a framework for facial DeepFake detection. It leverages multi-domain synergistic representation to fuse spatial and frequency-domain features, enabling comprehensive capture of diverse forgery traces. Additionally, it employs a dual continual learning mechanism that combines Real/Fake-aware Elastic Weight Consolidation (RF-EWC) and Domain-wise Orthogonal Gradient Constraint (D-OGC). RF-EWC distinguishes the parameter importance for real versus fake samples, while D-OGC ensures that updates to task-specific expert modules do not interfere with previously learned knowledge. This synergy allows the model to achieve a dynamic balance between robust anti-forgetting capabilities and agile adaptability to emerging facial forgery paradigms, all without relying on historical data replay. Extensive experiments demonstrate that our method surpasses current state-of-the-art (SOTA) approaches in both stability and plasticity, achieving a 60.7% relative reduction in the average detection error rate. On unseen forgery domains, it further improves the average detection AUC by 7.9% compared to the current SOTA method.
Yushuo Zhang, Yu Cheng, Yongkang Hu +3
East China Normal University, Shanghai, China · Shanghai Innovation Institute, Shanghai, China