Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

This paper claims that training an incremental deepfake detector on synthesized 'real' replays—without storing any original images—can beat sample-replay baselines, provided a confusion score adaptively switches between direct supervision a

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Adaptively weighting direct supervision versus a relative-separation loss makes generative replay work for incremental deepfake detection.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A genuine idea in a sloppy wrapper: the domain-safe/domain-risky split is worth a referee, but the adaptive alpha needs a detach statement and multi-seed evidence. the 4 major comments →

arxiv 2511.18436 v2 pith:L7U2B7BQ submitted 2025-11-23 cs.CV

When Generative Replay Meets Evolving Deepfakes: Dual Confusion-Aware Regularization for Incremental Face Forgery Detection

Hao Shen , Jikang Cheng , Renye Yan , Zhongyuan Wang , Wei Peng , Baojin Huang This is my paper
classification cs.CV
keywords incremental face forgery detectiongenerative replaycatastrophic forgettingdomain confusionrelative separation lossdiffusion-based forgery detectiondeepfake detectioncontinual learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that generative replay is a viable, privacy-preserving substitute for storing original data in incremental face forgery detection. It identifies a failure mode it calls the Domain Confusion Effect: when the replay generator resembles the current task's forgery generator, the synthesized real samples carry the same generative artifacts as the fakes, so a detector trained on them as 'real' learns a distorted boundary. The proposed strategy, DARW, separates replayed reals into domain-safe and domain-risky samples, supervises the safe ones directly, and applies a Relative Separation Loss to the risky ones while a Domain-aware Confusion Score adaptively balances the two. If the central claim is right, detectors can keep learning new forgery types indefinitely without retaining old images, with replay diversity as a free by-product of generation.

Core claim

The paper's central claim is that the bottleneck in generative replay for forgery detection is not the replayed fake samples but the replayed real samples, which are only sometimes safe to treat as real. Its experiments show that when the replay generator is close in distribution to the forgery generator, generated-real samples fall near fakes in feature space; when the generator is different, they remain distinguishable. DARW exploits this by measuring confusion with a DC Score—the L2 distance between the centroid of all generated-real replays and the centroid of the current task's fakes, normalized by tanh—and using that score to weight a direct cross-entropy loss against a Relative Separa

What carries the argument

The central mechanism is the pair of losses and the adaptive scalar between them. The DC Score is a normalized (tanh) L2 distance between the feature centroid of generated-real replay samples from all past tasks and the feature centroid of the current task's fake samples; a small distance signals high domain confusion. The Relative Separation Loss, applied to the generated-real samples when confusion is high, minimizes the average cosine similarity between each generated-fake feature and the generated-real centroid, so the detector still learns a relative real-vs-fake signal without treating the synthetic real as a hard label. The overall loss is a weighted combination: direct cross-entropy

Load-bearing premise

The load-bearing premise is that the L2 distance between the centroid of generated-real replays and the centroid of the current task's fakes, computed in the rolling feature space of the detector being trained and normalized by a tanh function, is a faithful meter of how much the replay reals will confuse the detector; if that proxy is miscalibrated, the adaptive weighting collapses toward fixed-weight baselines and the headline gain largely disappears.

What would settle it

Take a replay generator that produces high-FID images and is architecturally unrelated to the current fake generator, and measure the DC Score's predicted domain risk against an independent held-out test: train on the replay reals as 'real' and check how often they are misclassified as fake. If the DC Score calls a task domain-safe while that held-out real-classification accuracy collapses—or calls it risky while a fixed pretrained feature extractor shows large centroid separation—then the proxy is not tracking the confusion it is supposed to meter, and DARW's adaptive advantage should vanish.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If DARW's central claim holds, incremental deepfake detectors no longer need to store original forged images, removing the main privacy and memory bottleneck in continual forgery detection.
  • The same detector can keep absorbing new forgery types as they appear—classical face swaps and diffusion-based synthesis—without retraining from scratch on all past data.
  • Because a single generator can produce unlimited replay samples, replay diversity becomes essentially free, and the paper shows performance scales with replay size.
  • The adaptive weighting matters: fixed-weight versions of the loss are measurably worse on the longer protocol, so a confusion meter, not just a good loss, is part of the recipe.
  • The improvement is not tied to one generator; similar results appear across different diffusion-based replay generators, so the strategy can ride along with better generators as they appear.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The DC Score is computed in the feature space of the very backbone being optimized at that step, so its calibration can drift as the detector changes; if it mis-meters confusion, DARW degrades toward its own fixed-alpha baselines—the paper's numbers suggest that gap is a couple of AUC points, not a chasm.
  • The privacy story assumes that generated replay images do not leak enough identity information to count as personal data; that is plausible but not demonstrated and could be tested with face-recognition matching between generated reals and training identities.
  • The same two-supervision recipe might transfer to other incremental detection problems where synthetic 'real' data is self-referential—for example, detecting AI-generated audio or images, where the replay generator and the attack model share a pipeline.
  • A cheaper test of the mechanism: replace the centroid-L2 confusion score with a fixed, untrained feature extractor; if the gains evaporate, the score is learning to track exactly the artifacts it is meant to suppress.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper addresses incremental face forgery detection under evolving deepfake types by proposing a generative replay framework (DARW in the full text; 'Dual-CARE' in the arXiv abstract). Instead of storing real samples, it trains latent diffusion models to replay previous-task real and fake images. It identifies 'domain-risky' generated real samples that overlap with the current task's fake distribution and 'domain-safe' samples that do not. A Domain-aware Confusion Score (DC Score, Eq. 4–5) computed from feature-centroid distances between replayed real samples and current fakes adaptively weights a cross-entropy term against a Relative Separation Loss (Eq. 3). The method is evaluated on two incremental protocols (P1 and P2) against sample-replay and regularization baselines, reporting state-of-the-art average AUCs (e.g., P1 avg AUC 0.9574 vs. 0.9478 for SUR-LID). Ablations show drops when the Gen-Real supervision or RS loss is removed, and a fixed-α comparison in Tab. 4.

Significance. The paper's motivating question is timely and important: generative replay is a promising privacy-preserving alternative to sample replay, but its use in forgery detection is complicated by the fact that 'real' replay samples are synthetic. The observation that replay real samples can be 'domain-risky' when the replay generator resembles the current forgery generator is a useful conceptual contribution, and the experimental finding that fake-only replay suffers while full replay can hurt is informative. If the adaptive weighting mechanism is shown to be robust and not an artifact of the optimization, the work could provide a practical baseline for privacy-preserving incremental deepfake detection. The paper also ships the entire evaluation framework described in Sec. 5.1, which is a strength.

major comments (4)
  1. [§4.4–4.5, Eq. (4)–(7)] The DC Score α is computed with the very backbone f being optimized. The paper does not state whether gradients are detached when computing the centroids in Eq. (4) and the normalization in Eq. (5). If α is not detached, the model can reduce its value by shrinking the centroid distance, thereby down-weighting the hard cross-entropy supervision on generated-real samples and substituting the easier RS loss. The supplement's α trajectory (Supp. Tab. 1: α=0.2599 at T4 even though DDPM AUC is 0.9999) is consistent with this risk, and Tab. 4 shows fixed-α variants are only ~2 AUC points behind adaptive (0.9372/0.9312 vs. 0.9574). The central claim that adaptive weighting quantifies domain confusion is load-bearing; please state explicitly whether α is detached, and if not, re-run with stop-gradient and report both α trajectories and final AUCs.
  2. [§5.2, Tabs. 1–4 and §5.6] All tables report single-seed AUC/ACC with four decimals and no variance. The claimed margins over the strongest baseline are small (P1: 0.9574 vs. 0.9478 for SUR-LID; P2: 0.9429 vs. 0.9047), and the adaptive-vs-fixed margin is about 0.02. Without repeated runs and standard deviations, the headline claim that DARW 'consistently outperforms all baselines on both P1 and P2' is not statistically supported. Please provide mean ± std over at least 3–5 seeds, and ideally a paired significance test.
  3. [§5.3 and Supp. §3.2] The distance-metric pairing (L2 for DCS vs. cosine for RS) and the normalization function (tanh) are selected by comparing full test-set results across variants (e.g., Supp. Tab. 1: tanh vs. sigmoid vs. d/5). This is test-set selection bias; the reported configuration is the best on the evaluation data, so the reported numbers overstate expected performance. Please use a validation split for hyperparameter selection, or clearly disclose that the configuration was chosen on the test set.
  4. [Abstract vs. full text] The abstract in the arXiv record describes a method called 'Dual-CARE' that 'dual-modulates the optimization of both replay generators and the incremental detector.' The full text consistently describes 'DARW' and never updates the replay generator based on DC Score; Sec. 4.2 fixes the generator after training with Eq. (1). These are two different method descriptions. This inconsistency must be resolved before publication so readers know which method is actually proposed and evaluated.
minor comments (5)
  1. [Title/terminology] The method name differs between the abstract ('Dual-CARE') and the body ('DARW'). Pick one name and use it consistently.
  2. [Eq. (4)] The notation Cent(f(R_all_past_g)) is confusing because Eq. (2) defines Cent on raw inputs, not features. Clarify that the centroid is computed on the feature representations.
  3. [Fig. 4 (right)] 'LDM and DDIM, as the latest Diffusion-based detectors' should read 'generators'.
  4. [Sec. 5.3] Typo: 'reduces performance significanly' should be 'significantly'. Also the row 'w/o L_rs' in Tab. 3 has inconsistent spacing; check formatting.
  5. [References] Refs [35] and [36] are the same LDM paper but [36] is cited as SDv21; this appears to be an incorrect citation. Also, the paper does not discuss the computational overhead of training a separate LDM per task, which is relevant to the claimed scalability.

Circularity Check

0 steps flagged

No significant circularity found: the DC Score is a data-dependent weighting term, not a fitted quantity renamed as a prediction, and the reported gains are evaluated on held-out test data.

full rationale

The derivation chain is not circular. The DC Score alpha in Eq. 5 is computed from feature centroids of generated real replays and current-task fakes (Eq. 4); it is a data-dependent heuristic, not a fitted parameter renamed as a prediction. The detector loss Eq. 7 uses alpha only as a weighting coefficient, and the reported AUC is measured on held-out test data under the two protocols, not on the quantities used to set alpha. The Relative Separation Loss (Eq. 3) is an independent objective that separates generated real centroids from generated fake features; it does not reduce to the cross-entropy term or to alpha. The replay generators are trained with the standard diffusion objective (Eq. 1) and then used to produce fixed replay samples; no replay sample is constructed from test labels. Self-citations such as SUR-LID [8] provide baselines and protocol alignment, but the method's improvement is empirically demonstrated against those baselines within a standardized benchmark, so no load-bearing claim rests solely on an unverified self-citation. The choice of tanh normalization, distance metrics, and replay batch sizes is an empirical tuning choice on the same benchmark; this is an evaluation/overfitting concern, not a definitional circularity, and the paper does not present these choices as derived predictions. The potential issue that alpha depends on the very backbone being optimized and that the paper does not state whether gradients are detached is a training-dynamics/implementation concern, not a circularity, and there is no quoted step showing a claimed result reducing to its input by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The contribution is a heuristic loss-reweighting scheme, so the ledger is dominated by loss-form hyperparameters, all tuned on the same protocols used for the headline numbers. No new physical or architectural entities are postulated: 'DC Score' and 'RS Loss' are named loss components defined entirely by the paper's own equations, with no independent falsifiable handle. The assumptions that matter are the domain assumptions about what feature-space geometry means; none has external validation.

free parameters (5)
  • Replay batch sizes B_fake/B_real = 12/12 per 32-image batch
    Chosen by hand to maintain class balance (supp. §1.2); not swept, affects the balance between current-task and replay supervision.
  • α normalization function = tanh (sigmoid and d/5 variants compared)
    Selected on the same P1/P2 protocols where final numbers are reported (supp. §3.2); sigmoid costs ~2 AUC points on P1.
  • Distance/metric pairing (DC Score and RS Loss) = L2 for DC Score, cosine for RS Loss
    Selected via ablation on P1 (Tab. 3); the three other pairings score 0.9025–0.9291 vs 0.9498 for the chosen one.
  • Detector training budget = 5 epochs, Adam lr 2e-4, batch 32
    Implementation choice (§5.1); no sensitivity analysis reported.
  • Replay sampling configuration = DDIM, 250 steps, η=1.0
    Supp. §1.3; η=1.0 makes it strided DDPM, chosen to preserve diversity.
axioms (4)
  • domain assumption Feature-centroid L2 distance between generated-real replays and current-task fakes quantifies domain confusion (Eq. 4–5).
    Load-bearing for the adaptive mechanism: if this proxy is wrong, α mis-weights supervision. Only indirect evidence (ablations of normalization/metric variants) is provided.
  • domain assumption Generated fake replays are confusion-free and can be label-supervised regardless of replay generator (Sec. 3, §4.5).
    Justifies including X^f_g in L_cf; the motivation text argues replayed fakes do not affect detection, but Fig. 2 does not isolate this claim.
  • domain assumption Domain-risky generated real replays still encode useful previous-task information (Sec. 4.3).
    Motivates RS Loss; asserted ('similar learning effectiveness... suggests') without a shown supporting experiment.
  • domain assumption LDM trained with the L2 denoising objective (Eq. 1) adequately simulates each previous task's real and fake domains.
    The entire replay premise; quality is verified only by an asserted FID match in Fig. 2 and qualitative supplement samples.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of When Generative Replay Meets Evolving Deepfakes: Dual Confusion-Aware Regularization for Incremental Face Forgery Detection." pith.science (2026). https://pith.science/paper/L7U2B7BQ

@misc{pith2026251118436,
  author       = {Pith},
  title        = {Pith review of: When Generative Replay Meets Evolving Deepfakes: Dual Confusion-Aware Regularization for Incremental Face Forgery Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L7U2B7BQ}},
  note         = {Machine review of arXiv:2511.18436}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The rapid advancement of face generation techniques has introduced an increasing variety of forgery methods, making incremental deepfake detection essential for maintaining robust detectors. While generative replay provides a promising solution to alleviate catastrophic forgetting without storing historical data, its effectiveness is hindered by \textbf{domain confusion} between generated samples and real data. We observe that replay samples fall into two categories: when the replay generator closely resembles the newly introduced forgery model, generated real samples overlap with the fake domain and become \textbf{domain-risky}; when the generator differs significantly, generated samples maintain clearer domain separation and can be treated as \textbf{domain-safe}. To address this challenge, we propose a Dual \textbf{C}onfusion-\textbf{A}ware \textbf{RE}gularization strategy, termed \textbf{Dual-CARE}. A Domain-aware Confusion Score (DC Score) is introduced to quantify domain confusion and \textbf{dual-modulate} the optimization of both replay generators and the incremental detector. Guided by DC Score, replay generators are updated to better approximate previous-task distributions, while the detector adopts different supervision strategies: domain-safe samples are directly supervised, whereas domain-risky samples are regulated using a Relative Separation Loss to balance supervision and confusion. Extensive experiments demonstrate that Dual-CARE effectively exploits generative replay and improves incremental deepfake detection under evolving forgery scenarios.

Figures

Figures reproduced from arXiv: 2511.18436 by Baojin Huang, Hao Shen, Jikang Cheng, Renye Yan, Wei Peng, Zhongyuan Wang.

Figure 1
Figure 1. Figure 1: Left: Comparison between traditional sample replay and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Influence of distribution similarity on generative replay. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The overall framework of our method. 4.3. Relative Separation Loss As previously discussed, domain confusion among genera￾tive real replays, actual real samples, and actual fake sam￾ples may mislead the detector when direct label-based su￾pervision is conducted. However, the similar learning effec￾tiveness between domain-safe&real and domain-risky&real suggests that the domain-risky samples also contain pr… view at source ↗
Figure 4
Figure 4. Figure 4: Analysis of generative replay strategy. (Left) Perfor [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: UMAP visualization of T1 (LDM) features. (Left) Ini [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation on Relative Separation Loss: Sample-wise [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: Visualization of LDM-generated replay samples across [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: UMAP [30] visualization of the Domain-Safe scenario (DFDCP). 3 [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Pixel-Space Diffusion Transformers

    cs.CV 2026-07 conditional novelty 3.0

    A systematic review of pixel-space diffusion transformers, categorizing architectures and challenges for end-to-end image generation without latent compression.

Reference graph

Works this paper leans on

57 extracted references · 6 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Memory aware synapses: Learning what (not) to forget

    Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. InEuropean Con- ference on Computer Vision, pages 139–154, 2018. 2

    2018

  2. [2]

    Albumentations: fast and flexible image augmenta- tions.Information, 11(2):125, 2020

    Alexander Buslaev, Vladimir I Iglovikov, Eugene Khved- chenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A Kalinin. Albumentations: fast and flexible image augmenta- tions.Information, 11(2):125, 2020. 1

    2020

  3. [3]

    End-to-end reconstruction- classification learning for face forgery detection

    Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction- classification learning for face forgery detection. InIEEE Conference on Computer Vision and Pattern Recognition, pages 4113–4122, 2022. 1, 2

    2022

  4. [4]

    Self-supervised learning of adversarial exam- ple: Towards good generalizations for deepfake detection

    Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial exam- ple: Towards good generalizations for deepfake detection. InIEEE Conference on Computer Vision and Pattern Recog- nition, pages 18710–18719, 2022. 2

    2022

  5. [5]

    Local relation learning for face forgery detection

    Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Local relation learning for face forgery detection. InAAAI Conference on Artificial Intelligence, pages 1081–1088, 2021. 2

    2021

  6. [6]

    Diffusionface: Towards a comprehensive dataset for diffusion-based face forgery analysis.arXiv preprint arXiv:2403.18471, 2024

    Zhongxi Chen, Ke Sun, Ziyin Zhou, Xianming Lin, Xi- aoshuai Sun, Liujuan Cao, and Rongrong Ji. Diffusionface: Towards a comprehensive dataset for diffusion-based face forgery analysis.arXiv preprint arXiv:2403.18471, 2024. 5, 2

    Pith/arXiv arXiv 2024

  7. [7]

    Can we leave deepfake data behind in training deepfake detector?Advances in Neural Information Processing Systems, 37:21979–21998, 2024

    Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector?Advances in Neural Information Processing Systems, 37:21979–21998, 2024. 1

    2024

  8. [8]

    Stacking brick by brick: Aligned feature isolation for incremental face forgery detection

    Jikang Cheng, Zhiyuan Yan, Ying Zhang, Li Hao, Jiaxin Ai, Qin Zou, Chen Li, and Zhongyuan Wang. Stacking brick by brick: Aligned feature isolation for incremental face forgery detection. InIEEE Conference on Computer Vision and Pat- tern Recognition, pages 13927–13936, 2025. 1, 3, 5, 2

    2025

  9. [9]

    Ed ˆ4: Explicit data- level debiasing for deepfake detection.IEEE Transactions on Image Processing, 2025

    Jikang Cheng, Ying Zhang, Qin Zou, Zhiyuan Yan, Chao Liang, Zhongyuan Wang, and Chen Li. Ed ˆ4: Explicit data- level debiasing for deepfake detection.IEEE Transactions on Image Processing, 2025. 1

    2025

  10. [10]

    Xception: Deep learning with depthwise separable convolutions

    Franc ¸ois Chollet. Xception: Deep learning with depthwise separable convolutions. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 1251–1258, 2017. 2

    2017

  11. [11]

    Forensics adapter: Adapting clip for generalizable face forgery detection

    Xinjie Cui, Yuezun Li, Ao Luo, Jiaran Zhou, and Junyu Dong. Forensics adapter: Adapting clip for generalizable face forgery detection. InIEEE Conference on Computer Vision and Pattern Recognition Conference, pages 19207– 19217, 2025. 2

    2025

  12. [12]

    A continual learning survey: Defying for- getting in classification tasks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3366–3385, 2021

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ale ˇs Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying for- getting in classification tasks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3366–3385, 2021. 2

    2021

  13. [13]

    com / c / deepfake - detection - challengeAc- cessed 2021-04-24

    Deepfake detection challenge.https://www.kaggle. com / c / deepfake - detection - challengeAc- cessed 2021-04-24. 5, 2

    2021

  14. [14]

    Implicit identity leakage: The stum- bling block to improving deepfake detection generalization

    Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stum- bling block to improving deepfake detection generalization. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3994–4004, 2023. 2

    2023

  15. [15]

    Ddgr: Continual learning with deep diffusion-based generative replay

    Rui Gao and Weiwei Liu. Ddgr: Continual learning with deep diffusion-based generative replay. InInternational Conference on Machine Learning, pages 10744–10763,

  16. [16]

    Learning meta face recognition in un- seen domains

    Jianzhu Guo, Xiangyu Zhu, Chenxu Zhao, Dong Cao, Zhen Lei, and Stan Z Li. Learning meta face recognition in un- seen domains. InIEEE Conference on Computer Vision and Pattern Recognition, pages 6163–6172, 2020. 2

    2020

  17. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2

    2016

  18. [18]

    Denoising diffu- sion probabilistic models.Advances in Neural Information Processing Systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.Advances in Neural Information Processing Systems, 33:6840–6851, 2020. 2, 5, 1

    2020

  19. [19]

    Implicit identity driven deepfake face swapping detection

    Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. InIEEE Conference on Computer Vision and Pattern Recognition, pages 4490– 4499, 2023. 1, 2

    2023

  20. [20]

    Freqdebias: Towards generalizable deepfake detec- tion via consistency-driven frequency debiasing

    Hossein Kashiani, Niloufar Alipour Talemi, and Fatemeh Afghah. Freqdebias: Towards generalizable deepfake detec- tion via consistency-driven frequency debiasing. InIEEE Conference on Computer Vision and Pattern Recognition, pages 8775–8785. IEEE, 2025. 2

    2025

  21. [21]

    Sddgr: Stable diffusion-based deep generative replay for class incremental object detection

    Junsu Kim, Hoseong Cho, Jihyeon Kim, Yihalem Yimolal Tiruneh, and Seungryul Baek. Sddgr: Stable diffusion-based deep generative replay for class incremental object detection. InIEEE Conference on Computer Vision and Pattern Recog- nition, pages 28772–28781, 2024. 3

    2024

  22. [22]

    Cored: Gen- eralizing fake media detection with continual representation using distillation

    Minha Kim, Shahroz Tariq, and Simon S Woo. Cored: Gen- eralizing fake media detection with continual representation using distillation. InACM International Conference on Mul- timedia, pages 337–346, 2021. 3, 5

    2021

  23. [23]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Pith/arXiv arXiv

  24. [24]

    Overcoming catastrophic forgetting in neu- ral networks.Proceedings of the National Academy of Sci- ences, 114(13):3521–3526, 2017

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska- Barwinska, et al. Overcoming catastrophic forgetting in neu- ral networks.Proceedings of the National Academy of Sci- ences, 114(13):3521–3526, 2017. 2

    2017

  25. [25]

    Face x-ray for more gen- 9 eral face forgery detection

    Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more gen- 9 eral face forgery detection. InIEEE Conference on Computer Vision and Pattern Recognition, pages 5001–5010, 2020. 2

    2020

  26. [26]

    Celeb-df: A new dataset for deepfake forensics

    Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A new dataset for deepfake forensics. InIEEE Conference on Computer Vision and Pattern Recognition,

  27. [27]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 40(12):2935–2947, 2017. 2, 5

    2017

  28. [28]

    Exploring disentangled content information for face forgery detection

    Jiahao Liang, Huafeng Shi, and Weihong Deng. Exploring disentangled content information for face forgery detection. InEuropean Conference on Computer Vision, pages 128–

  29. [29]

    Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learn- ing

    Zheda Mai, Ruiwen Li, Hyunwoo Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learn- ing. InIEEE Conference on Computer Vision and Pattern Recognition, pages 3589–3599, 2021. 3

    2021

  30. [30]

    Umap: Uniform manifold approximation and projection for dimen- sion reduction.arXiv preprint arXiv:1802.03426, 2018

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimen- sion reduction.arXiv preprint arXiv:1802.03426, 2018. 7, 3

    Pith/arXiv arXiv 2018

  31. [31]

    Dfil: Deepfake incremental learning by exploiting domain-invariant forgery clues

    Kun Pan, Yifang Yin, Yao Wei, Feng Lin, Zhongjie Ba, Zhenguang Liu, Zhibo Wang, Lorenzo Cavallaro, and Kui Ren. Dfil: Deepfake incremental learning by exploiting domain-invariant forgery clues. InACM International Con- ference on Multimedia, pages 8035–8046, 2023. 1, 3, 5, 2

    2023

  32. [32]

    Scalable diffusion mod- els with transformers

    William Peebles and Saining Xie. Scalable diffusion mod- els with transformers. InIEEE International Conference on Computer Vision, pages 4195–4205, 2023. 5

    2023

  33. [33]

    Thinking in frequency: Face forgery detection by mining frequency-aware clues

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. InEuropean Conference on Computer Vision, pages 86–103. Springer, 2020. 2

    2020

  34. [34]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. InIEEE Conference on Com- puter Vision and Pattern Recognition, pages 2001–2010,

    2001

  35. [35]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InIEEE Conference on Computer Vision and Pattern Recognition, pages 10684– 10695, 2022. 2, 5, 1

    2022

  36. [36]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InIEEE Conference on Computer Vision and Pattern Recognition, pages 10684– 10695, 2022. 5, 3

    2022

  37. [37]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical image com- puting and computer-assisted intervention, pages 234–241. Springer, 2015. 1

    2015

  38. [38]

    Faceforen- sics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Chris- tian Riess, Justus Thies, and Matthias Nießner. Faceforen- sics++: Learning to detect manipulated facial images. In IEEE International Conference on Computer Vision, pages 1–11, 2019. 5, 2

    2019

  39. [39]

    Continual learning with deep generative replay.Advances in Neural Information Processing Systems, 30, 2017

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay.Advances in Neural Information Processing Systems, 30, 2017. 3

    2017

  40. [40]

    Detecting deep- fakes with self-blended images

    Kaede Shiohara and Toshihiko Yamasaki. Detecting deep- fakes with self-blended images. InIEEE Conference on Computer Vision and Pattern Recognition, pages 18720– 18729, 2022. 2

    2022

  41. [41]

    Always be dreaming: A new approach for data-free class-incremental learning

    James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. InIEEE International Conference on Computer Vision, pages 9374– 9384, 2021. 3

    2021

  42. [42]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 2, 1

    Pith/arXiv arXiv 2010

  43. [43]

    Dual contrastive learning for general face forgery detection

    Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. InAAAI Conference on Artificial Intelli- gence, pages 2316–2324, 2022. 1, 2

    2022

  44. [44]

    Continual face forgery detection via historical distribution preserving.International Journal of Computer Vision, 133(3):1067–1084, 2025

    Ke Sun, Shen Chen, Taiping Yao, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. Continual face forgery detection via historical distribution preserving.International Journal of Computer Vision, 133(3):1067–1084, 2025. 1, 3, 5

    2025

  45. [45]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational Conference on Machine Learning, pages 6105–6114, 2019. 5

    2019

  46. [46]

    Dynamic mixed-prototype model for incremental deepfake detection

    Jiahe Tian, Cai Yu, Xi Wang, Peng Chen, Zihao Xiao, Jizhong Han, and Yesheng Chai. Dynamic mixed-prototype model for incremental deepfake detection. InACM Interna- tional Conference on Multimedia, pages 8129–8138, 2024. 1, 3

    2024

  47. [47]

    Representative forgery mining for fake face detection

    Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. InIEEE Conference on Com- puter Vision and Pattern Recognition, pages 14923–14932,

  48. [48]

    Der: Dynam- ically expandable representation for class incremental learn- ing

    Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynam- ically expandable representation for class incremental learn- ing. InIEEE Conference on Computer Vision and Pattern Recognition, pages 3014–3023, 2021. 5

    2021

  49. [49]

    Ucf: Uncovering common features for generalizable deep- fake detection

    Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deep- fake detection. InIEEE International Conference on Com- puter Vision, pages 22412–22423, 2023. 2

    2023

  50. [50]

    Deepfakebench: A comprehensive benchmark of deepfake detection.arXiv preprint arXiv:2307.01426,

    Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. Deepfakebench: A comprehensive benchmark of deepfake detection.arXiv preprint arXiv:2307.01426,

    Pith/arXiv arXiv

  51. [51]

    Transcending forgery specificity with latent space augmentation for generalizable deepfake detection

    Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 8984–8994, 2024. 1, 2

    2024

  52. [52]

    Df40: Toward next-generation deepfake detection.arXiv preprint arXiv:2406.13495, 2024

    Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Li Yuan, Chengjie 10 Wang, Shouhong Ding, et al. Df40: Toward next-generation deepfake detection.arXiv preprint arXiv:2406.13495, 2024. 5

    Pith/arXiv arXiv 2024

  53. [53]

    Effort: Efficient orthogonal mod- eling for generalizable ai-generated image detection

    Zhiyuan Yan, Jiangming Wang, Zhendong Wang, Peng Jin, Ke-Yue Zhang, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, and Li Yuan. Effort: Efficient orthogonal mod- eling for generalizable ai-generated image detection. InIn- ternational Conference on Machine Learning, 2025. 2

    2025

  54. [54]

    Multi-attentional deep- fake detection

    Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep- fake detection. InIEEE Conference on Computer Vision and Pattern Recognition, pages 2185–2194, 2021. 2 11 When Generative Replay Meets Evolving Deepfakes: Domain-Aware Relative Weighting for Incremental Face Forgery Detection Supplementary Material

    2021

  55. [55]

    Detailed Implementation Settings 1.1. Preprocessing Data Preprocessing.Following the standard Deep- FakeBench [50] protocol, all video frames undergo face de- tection, extraction, and alignment before being resized to 256×256. For input normalization, we adopt a mean of [0.5,0.5,0.5]and a standard deviation of[0.5,0.5,0.5]for the three RGB channels. Durin...

  56. [56]

    1, we present qualitative examples of re- play samples generated by our LDM [35] generator

    Sample Visualizations of Generative Replay As shown in Fig. 1, we present qualitative examples of re- play samples generated by our LDM [35] generator. The 1 generated images are visualized across three representa- tive datasets: DiffusionFace [6], FaceForensics++ [38], and DFDCP [13]. For each dataset, we show samples from both the Real and corresponding...

  57. [57]

    domain-safe

    Further Analysis and Ablations 3.1. Impact of Sample-wise Constraint inL rs To assess the importance of the fine-grained sample-wise constraint in our Relative Separation Loss (L rs), we com- pare it with a coarse-grained centroid-based variant. As shown in Fig. 2, both approaches perform similarly on the shorter Protocol 2. However, under the longer Prot...

    1947

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.