Pith. sign in

REVIEW 4 major objections 6 minor 70 references

The paper argues that deepfake detection cannot be generalized through static training: across 618 chronological configurations and four lightweight models, every continual-learning strategy scored near chance (FWT-AUC ≈ 0.5) on future gene

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 ↗

A chronological continual learning study finds deepfake detectors retain past knowledge but generalize to future generators at near-random AUC around 0.5.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A systematic and useful CL benchmark for deepfake detection, but the headline hypothesis about non-transferable generator imprints is overclaimed because the evaluation confounds generator identity with content domain and data volume. the 4 major comments →

arxiv 2509.07993 v1 pith:YJRJQP7O submitted 2025-08-29 cs.LG cs.AIcs.CVcs.GR

Revisiting Deepfake Detection: Chronological Continual Learning and the Limits of Generalization

Federico Fontana , Anxhelo Diko , Romeo Lanzino , Marco Raoul Marini , Bachir Kaddar , Gian Luca Foresti , Luigi Cinque This is my paper
classification cs.LG cs.AIcs.CVcs.GR
keywords deepfake detectioncontinual learningdomain-incremental learningforward transfercatastrophic forgettingnon-universal deepfake distribution hypothesisAUC evaluationchronological simulation
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

This paper tries to establish that a deepfake detector trained on past generators cannot, without additional training, recognize fakes made by future generators. To test this, it reframes deepfake detection as a continual-learning problem ordered by real release dates, streaming six datasets from 2018 to 2024 through eight continual-learning strategies and four lightweight models in over 600 simulations. The headline result is that retention of historical knowledge works, with the best configuration reaching a mean C-AUC around 0.915, but forward transfer does not: FWT-AUC stays between 0.49 and 0.57, essentially random. From this the paper proposes the Non-Universal Deepfake Distribution Hypothesis: each generator leaves a unique, non-transferable signature, so no static training set can cover the future. The practical upside is efficiency, with continual updating using about 155 times less GPU time than full retraining, which matters for real-time social-media monitoring.

Core claim

The paper's central claim is the Non-Universal Deepfake Distribution Hypothesis: there is no learnable 'distribution of deepfakes' shared across generators; each generator imprints artifacts specific to itself, so a detector trained on past generators performs no better than random on the next unseen one, with FWT-AUC between 0.49 and 0.57 across all 618 configurations. The supporting measurements are the best excess-over-chance transfer, Tmax = 0.094, the per-step decay, Tdecay = 0.54, and the resulting compounded degradation, Tcomp = 0.515 in three time steps, together with the empirical decorrelation between C-AUC and FWT-AUC. The paper also claims this limitation is not a defect of any p

What carries the argument

The load-bearing machinery is a time-ordered domain-incremental simulation: six datasets spanning July 2018 to February 2024 are streamed month by month, selected by an exponentially weighted reverse-chronological protocol, so the classifier sees realistic mixtures of old and new data. Performance is measured with two AUC-based metrics, C-AUC for retention on seen datasets and FWT-AUC for transfer to unseen future datasets, and the failure pattern is condensed into Tmax, the best excess AUC over 0.5 on the next generator, and Tdecay, the per-step erosion of that excess, yielding a compounded degradation model.

Load-bearing premise

The central claim assumes that the six time-ordered datasets are all instances of the same deepfake-detection task, so that near-chance scores on the later diffusion-based, non-face datasets reflect failure to transfer between generators rather than simply measuring a different problem.

What would settle it

Train a detector on face-manipulation deepfake datasets at a matched data volume and test it separately on diffusion-generated faces and on non-face diffusion images; if it scores significantly above chance on the faces while failing only on non-face images, the observed FWT-AUC floor is a content or domain artifact, not a property of generators. Alternatively, rerun the continual protocol with the full retraining budget of 5.77M unique samples; if FWT-AUC then exceeds 0.57, the reported ceiling is a data-efficiency artifact.

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

If this is right

  • Detectors deployed in the wild must be updated continuously; a detector trained once on past deepfakes behaves near randomly on next-generation fakes within roughly three generator generations.
  • Evaluation of continual deepfake detection should report both C-AUC and FWT-AUC, since historical retention and future transfer are empirically decorrelated objectives.
  • Efficient architectures plus replay and regularization make update pipelines about 155 times cheaper in GPU time than full retraining, enabling frequent updates at scale.
  • The chronological protocol can replace arbitrary or affinity-based generator sequences when benchmarking deepfake detection, because it reflects real release order.
  • Static training baselines should no longer be treated as a sufficient deployment strategy for deepfake detectors under evolving generators.

Where Pith is reading between the lines

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

  • Inference: the hypothesis suggests a more general statement beyond this testbed: any train-once-deploy-forever detector will fail as generator technology shifts, so the useful research target becomes fast adaptation and test-time updating rather than universal features.
  • Inference: because the later datasets are diffusion-generated and non-face while the early ones are GAN face manipulations, the near-random FWT-AUC may partly reflect task-domain shift; a cleaner test of the hypothesis would hold content and architecture constant while varying only the generator.
  • Inference: the efficiency numbers imply that a decentralized, frequently updated detection ecosystem is feasible, since runs could be executed on a single consumer GPU and small teams could maintain detectors without centralized infrastructure.
  • Inference: a testable extension would train on all six generators with the full retraining budget of 5.77M unique samples and then measure FWT-AUC on a seventh held-out generator; if it rises well above 0.57, the reported ceiling is partly a small-data artifact rather than fundamental non-transferability.
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 / 6 minor

Summary. The paper reframes deepfake detection as a domain-incremental continual learning problem with a chronological data stream spanning 2018–2025, using six public datasets and four lightweight backbones. It evaluates eight continual-learning strategies under three monthly-batch schedules and proposes two new metrics: Continual AUC (C-AUC) for historical retention and Forward Transfer AUC (FWT-AUC) for future generalization. Across 618 configurations, the authors find that C-AUC can be high with efficient training (≈155× less GPU time than full retraining), but FWT-AUC remains near 0.5 for all methods and models. They interpret this as evidence for the Non-Universal Deepfake Distribution Hypothesis—that each deepfake generator leaves a unique, non-transferable signature—and fit an exponential decay model (Tmax=0.094, Tdecay=0.54) to argue that static detectors degrade to chance within three time steps.

Significance. If the central claim is confirmed, the paper provides an important negative result for deepfake detection: no static detector can generalize to future generators, so continual updating is essential. The benchmark itself is a useful contribution: a chronological, realistic simulation protocol with six datasets, lightweight architectures, and a large grid of CL strategies. The C-AUC and FWT-AUC metrics are sensible for imbalanced continual-learning evaluation and may be reused by the community. The efficiency analysis (orders-of-magnitude reduction in GPU time versus full retraining) is practically relevant. However, the paper's interpretive claim—that near-chance forward transfer proves generator-specific 'non-transferable imprints'—is not yet supported because the evaluation confounds generator identity with content-domain shift and data volume. The empirical regularity is solid, but the causal inference is not.

major comments (4)
  1. [Sec. 4.5 / Table 1] The FWT-AUC ≈ 0.5 result is confounded by content-domain shift. The four training datasets (DeepfakeTIMIT, WildDeepfake, DFFD, FakeAVCeleb) are face-manipulation media, while the held-out 'future' datasets COCOFake and CIFAKE are Stable Diffusion images of general COCO scenes and synthetic CIFAR-10 images. A face-trained detector failing on non-face diffusion images is expected from task shift alone and does not support the conclusion that each generator leaves a unique, non-transferable signature. To support the Non-Universal Deepfake Distribution Hypothesis, the authors should include a control condition in which the content domain is held constant (e.g., face-only diffusion generators or multiple face-manipulation generators from different years) so that only the generator identity changes. Without such a control, the central claim of Sec. 4.6 is not established.
  2. [Sec. 4.5 / Table 3] The forward-transfer evaluation is performed on models that see at most 64K unique samples (monthly batches = 50) compared with 5.77M for full retraining. The near-random FWT-AUC could reflect undertraining or insufficient data rather than inherent non-transferability. The paper does not include a matched-data baseline (e.g., full retraining on the same 64K samples, or CL methods trained with larger buffers or more epochs). This confound affects the interpretation of every FWT-AUC value in Table 2 and the Tmax/Tdecay estimates in Eqs. (10)–(11). At minimum, report a learning-curve analysis showing that FWT-AUC is flat as data volume increases.
  3. [Sec. 4.6.1, Eqs. (7)–(12)] The exponential decay model is fitted to the data and then used as a derivation. Tmax = 0.094 and Tdecay = 0.54 are computed from the experimental FWT-AUC values (Eqs. 10–11), and then Eq. (12) 'derives' Tcomp = 0.515. This is a restatement of the fitted parameters, not an independent prediction. The claim that 'even the best strategy turns to random guessing in 3 time units' follows by construction from the assumed exponential form. Additionally, Tdecay is computed only over experiments with eval AUC ≥ 0.75; no justification is given for this selection threshold, and the ratio in Eq. (8) becomes unstable when the denominator is near zero. The authors should present this as a phenomenological fit, include confidence intervals, and compare against alternative dynamics (e.g., linear or power-law decay) before drawing the conclusion.
  4. [Sec. 4.6, Hypothesis Statement] The hypothesis 'Deepfake detection cannot be generalized through static training' is too broad for the evidence presented. The evaluation covers only six datasets released between 2018 and 2024, with two of the six being non-face diffusion images. The results support a more limited claim: in this chronological setup and under these training budgets, none of the tested CL strategies achieve forward transfer above chance. A universal negative claim requires evidence across a substantially broader range of generators and content types, or a formal bound. As written, the central claim overstates the empirical support and invites misinterpretation.
minor comments (6)
  1. [Sec. 3.2.1, Eq. (3)] Equation (3) is difficult to parse; the notation involving '0.5i' and the normalization denominator is garbled. Also, the base 0.5 is arbitrary and no sensitivity analysis is provided for this choice. Rewrite as a truncated geometric distribution and report whether the main results change for other base values.
  2. [Sec. 4.5 / Table 3] The claim of '155 times less GPU time' is not directly derivable from Table 3. Clarify the computation: compare which full-retraining scenario (N=1? N=2?) to which continual-learning configuration? Also, the repeated columns in Table 3 for the three monthly-batch settings make the table hard to read; consider splitting or labeling more clearly.
  3. [Sec. 4.4] The sentence 'each method processes a total of 80 × monthly batches × 16 samples' is unclear. Where does the factor 80 come from? If it is the number of simulated months, state so explicitly and define the timeline length.
  4. [Table 1 / Figure 1] There are typos in the dataset names: 'FakeA VCelebv2' in Table 1 and 'CIF AKECOCOFake' in Figure 1. Additionally, references [1] and [2] are the same paper, and several reference entries are incomplete (e.g., 'et al. Ju' and 'et al. Yan'). A thorough proofread is needed.
  5. [Sec. 4.6.1] The notation uses the same symbols for theoretical quantities and empirical estimates (Tmax, Tdecay, etc.). This makes the 'Empirical Parameterization' paragraph confusing. Use hats or distinct subscripts for the estimates.
  6. [General] The manuscript does not include a limitations section. Given the confounds identified in the major comments, a discussion of the scope and possible alternative explanations for the observed FWT-AUC plateau is essential.

Circularity Check

1 steps flagged

Tcomp=0.515 is a restatement of fitted Tmax/Tdecay under an assumed exponential decay; the raw FWT-AUC measurement is non-circular but the formalized decay 'prediction' reduces by construction.

specific steps
  1. fitted input called prediction [Section 4.6.1, Eqs. (10)-(12)]
    "The empirical value Tdecay = 0.54, calculated across all CL strategies and hyperparameters (selecting only experiments where the classifier achieved at least 0.75 on evaluation AUC), indicates that residual classification capacity decays exponentially. This leads to a compounding degradation process: Tcomp ≈ Tcomp = 0.5 + (Tmax) · (Tdecay)k = = 0.515 ≈ Random guessing"

    Tmax (Eq. 10) and Tdecay (Eq. 11) are descriptive statistics computed from the very FWT-AUC curves that the hypothesis is meant to explain. The exponential/geometric form in Eq. 9 is assumed, not derived. Substituting Tmax=0.094 and Tdecay=0.54 into Eq. 9 with k=3 yields Tcomp=0.5+0.094*0.54^3=0.515, which is exactly a re-expression of the observed near-0.5 FWT-AUC. The 'derived' conclusion that static training degrades to random guessing in 3 time steps is therefore contained in the fitted parameters rather than an independent prediction.

full rationale

The paper's central empirical claim—that FWT-AUC remains near 0.5 across 618 configurations—is a direct measurement and is not circular by itself. The circularity is localized to Sec. 4.6.1, where Tmax and Tdecay are defined as summary statistics over exactly those FWT-AUC observations, an exponential decay model is assumed in Eq. 9, and the fitted values are then substituted into Eq. 12 to 'derive' Tcomp≈0.515 and the conclusion that the best strategy turns to random guessing in 3 time steps. This is a fitted parameter renamed as a prediction: the 3-step decay is a re-expression of the already-observed near-0.5 AUC under an assumed parametric form, not an independent consequence of the hypothesis. I did not find load-bearing self-citation or an imported uniqueness theorem; the references to the authors' prior work [27] are not used to justify the central claim. The data-volume and domain-shift confounds (COCOFake/CIFAKE are diffusion-based non-face images; CL models see at most 64K unique samples vs 5.77M for full retraining) are serious validity threats to the 'unique generator signature' interpretation, but they are not circularity and are not scored as such. Overall: partial circularity in the formalized decay model, with an independent (though confounded) empirical core, so score 6.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on the timeline protocol, the dataset domain assumption, and the fitted decay model. The paper introduces no physical entities, but it postulates a 'unique generator imprint' as the explanation for observed FWT-AUC, and it hand-choses the exponential sampling weights and the decay model.

free parameters (5)
  • Tmax (maximum temporal transferability) = 0.094
    Computed as max over strategies/times of (AUC(C_s, D_{t+1}) - 0.5) in Eq 10; used as the scale factor in the Tcomp exponential model.
  • Tdecay (transfer decay factor) = 0.54
    Average ratio of excess AUC at t+2 to t+1 over strategies/times (Eq 11), selecting only experiments with eval AUC >= 0.75; used as the exponential base in Tcomp.
  • Exponential sampling weight base 0.5 = 0.5
    Hand-chosen base in Eq 3 that controls how quickly older datasets are downweighted in the timeline protocol; no evidence it matches real social media data streams.
  • Monthly batch count (10/20/50)
    Protocol hyperparameter that controls data volume; FWT-AUC is robust across these, but C-AUC varies, so this is a tunable element of the framework.
  • CL method hyperparameters (LR, lambda, buffer size) = per-method values in Sec 4.4
    Tuned per method and per batch setting; the hypothesis claims robustness across these configurations, but the tuning itself is part of the experimental setup.
axioms (5)
  • domain assumption The six datasets and their release dates represent the real-world chronological evolution of deepfake generators.
    Sec 3.2 and Table 1 assign dates (some N/A) and treat DFFD/FakeAVCeleb as era markers; the entire chronological simulation rests on this.
  • ad hoc to paper The exponentially weighted random selection protocol (Eq 3) simulates real-world data streams at time t.
    Sec 3.2.1; no validation against actual social media deepfake statistics; the 0.5 weighting is arbitrary.
  • standard math AUC on held-out test sets of future datasets is an appropriate measure of temporal generalization.
    Sec 3.3.2; AUC is a standard imbalance-robust metric; the assumption is that aggregate average over future datasets captures generalization capability.
  • ad hoc to paper The exponential decay model Tcomp = 0.5 + Tmax * (Tdecay)^k describes the temporal evolution of detection capacity.
    Sec 4.6.1 Eq 9; this functional form is assumed, not derived, and fitted to the observed Tmax/Tdecay.
  • domain assumption COCOFake and CIFAKE count as 'deepfakes' comparable to face-manipulation generators.
    Table 1; these are Stable Diffusion synthetic images, not face-swap deepfakes; the paper treats them as later deepfake generations without justifying comparability.
invented entities (1)
  • Unique non-transferable generator imprint no independent evidence
    purpose: Explains why FWT-AUC is near 0.5; each generator leaves a signature that does not transfer to future generators.
    No direct measurement of an imprint; it is inferred from low cross-generator AUC, which could also be explained by domain shift or data scarcity.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Deepfake Detection: Chronological Continual Learning and the Limits of Generalization." pith.science (2026). https://pith.science/paper/YJRJQP7O

@misc{pith2026250907993,
  author       = {Pith},
  title        = {Pith review of: Revisiting Deepfake Detection: Chronological Continual Learning and the Limits of Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJRJQP7O}},
  note         = {Machine review of arXiv:2509.07993}
}
Share X Bluesky LinkedIn Reddit HN
abstract

The rapid evolution of deepfake generation technologies poses critical challenges for detection systems, as non-continual learning methods demand frequent and expensive retraining. We reframe deepfake detection (DFD) as a Continual Learning (CL) problem, proposing an efficient framework that incrementally adapts to emerging visual manipulation techniques while retaining knowledge of past generators. Our framework, unlike prior approaches that rely on unreal simulation sequences, simulates the real-world chronological evolution of deepfake technologies in extended periods across 7 years. Simultaneously, our framework builds upon lightweight visual backbones to allow for the real-time performance of DFD systems. Additionally, we contribute two novel metrics: Continual AUC (C-AUC) for historical performance and Forward Transfer AUC (FWT-AUC) for future generalization. Through extensive experimentation (over 600 simulations), we empirically demonstrate that while efficient adaptation (+155 times faster than full retraining) and robust retention of historical knowledge is possible, the generalization of current approaches to future generators without additional training remains near-random (FWT-AUC $\approx$ 0.5) due to the unique imprint characterizing each existing generator. Such observations are the foundation of our newly proposed Non-Universal Deepfake Distribution Hypothesis. \textbf{Code will be released upon acceptance.}

Figures

Figures reproduced from arXiv: 2509.07993 by Anxhelo Diko, Bachir Kaddar, Federico Fontana, Gian Luca Foresti, Luigi Cinque, Marco Raoul Marini, Romeo Lanzino.

Figure 1
Figure 1. Figure 1: Illustration of the simulation framework spanning from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The lines represent the 8 CL methods evaluated on MobileNetV4, each configured with its optimal hyperparameters found in the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Spider charts showing the evaluation AUC on each dataset for each evaluation step using different strategies. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

70 extracted references · 57 canonical work pages · 5 internal anchors

  1. [2]

    Parents and children: Distinguishing multimodal deepfakes from natural images

    Roberto Amoroso, Davide Morelli, Marcella Cornia, Lorenzo Baraldi, Alberto Del Bimbo, and Rita Cucchiara. Parents and children: Distinguishing multimodal deepfakes from natural images. Arxiv Preprint Arxiv:2304.00500 ,

    Pith/arXiv arXiv

  2. [3]

    Learn- ing fast, learning slow: A general continual learning method based on complementary learning system

    Elahe Arani, Fahad Sarfraz, and Bahram Zonooz. Learn- ing fast, learning slow: A general continual learning method based on complementary learning system. Arxiv Preprint Arxiv:2201.12604, 2022. 2, 5, 6

    Pith/arXiv arXiv 2022

  3. [4]

    Cifake: Image classifica- tion and explainable identification of ai-generated synthetic images

    Jordan J Bird and Ahmad Lotfi. Cifake: Image classifica- tion and explainable identification of ai-generated synthetic images. IEEE Access, 12:15642–15650, 2024. 5

    work page 2024

  4. [5]

    Dark experience for general continual learning: a strong, simple baseline

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances In Neural Information Processing Systems , 33:15920–15930,

    work page

  5. [6]

    Harnessing the power of text-image contrastive models for automatic detection of on- line misinformation

    Hao Chen, Peng Zheng, Xin Wang, Shu Hu, Bin Zhu, Jin- rong Hu, Xi Wu, and Siwei Lyu. Harnessing the power of text-image contrastive models for automatic detection of on- line misinformation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 923–932, 2023. 3

    work page 2023

  6. [7]

    Masked Conditional Diffusion Model for Enhancing Deepfake Detection

    Tiewen Chen, Shanmin Yang, Shu Hu, Zhenghan Fang, Ying Fu, Xi Wu, and Xin Wang. Masked conditional diffusion model for enhancing deepfake detection. Arxiv Preprint Arxiv:2402.00541, 2024. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [8]

    Fakecatcher: Detection of synthetic portrait videos using biological sig- nals

    Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological sig- nals. Ieee Transactions On Pattern Analysis And Machine Intelligence, 2020. 2

    work page 2020

  8. [9]

    Forensictrans- fer: Weakly-supervised domain adaptation for forgery detec- tion

    Davide Cozzolino, Justus Thies, Andreas R ¨ossler, Christian Riess, Matthias Nießner, and Luisa Verdoliva. Forensictrans- fer: Weakly-supervised domain adaptation for forgery detec- tion. Conference On Computer Vision And Pattern Recogni- tion, 2018. 2

    work page 2018

  9. [10]

    On the detection of digital face manipulation

    Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 5781–5790, 2020. 5

    work page 2020

  10. [11]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 5, 6

    Pith/arXiv arXiv 2010

  11. [12]

    Attacking identity semantics in deepfakes via deep feature fusion

    Bing Fan, Zihan Jiang, Shu Hu, and Feng Ding. Attacking identity semantics in deepfakes via deep feature fusion. In 2023 IEEE 6th International Conference on Multimedia In- formation Processing and Retrieval (MIPR), pages 114–119. IEEE, 2023. 3

    work page 2023

  12. [13]

    Synthesizing black-box anti-forensics deepfakes with high visual quality

    Bing Fan, Shu Hu, and Feng Ding. Synthesizing black-box anti-forensics deepfakes with high visual quality. Icassp, 2024

    work page 2024

  13. [14]

    Detection, Attribution and Localization of GAN Generated Images

    Michael Goebel, Lakshmanan Nataraj, Tejaswi Nanjun- daswamy, Tajuddin Manhar Mohammed, Shivkumar Chan- drasekaran, and BS Manjunath. Detection, attribution and localization of gan generated images. Arxiv Preprint Arxiv:2007.10466, 2020. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2007

  14. [15]

    An empirical investigation of catas- trophic forgetting in gradient-based neural networks

    Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catas- trophic forgetting in gradient-based neural networks. Arxiv Preprint Arxiv:1312.6211, 2013. 2

    Pith/arXiv arXiv 2013

  15. [16]

    Studies of mind and brain: Neural principles of learning, perception, development, cognition, and motor control

    Stephen T Grossberg. Studies of mind and brain: Neural principles of learning, perception, development, cognition, and motor control . Springer Science & Business Media,

    work page

  16. [17]

    Deepfake detection by analyzing convolutional traces

    Luca Guarnera, Oliver Giudice, and Sebastiano Battiato. Deepfake detection by analyzing convolutional traces. In Conference on Computer Vision and Pattern Recognition Workshops, pages 666–667, 2020. 2

    work page 2020

  17. [18]

    Robust attentive deep neural network for detecting gan- generated faces

    Hui Guo, Shu Hu, Xin Wang, Ming-Ching Chang, and Siwei Lyu. Robust attentive deep neural network for detecting gan- generated faces. Ieee Access, 10:32574–32583, 2022. 3

    work page 2022

  18. [19]

    Deep fake image detection based on pairwise learning

    Chih-Chung Hsu, Yi-Xiu Zhuang, and Chia-Yen Lee. Deep fake image detection based on pairwise learning. Applied Science., 10(1):370, 2020. 2

    work page 2020

  19. [20]

    Improving the generalization ability of deepfake detection via disentangled representation learning

    Jiashang Hu, Shilin Wang, and Xiaoyong Li. Improving the generalization ability of deepfake detection via disentangled representation learning. In 2021 IEEE International Confer- ence on Image Processing (ICIP) , pages 3577–3581. IEEE,

    work page 2021

  20. [21]

    De- tecting cnn-generated facial images in real-world scenarios

    Nils Hulzebosch, Sarah Ibrahimi, and Marcel Worring. De- tecting cnn-generated facial images in real-world scenarios. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops , pages 642–643,

    work page

  21. [22]

    et al. Ju. Improving fairness in deepfake detection. In Pro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV), 2024. Open Access ver- sion available on the CVF website. 3

    work page 2024

  22. [23]

    Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S. Woo. FakeA VCeleb: A novel audio-video multimodal deep- fake dataset. In Thirty-fifth Conference on Neural Infor- mation Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5

    work page 2021

  23. [24]

    Overcoming catastrophic forgetting in neu- ral networks

    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, 5, 6

    work page 2017

  24. [25]

    Deepfakes: a new threat to face recognition? assessment and detection

    Pavel Korshunov and S ´ebastien Marcel. Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685, 2018. 5

    Pith/arXiv arXiv 2018

  25. [26]

    Towards generalized deepfake detection with continual learning on limited new data

    Raj Kumar, Sunil Patel, Minho Lee, and Li Chen. Towards generalized deepfake detection with continual learning on limited new data. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME) , pages 1234– 1239, 2024. 1, 2, 3

    work page 2024

  26. [27]

    Faster than lies: Real-time deepfake detection using binary neural networks

    Romeo Lanzino, Federico Fontana, Anxhelo Diko, Marco Raoul Marini, and Luigi Cinque. Faster than lies: Real-time deepfake detection using binary neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3771–3780, 2024. 3

    work page 2024

  27. [28]

    How Generalizable are Deepfake Image Detectors? An Empirical Study

    Boquan Li, Jun Sun, Christopher M Poskitt, and Xingmei Wang. How generalizable are deepfake image detectors? an empirical study. Arxiv Preprint Arxiv:2308.04177, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [29]

    A continual deepfake detection benchmark: Dataset, meth- ods, and essentials

    Chuqiao Li, Zhiwu Huang, Danda Pani Paudel, Yabin Wang, Mohamad Shahbazi, Xiaopeng Hong, and Luc Van Gool. A continual deepfake detection benchmark: Dataset, meth- ods, and essentials. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1339– 1349, 2023. 1, 2, 3

    work page 2023

  29. [30]

    In ictu oculi: Exposing ai created fake videos by detecting eye blinking

    Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In International Workshop on Information Forensics and Secu- rity (WIFS), pages 1–7. IEEE, 2018. 2

    work page 2018

  30. [31]

    Celeb-df: A large-scale challenging dataset for deep- fake forensics

    Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deep- fake forensics. In Conference on Computer Vision and Pat- tern Recognition, pages 3207–3216, 2020. 2

    work page 2020

  31. [32]

    Exploring disentangled content information for face forgery detection

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

    work page

  32. [33]

    Detecting multimedia generated by large ai models: A survey

    Li Lin, Neeraj Gupta, Yue Zhang, Hainan Ren, Chun-Hao Liu, Feng Ding, Xin Wang, Xin Li, Luisa Verdoliva, and Shu Hu. Detecting multimedia generated by large ai models: A survey. Arxiv Preprint Arxiv:2402.00045, 2024. 3

    Pith/arXiv arXiv 2024

  33. [34]

    Preserving fairness generalization in deepfake detec- tion

    Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detec- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 16815– 16825, 2024. 3

    work page 2024

  34. [35]

    Global tex- ture enhancement for fake face detection in the wild

    Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global tex- ture enhancement for fake face detection in the wild. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060–8069, 2020. 3

    work page 2020

  35. [36]

    Gradient episodic memory for continual learning

    David Lopez-Paz and Marc’Aurelio Ranzato. Gradient episodic memory for continual learning. Advances In Neural Information Processing Systems, 30, 2017. 4

    work page 2017

  36. [37]

    Detecting images generated by deep diffusion models using their lo- cal intrinsic dimensionality

    Peter Lorenz, Ricard L Durall, and Janis Keuper. Detecting images generated by deep diffusion models using their lo- cal intrinsic dimensionality. In International Conference on Computer Vision, pages 448–459, 2023. 3

    work page 2023

  37. [38]

    Incremental learning for the detection and clas- sification of gan-generated images

    Francesco Marra, Cristiano Saltori, Giulia Boato, and Luisa Verdoliva. Incremental learning for the detection and clas- sification of gan-generated images. In 2019 IEEE inter- national workshop on information forensics and security (WIFS), pages 1–6. IEEE, 2019. 3

    work page 2019

  38. [39]

    Deepfakes genera- tion and detection: State-of-the-art, open challenges, coun- termeasures, and way forward

    Momina Masood, Mariam Nawaz, Khalid Mahmood Malik, Ali Javed, Aun Irtaza, and Hafiz Malik. Deepfakes genera- tion and detection: State-of-the-art, open challenges, coun- termeasures, and way forward. Appl. Intell. , pages 1–53,

    work page

  39. [40]

    Gbdf: gender balanced deepfake dataset towards fair deepfake detection

    Aakash Varma Nadimpalli and Ajita Rattani. Gbdf: gender balanced deepfake dataset towards fair deepfake detection. Arxiv Preprint Arxiv:2207.10246, 2022. 3

    Pith/arXiv arXiv 2022

  40. [41]

    Multi-task learning for detecting and segment- ing manipulated facial images and videos

    Huy H Nguyen, Fuming Fang, Junichi Yamagishi, and Isao Echizen. Multi-task learning for detecting and segment- ing manipulated facial images and videos. In International Conference on Biometrics Theory, Applications and Systems, pages 1–8. IEEE, 2019. 2

    work page 2019

  41. [42]

    Glide: Towards photorealistic image genera- tion and editing with text-guided diffusion models

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image genera- tion and editing with text-guided diffusion models. Arxiv Preprint Arxiv:2112.10741, 2021. 3

    Pith/arXiv arXiv 2021

  42. [43]

    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. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8035– 8046, 2023. 3

    work page 2023

  43. [44]

    Learning a deep dual-level network for robust deepfake detection

    Wenbo Pu, Jing Hu, Xin Wang, Yuezun Li, Shu Hu, Bin Zhu, Rui Song, Qi Song, Xi Wu, and Siwei Lyu. Learning a deep dual-level network for robust deepfake detection. Pattern Recogn., 130:108832, 2022. 3

    work page 2022

  44. [45]

    Mobilenetv4: universal models for the mobile ecosystem

    Danfeng Qin, Chas Leichner, Manolis Delakis, Marco Fornoni, Shixin Luo, Fan Yang, Weijun Wang, Colby Ban- bury, Chengxi Ye, Berkin Akin, et al. Mobilenetv4: universal models for the mobile ecosystem. In European Conference on Computer Vision, pages 78–96. Springer, 2024. 4, 5, 6

    work page 2024

  45. [46]

    Hierarchical text-conditional image gener- ation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents. Arxiv Preprint Arxiv:2204.06125, 1 (2):3, 2022. 3

    Pith/arXiv arXiv 2022

  46. [47]

    Imagenet-21k pretraining for the masses

    Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. Arxiv Preprint Arxiv:2104.10972, 2021. 6

    Pith/arXiv arXiv 2021

  47. [48]

    Experience replay for continual learning

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lil- licrap, and Gregory Wayne. Experience replay for continual learning. Advances In Neural Information Processing Sys- tems, 32, 2019. 2, 5, 6

    work page 2019

  48. [49]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 2

    work page 2022

  49. [50]

    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 Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2

    work page 2019

  50. [51]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Adv. Neur. In., 35:36479–36494,

    work page

  51. [52]

    Error sensitivity modulation based experience replay: Mitigating abrupt representation drift in continual learning

    Fahad Sarfraz, Elahe Arani, and Bahram Zonooz. Error sensitivity modulation based experience replay: Mitigating abrupt representation drift in continual learning. In The Eleventh International Conference on Learning Representa- tions, 2023. 2, 5, 6

    work page 2023

  52. [53]

    Con- tinuous fake media detection: Adapting deepfake detectors to new generative techniques.Comput

    Francesco Tassone, Luca Maiano, and Irene Amerini. Con- tinuous fake media detection: Adapting deepfake detectors to new generative techniques.Comput. Vis. Image Und., 249: 104143, 2024. 1, 2, 3

    work page 2024

  53. [54]

    An examination of fairness of ai models for deepfake detection

    Loc Trinh and Yan Liu. An examination of fairness of ai models for deepfake detection. Ijcai, 2021. 3

    work page 2021

  54. [55]

    En- hancing generalization ability in deepfake detection via con- tinual learning

    Shaheen Usmani, Sunil Kumar, and Debanjan Sadhya. En- hancing generalization ability in deepfake detection via con- tinual learning. In Proceedings of the Fifteenth Indian Con- ference on Computer Vision, Graphics and Image Processing (ICVGIP), 2024. 3

    work page 2024

  55. [56]

    Three scenar- ios for continual learning

    Gido M Van de Ven and Andreas S Tolias. Three scenar- ios for continual learning. Arxiv Preprint Arxiv:1904.07734,

    Pith/arXiv arXiv 1904

  56. [57]

    Fastvit: A fast hybrid vision transformer using structural reparameterization

    Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, On- cel Tuzel, and Anurag Ranjan. Fastvit: A fast hybrid vision transformer using structural reparameterization. In Proceed- ings of the IEEE/CVF international conference on computer vision, pages 5785–5795, 2023. 4, 5, 6

    work page 2023

  57. [58]

    Cnn-generated images are sur- prisingly easy to spot

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are sur- prisingly easy to spot... for now. InConference on Computer Vision and Pattern Recognition, pages 8695–8704, 2020. 3

    work page 2020

  58. [59]

    Deepfake detectors and datasets exhibit racial and gender bias, usc study shows

    Kyle Wiggers. Deepfake detectors and datasets exhibit racial and gender bias, usc study shows. In VentureBeat, https: //tinyurl.com/ms8zbu6f, 2021. 3

    work page 2021

  59. [60]

    Con- vnext v2: Co-designing and scaling convnets with masked autoencoders

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Con- vnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16133– 16142, 2023. 4, 5, 6

    work page 2023

  60. [61]

    A comprehensive analysis of ai biases in deepfake de- tection with massively annotated databases

    Ying Xu, Philipp Terh ¨orst, Kiran Raja, and Marius Peder- sen. A comprehensive analysis of ai biases in deepfake de- tection with massively annotated databases. Arxiv Preprint Arxiv:2208.05845, 2022. 3

    Pith/arXiv arXiv 2022

  61. [62]

    Yan et al

    et al. Yan et al. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR) , 2024. Open Access version available on the CVF website. 3

    work page 2024

  62. [63]

    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. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV) , pages 22412–22423, 2023. 3

    work page 2023

  63. [64]

    Generalizing deep- fake video detection with plug-and-play: Video-level blend- ing and spatiotemporal adapter tuning

    Zhiyuan Yan, Yandan Zhao, Shen Chen, Xinghe Fu, Taip- ing Yao, Shouhong Ding, and Li Yuan. Generalizing deep- fake video detection with plug-and-play: Video-level blend- ing and spatiotemporal adapter tuning. Arxiv Preprint Arxiv:2408.17065, 2024. 3

    Pith/arXiv arXiv 2024

  64. [65]

    CrossDF: Improving Cross-Domain Deepfake Detection with Deep Information Decomposition

    Shanmin Yang, Shu Hu, Bin Zhu, Ying Fu, Siwei Lyu, Xi Wu, and Xin Wang. Improving cross-dataset deepfake de- tection with deep information decomposition. Arxiv Preprint Arxiv:2310.00359, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  65. [66]

    Exposing deep fakes using inconsistent head poses

    Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In International Conference on Acoustics, Speech and Signal Processing , pages 8261–

    work page

  66. [67]

    Towards understanding the generalization of deepfake detectors from a game-theoretical view

    Kelu Yao, Jin Wang, Boyu Diao, and Chao Li. Towards understanding the generalization of deepfake detectors from a game-theoretical view. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2031–2041,

    work page 2023

  67. [68]

    Face anti-spoofing via disentangled represen- tation learning

    Ke-Yue Zhang, Taiping Yao, Jian Zhang, Ying Tai, Shouhong Ding, Jilin Li, Feiyue Huang, Haichuan Song, and Lizhuang Ma. Face anti-spoofing via disentangled represen- tation learning. In Computer Vision–ECCV 2020: 16th Eu- ropean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part XIX 16, pages 641–657. Springer, 2020. 3

    work page 2020

  68. [69]

    X-Transfer: A Transfer Learning-Based Framework for GAN-Generated Fake Image Detection

    Lei Zhang, Hao Chen, Shu Hu, Bin Zhu, Xi Wu, Jinrong Hu, and Xin Wang. X-transfer: A transfer learning-based frame- work for robust gan-generated fake image detection. Arxiv Preprint Arxiv:2310.04639, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  69. [70]

    Few-shot learning for misinformation detection based on contrastive models

    Peng Zheng, Hao Chen, Shu Hu, Bin Zhu, Jinrong Hu, Ching-Sheng Lin, Xi Wu, Siwei Lyu, Guo Huang, and Xin Wang. Few-shot learning for misinformation detection based on contrastive models. Electronics, 13(4):799, 2024. 3

    work page 2024

  70. [71]

    Wilddeepfake: A challenging real-world dataset for deepfake detection

    Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2382– 2390, 2020. 5

    work page 2020

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.