Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

Adaptive Meta-Learning for Robust Deepfake Detection: A Multi-Agent Framework to Data Drift and Model Generalization

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an adversarial meta-learning algorithm with a refinement phase, combined with a hierarchical multi-agent retrieval-augmented generation workflow for custom sample synthesis, makes deepfake detectors generalize to…

desk verdict Interesting integration of meta-learning and agent-based sample synthesis, but the headline experiments conflate algorithm with training-data scale; the generalization claim needs a controlled baseline. read the letter →

arxiv 2411.08148 v1 pith:NXG4DYPD submitted 2024-11-12 cs.AI

Dinesh Srivasthav P , Badri Narayan Subudhi This is my paper · ORCID
classification cs.AI
keywords deepfakedetectionmeta-learninggeneralizationadversarialrobustnesssamplesynthesisdatadriftretrieval-augmentedgenerationmulti-agentworkflow
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Deepfake detectors trained on static benchmarks fail in three practical ways: they do not generalize to unseen or cross-domain forgeries, they are brittle under tiny adversarial perturbations, and they do not adapt as new generation techniques appear. This paper claims to address all three at once by wrapping the Reptile meta-learning algorithm in a refinement phase that, for every task, generates synthetic samples for the examples the model gets wrong and adversarial samples for the examples it gets right, then trains with a unified loss adding consistency regularization and a margin ranking term. On top of that, it introduces a hierarchical multi-agent retrieval-augmented generation pipeline that synthesizes custom deepfake images from current attack patterns, so the training set can be refreshed as data drift occurs. Experimental results show the meta-trained model reaching 61.5 percent accuracy on a completely unseen dataset where fully trained transformer baselines fall below 50 percent, and staying consistent on two other held-out datasets.

What carries the argument

The load-bearing mechanism is a refinement phase appended to Reptile, a first-order meta-learning algorithm whose outer loop moves model weights toward the solutions found in inner task updates. In each inner step, the support samples are ranked by $M_{\text{adaptive}} = -(p_y - H(p) + \text{Margin} - 2 \cdot i_{\text{misclassified}} \cdot (1 + \text{Margin}))$, where $p_y$ is the probability of the true class, $H(p)$ is the entropy of the prediction, and $\text{Margin}$ is the gap between the top and second prediction. Misclassified samples with large margin are ranked first and get synthetic augmentations; correctly classified samples with large margin are ranked first and get adversarial perturbations. The new samples are added to the task's support set, and training minimizes $L_{\text{unified}} = L_{\text{base}} + \lambda_1 L_{\text{TSAC}} + \lambda_2 L_{\text{ADV}}$, where $L_{\text{base}}$ is Reptile's inner loss, $L_{\text{TSAC}}$ is a weighted contrastive loss pulling feature representations of original and synthetic samples together for same-class pairs, and $L_{\text{ADV}}$ is a margin ranking loss $\max(0, m - (f(x_i) - f(x_{\text{adv},i})))$ on original versus adversarial representations. The second mechanism is a hierarchical multi-agent retrieval-augmented generation workflow whose agents produce attack patterns and few-shot prompts, which a sample-synthesis module uses to generate custom deepfake images and keep the training distribution current.

What would settle it

Ablate the refinement phase: train the same architecture with Reptile alone on the same five datasets and evaluate on the OpenForensics-based test set; if the accuracy is close to the reported 0.6151, the claimed gains come from the meta-training dataset mixture rather than from the new losses and sample-selection mechanism.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a few-shot meta-trained model can outperform models fully trained on the same benchmark when all are evaluated on data none of them saw during training. The meta model uses a refinement phase atop Reptile: for each inner task it ranks support samples with the Madaptive score, augments the hardest misclassified samples, adversarially perturbs the most confidently classified ones, adds those new samples to the support set, and optimizes a unified loss made of Reptile's inner loss, a weighted contrastive consistency loss, and an adversarial margin ranking loss. On the unseen OpenForensics-based test set the meta model obtains accuracy 0.6151, AUC 0.6042, and F1 0.6319, while the same architecture trained without meta-learning reaches 46.49 percent accuracy and the other fully trained transformers stay near or below 50 percent. The paper interprets this as evidence that the refinement phase and the diverse multi-dataset meta-training jointly improve generalization and robustness, and that the retrieval-augmented generation and agent workflow can supply synthetic few-shot samples to keep the model aligned with emerging attack patterns.

Load-bearing premise

The method assumes that if the model is trained on extra copies of its hardest mistakes and adversarial copies of its most confident successes, it will become both more general and more robust, even though the score it is told to separate is never concretely defined.

Editorial extensions

If this is right

  • A few-shot meta-trained detector can outperform fully trained models on a benchmark no model has seen, so expensive full training on a single dataset may not be the best route to cross-domain deepfake detection.
  • The same refinement phase, with sample synthesis and consistency regularization, transfers in principle to any classification task where generalization and adversarial robustness are the bottlenecks.
  • The hierarchical multi-agent retrieval-augmented generation workflow offers a way to inject newly emerging attack patterns into training without manual dataset collection, directly addressing data drift.
  • Meta-task design matters: increasing the number of classes per task lowers query accuracy, while increasing shots per class raises it, giving a practical guideline for building the meta-dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct ablation separating Reptile-only from the full refinement phase would be needed to know how much of the 15-point gain comes from the new losses, because the comparison baselines are trained on DeepFakeFace alone rather than on the same five-dataset mixture.
  • The margin ranking loss, as written, rewards adversarial samples with lower feature values than clean samples; that is a sensitivity signal rather than an invariance signal, so the robustness interpretation would be stronger if tested with feature-distance or certified-robustness measures.
  • Because the reported AUC values are around 0.60, the practical implication is that meta-learning plus synthetic sample generation is a promising direction rather than a deployable detector; future work might target confidence calibration and higher-capacity training.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes a framework for deepfake detection combining an adversarial meta-learning algorithm (built on Reptile) with a hierarchical multi-agent retrieval-augmented generation (RAG) workflow for synthetic sample generation. The meta-learning component includes task-specific adaptive sample selection via a proposed M_adaptive metric, sample synthesis/augmentation for misclassified samples, adversarial perturbation for correctly classified samples, and two auxiliary losses: a weighted contrastive consistency loss (Eq. 1) and a margin-ranking adversarial loss (Eq. 2). The RAG workflow is designed to produce few-shot prompts and generate synthetic deepfake images to address data drift. The authors claim the framework boosts both robustness and generalization, and report experiments on a meta-dataset of five training datasets with evaluation on an unseen OpenForensics-based set (Table VI) and on DGM and iFakeFaceDB test sets (Table VII), where the Meta model outperforms several transformer baselines trained only on DeepFakeFace.

Significance. If validated, the proposed direction—combining meta-learning with adaptive sample selection and agent-driven synthetic data generation—could be relevant for deepfake detection under data drift. The paper explicitly aims at three practically important challenges (generalization, adversarial robustness, and adaptability) and offers a system-level architecture. However, the empirical evidence as presented does not isolate the contribution of the proposed method: the main comparison conflates the algorithm with a much larger and more diverse training set, and no adversarial robustness evaluation is provided. The method's novelty is plausible, but its significance cannot be assessed from the current experiments. On the positive side, the paper documents a concrete experimental protocol, a large meta-dataset, and specific attack/augmentation lists, which would aid reproducibility if code and hyperparameters were provided.

major comments (5)
  1. [Section IV, Tables VI and VII] The central comparison is confounded by training data scale and composition. The Meta model is meta-trained on 596k samples from five datasets, while every baseline is trained only on the DeepFakeFace dataset. The reported gains on OpenForensics (Table VI: Meta Acc 0.6151 vs. CoaT 0.4649) and on DGM/iFakeFaceDB (Table VII) may therefore reflect training-set size and diversity rather than the proposed meta-learning algorithm. No control experiment trains a same-architecture baseline on the full meta-dataset, or ablates the proposed losses and selection mechanism. Without such a controlled comparison, the paper's claim that the algorithm 'boosts both robustness and generalization' is not established.
  2. [Section III-B, Eq. (2)] The adversarial margin ranking loss, max(0, m - (f(x_i) - f(x_adv_i))), requires f to be defined, but f is never specified. If f is a class score or feature scalar, the loss explicitly drives f(x_adv) below f(x) by a margin, which encourages the model to change its output under adversarial perturbation rather than to be invariant to it—the opposite of the usual goal of adversarial robustness. The paper also provides no adversarial evaluation on any test set, so the robustness claim is unsupported. Please define f, justify the loss against a standard robustness notion, and report accuracy/AUC under the training-time attacks (e.g., FGSM, PGD, AutoAttack).
  3. [Section III-B, M_adaptive definition] There is an internal inconsistency between the formula for M_adaptive and the claimed ranking order. For misclassified samples, the formula reduces to -p_y + H(p) + Margin + 2, which increases with Margin. The text states that misclassified samples are 'ranked in ascending order of M_adaptive' and that this yields 'Wrong predictions with large margin > Wrong predictions with small margin > ...' — but ascending order would place wrong predictions with small margin first. Either the intended ordering is descending, or the sign of the misclassification term is wrong. Since M_adaptive is the core sample-selection mechanism, this must be corrected and clarified.
  4. [Section IV, results] No ablation study isolates any component of the proposed method. The framework includes the M_adaptive selection, the consistency loss (Eq. 1), the margin ranking loss (Eq. 2), the sample synthesis/augmentation strategy, and the Reptile meta-learning itself. Tables VI–VII only report the full system versus standard training. As a result, even if the comparison were fair, the reader cannot tell which component drives the reported improvement. Please add ablations that remove each term or replace the selection with random sampling.
  5. [Section IV, adversarial robustness claim] The paper motivates the work with adversarial robustness and lists 15 adversarial attacks used during training (Table III), but the evaluation in Tables VI–VII uses only clean test images. There is no evidence that the trained model is robust to any perturbation, nor any comparison with a baseline trained without the adversarial loss. A robustness claim requires evaluation under adversarial perturbations, ideally with a defense-aware attack (e.g., AutoAttack).
minor comments (6)
  1. [Abstract] The abstract states 'The associated code is available here' but no URL or repository is provided in the text. Please include a working link.
  2. [Table V and text, Section IV] Table V lists 'Swin tl' as one of the model architectures, and the text mentions its OpenForensics accuracy (50.7%), but Tables VI and VII do not report Swin tl results. Please include these numbers or remove the discussion.
  3. [Section III-B, Eq. (1)] The notation 'W (i) t' is rendered confusingly; the superscript and subscript order should be clarified (e.g., W_t^{(i)}). Also, the ContrastiveLoss formula's second line appears to be missing a closing parenthesis and the square notation is inconsistent.
  4. [Section IV, Table III] The distinction between 'synthetic samples' and 'augmentations' is unclear. The sample synthesis described in the algorithm appears to be standard image augmentation (flip, rotation, color jitter), which is not the same as the generative-model-based synthesis in the RAG module. Please clarify terminology.
  5. [Various] Several typos occur, including 'bussinessman' in Figure 3, 'DeepFaceFace' for DeepFakeFace in Section IV, and an orphan citation '[46]' at the end of a paragraph. The paper should be carefully proofread.
  6. [Section V] The conclusion states that the proposed workflow 'helps in collecting real-time information...' but no experiment evaluates the multi-agent RAG workflow or the data-drift adaptation capability. The manuscript should either present such an evaluation or clearly state that this component is a proposed system without empirical validation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the proposed losses and M_adaptive metric are heuristic training objectives, the OpenForensics evaluation is on a genuinely unseen dataset, and no self-citation chain is load-bearing.

full rationale

The paper's derivation chain consists of: defining M_adaptive to rank hard and confident samples; generating synthetic and adversarial samples for those samples; adding them to task support sets; and training with the unified loss L_unified = L_base + λ1·L_TSAC + λ2·L_ADV. None of these components is defined in terms of the evaluation targets. L_TSAC and L_ADV are standard contrastive and margin-ranking losses with adaptive weights; they do not incorporate test-set labels or the reported accuracy/AUC/F1 values. The reported OpenForensics result is on a dataset excluded from meta-training, so it is genuinely external and not forced by construction. The DGM and iFakeFaceDB evaluations use test splits whose train splits were part of the 596k meta-training set; this is a comparison-protocol confound because the baselines were trained only on DeepFakeFace and never saw even the train splits of those datasets. However, that is an experimental-design weakness, not circularity: the test labels were not used for fitting, and the paper explicitly acknowledges that the Meta model was trained on the train splits. No equation reduces to another equation by construction, and no load-bearing claim depends on a self-citation or a uniqueness theorem imported from the authors. The unverified adversarial-robustness claim and the lack of controlled same-data baselines are correctness risks, not circular steps. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The framework rests on several assumptions inherited from prior work or introduced ad hoc: Reptile works as a meta-learner, generated synthetic images are valid deepfake examples, the M_adaptive metric ranks samples as claimed, and the margin ranking loss with an undefined f yields robustness. The free parameters are hyperparameters chosen without sensitivity analysis, and the M_adaptive constants are hand-picked. M_adaptive is a new scoring construct without independent evidence.

free parameters (5)
  • lambda1 = 0.5
    Weight on the consistency loss in Eq. (3); chosen by hand with no sensitivity analysis reported in Section IV.
  • lambda2 = 0.5
    Weight on the adversarial loss in Eq. (3); chosen by hand with no sensitivity analysis reported in Section IV.
  • margin m = not specified
    Margin in the contrastive loss and margin ranking loss, Eqs. (1) and (2); no value or tuning procedure is given.
  • top-k selection size = not specified
    Number of representational samples chosen for sample synthesis and adversarial generation per task; the paper does not state k or the sampling rule.
  • M_adaptive formula constants = 2 and 1
    Constants in the M_adaptive formula in Section III-B; they are chosen ad hoc to produce the claimed ranking order, which is not verified.
assumptions (4)
  • domain assumption Reptile meta-learning converges to a good initialization for few-shot adaptation.
    Cited as [20] and used as the base algorithm without re-derivation or verification in this paper's setting.
  • ad hoc to paper Synthetic images generated by Stable Diffusion inpainting with Grounding DINO and SAM masks are valid deepfake training examples.
    Section III-A3 assumes generated images are representative of real deepfake attacks; no quality filter or human verification is described.
  • ad hoc to paper The M_adaptive metric correctly ranks difficult and confident samples in the claimed order.
    Section III-B states the ranking order, but the formula and the stated order appear inconsistent, and the metric is not validated against any external measure.
  • ad hoc to paper The margin ranking loss with the undefined feature function f improves adversarial robustness.
    Eq. (2) is introduced without defining f; as written it appears to encourage separation between clean and adversarial samples, which is the opposite of invariance.
invented entities (1)
  • M_adaptive sample ranking metric
    purpose: Selects the most difficult misclassified samples for augmentation and the most confidently classified samples for adversarial perturbation.
    Introduced in Section III-B; no external validation of the ranking behavior is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Meta-Learning for Robust Deepfake Detection: A Multi-Agent Framework to Data Drift and Model Generalization." pith.science (2026). https://pith.science/paper/NXG4DYPD

@misc{pith2026241108148,
  author       = {Pith},
  title        = {Pith review of: Adaptive Meta-Learning for Robust Deepfake Detection: A Multi-Agent Framework to Data Drift and Model Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NXG4DYPD}},
  note         = {Machine review of arXiv:2411.08148}
}
read the original abstract

Pioneering advancements in artificial intelligence, especially in genAI, have enabled significant possibilities for content creation, but also led to widespread misinformation and false content. The growing sophistication and realism of deepfakes is raising concerns about privacy invasion, identity theft, and has societal, business impacts, including reputational damage and financial loss. Many deepfake detectors have been developed to tackle this problem. Nevertheless, as for every AI model, the deepfake detectors face the wrath of lack of considerable generalization to unseen scenarios and cross-domain deepfakes. Besides, adversarial robustness is another critical challenge, as detectors drastically underperform to the slightest imperceptible change. Most state-of-the-art detectors are trained on static datasets and lack the ability to adapt to emerging deepfake attack trends. These three crucial challenges though hold paramount importance for reliability in practise, particularly in the deepfake domain, are also the problems with any other AI application. This paper proposes an adversarial meta-learning algorithm using task-specific adaptive sample synthesis and consistency regularization, in a refinement phase. By focussing on the classifier's strengths and weaknesses, it boosts both robustness and generalization of the model. Additionally, the paper introduces a hierarchical multi-agent retrieval-augmented generation workflow with a sample synthesis module to dynamically adapt the model to new data trends by generating custom deepfake samples. The paper further presents a framework integrating the meta-learning algorithm with the hierarchical multi-agent workflow, offering a holistic solution for enhancing generalization, robustness, and adaptability. Experimental results demonstrate the model's consistent performance across various datasets, outperforming the models in comparison.

Figures

Figures reproduced from arXiv: 2411.08148 by the authors.

Figure 1
Figure 1. Proposed framework (all the olive green blocks are the additions we added to the Reptile algorithm) [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Hierarchical multi-agent workflow for custom deepfake sample synthesis [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Output of few-shot prompts from the multi-agent hierarchical workflow for the positive prompt: ”A bussinessman [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Expression Swap + Age Progression – An Image inpainted through the Sample Synthesis Module [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Attribute Swap – Few-shot samples generated from the Hierarchical workflow module based on few-shot prompts [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: N-Way vs K-Shot [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Sample of images post the application of ensemble [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Sample of images post the application of ensemble [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: N-Way vs Query Set Accuracy A total of six different datasets were considered as described in Table II. The OpenForensics-based dataset was kept for unseen test evaluation, and the other five datasets were used for meta-training. Each of the individual datasets have tw…
Figure 10
Figure 10. Figure 10: K-Shot vs Query Set Accuracy will require even more computational bandwidth than training a classifier in the typical way. Moreover, meta￾training in accordance to the proposed algorithm that further involves a refinement phase with the prospects of task-specific adap…
Figure 11
Figure 11. Figure 11: Comparison of test accuracy for different models on the primary test set (DeepFakeface), and on an unseen test [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Graph-based Fake Account Detection: A Survey

    cs.SI 2025-07 conditional novelty 4.0 of 10

    A structured survey of graph-based fake account detection methods, organizing classical, traditional machine learning, and deep learning approaches and their datasets.

  2. Unmasking Synthetic Realities in Generative AI: A Comprehensive Review of Adversarially Robust Deepfake Detection Systems

    cs.CR 2025-07 conditional novelty 3.0 of 10

    A systematic review of deepfake detection finds a pervasive lack of adversarial robustness evaluation across all modalities and calls for resilient, modality-agnostic detectors.

Reference graph

Works this paper leans on

56 extracted references · 55 canonical work pages · cited by 2 Pith papers

  1. [1]

    Deep fake video detection using transfer learning approach,

    S. Suratkar and F. Kazi, “Deep fake video detection using transfer learning approach,” Arabian journal for science and engineering, vol. 48, pp. 1–11, 10 2022

    work page 2022

  2. [2]

    Fretal: Generalizing deep- fake detection using knowledge distillation and representation learning,

    M. Kim, S. Tariq, and S. S. Woo, “Fretal: Generalizing deep- fake detection using knowledge distillation and representation learning,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 1001–1012, 2021

    work page 2021

  3. [3]

    Domain generalization for face forgery detection by style transfer,

    T. Kim, J. Choi, H. Cho, H. Lim, and J. Choi, “Domain generalization for face forgery detection by style transfer,” in 2024 IEEE International Conference on Consumer Electronics (ICCE), pp. 1–5, 2024

    work page 2024

  4. [4]

    Robustness and generalizability of deepfake detection: A study with diffusion models,

    H. Song, S. Huang, Y . Dong, and W.-W. Tu, “Robustness and generalizability of deepfake detection: A study with diffusion models,” 2023

    work page 2023

  5. [5]

    How generalizable are deepfake image detectors? an empirical study,

    B. Li, J. Sun, C. M. Poskitt, and X. Wang, “How generalizable are deepfake image detectors? an empirical study,” 2024

    work page 2024

  6. [6]

    A review of deep learning-based approaches for deepfake content detection,

    L. A. Passos, D. Jodas, K. A. P. Costa, L. A. Souza J ´unior, D. Rodrigues, J. Del Ser, D. Camacho, and J. P. Papa, “A review of deep learning-based approaches for deepfake content detection,” Expert Systems , vol. 41, Feb. 2024

    work page 2024

  7. [7]

    Revisiting generalizability in deepfake detection: Improving metrics and stabilizing transfer,

    S. Kamat, S. Agarwal, T. Darrell, and A. Rohrbach, “Revisiting generalizability in deepfake detection: Improving metrics and stabilizing transfer,” in Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV) Workshops , pp. 426–435, October 2023

    work page 2023

  8. [8]

    On the vulnerability of deepfake detectors to attacks generated by denoising diffusion models,

    M. Ivanovska and V . Struc, “On the vulnerability of deepfake detectors to attacks generated by denoising diffusion models,” in 2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW) , (Los Alamitos, CA, USA), pp. 1051–1060, IEEE Computer Society, jan 2024

    work page 2024

Show all 56 references
  1. [9]

    Evading deepfake-image detectors with white- and black-box attacks,

    N. Carlini and H. Farid, “Evading deepfake-image detectors with white- and black-box attacks,” in2020 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 2804–2813, 2020

    2020

  2. [10]

    Evad- ing deepfake detectors via adversarial statistical consistency,

    Y . Hou, Q. Guo, Y . Huang, X. Xie, L. Ma, and J. Zhao, “Evad- ing deepfake detectors via adversarial statistical consistency,” 2023

    2023

  3. [11]

    Adversarially robust deepfake video detection,

    A. Devasthale and S. Sural, “Adversarially robust deepfake video detection,” in 2022 IEEE Symposium Series on Compu- tational Intelligence (SSCI) , pp. 396–403, 2022

    2022

  4. [12]

    D4: Detection of adversarial diffusion deepfakes using disjoint ensembles,

    A. Hooda, N. Mangaokar, R. Feng, K. Fawaz, S. Jha, and A. Prakash, “D4: Detection of adversarial diffusion deepfakes using disjoint ensembles,” 2023

    2023

  5. [13]

    Adversari- ally robust deepfake detection via adversarial feature similarity learning,

    S. Khan, J.-C. Chen, W.-H. Liao, and C.-S. Chen, “Adversari- ally robust deepfake detection via adversarial feature similarity learning,” in MultiMedia Modeling: 30th International Confer- ence, MMM 2024, Amsterdam, The Netherlands, January 29 – February 2, 2024, Proceedings, ...

    2024

  6. [14]

    Adver- sarial threats to deepfake detection: A practical perspective,

    P. Neekhara, B. Dolhansky, J. Bitton, and C. C. Ferrer, “Adver- sarial threats to deepfake detection: A practical perspective,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pp. 923–932, 2021

    2021

  7. [15]

    2d- malafide: Adversarial attacks against face deepfake detection systems,

    C. Galdi, M. Panariello, M. Todisco, and N. Evans, “2d- malafide: Adversarial attacks against face deepfake detection systems,” 2024

    2024

  8. [16]

    Metamorphic testing-based adversarial attack to fool deepfake detectors,

    N. Lim, M. Y . Kuan, M. Pu, M. Lim, and C. Y . Chong, “Metamorphic testing-based adversarial attack to fool deepfake detectors,” in 2022 26th International Conference on Pattern Recognition (ICPR) , (Los Alamitos, CA, USA), pp. 2503– 2509, IEEE Computer Society, aug 2022

    2022

  9. [17]

    Evading deepfake detectors via high quality face pre-processing methods,

    J. Kim, T. Kim, J. Kim, and S. S. Woo, “Evading deepfake detectors via high quality face pre-processing methods,” in 2022 26th International Conference on Pattern Recognition (ICPR), (Los Alamitos, CA, USA), pp. 1937–1944, IEEE Computer Society, aug 2022

    2022

  10. [18]

    Making deepfakes more spurious: Evading deep face forgery detection via trace removal attack,

    C. Liu, H. Chen, T. Zhu, J. Zhang, and W. Zhou, “Making deepfakes more spurious: Evading deep face forgery detection via trace removal attack,” IEEE Transactions on Dependable and Secure Computing , vol. 20, pp. 5182–5196, nov 2023

    2023

  11. [19]

    Continuous fake media detection: Adapting deepfake detectors to new generative tech- niques,

    F. Tassone, L. Maiano, and I. Amerini, “Continuous fake media detection: Adapting deepfake detectors to new generative tech- niques,” Computer Vision and Image Understanding , vol. 249, p. 104143, 2024

    2024

  12. [20]

    On first-order meta- learning algorithms,

    A. Nichol, J. Achiam, and J. Schulman, “On first-order meta- learning algorithms,” 2018

    2018

  13. [21]

    Deep learning for deepfakes creation and de- tection: A survey,

    T. T. Nguyen, Q. V . H. Nguyen, D. T. Nguyen, D. T. Nguyen, T. Huynh-The, S. Nahavandi, T. T. Nguyen, Q.-V . Pham, and C. M. Nguyen, “Deep learning for deepfakes creation and de- tection: A survey,”Computer Vision and Image Understanding, vol. 223, p. 103525, Oct. 2022

    2022

  14. [22]

    The creation and detection of deep- fakes: A survey,

    Y . Mirsky and W. Lee, “The creation and detection of deep- fakes: A survey,” ACM Computing Surveys , vol. 54, p. 1–41, Jan. 2021

    2021

  15. [23]

    Deepfakes generation and detection: A short survey,

    Z. Akhtar, “Deepfakes generation and detection: A short survey,” Journal of Imaging , vol. 9, no. 1, 2023

    2023

  16. [24]

    Are face detection models biased?,

    S. Mittal, K. Thakral, P. Majumdar, M. Vatsa, and R. Singh, “Are face detection models biased?,” 2022

    2022

  17. [25]

    Df-platter: Multi-face heterogeneous deepfake dataset,

    K. Narayan, H. Agarwal, K. Thakral, S. Mittal, M. Vatsa, and R. Singh, “Df-platter: Multi-face heterogeneous deepfake dataset,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 9739–9748, 2023

    2023

  18. [26]

    Deephy: On deepfake phylogeny,

    K. Narayan, H. Agarwal, K. Thakral, S. Mittal, M. Vatsa, and R. Singh, “Deephy: On deepfake phylogeny,” 2022

    2022

  19. [27]

    Detecting and grounding multi-modal media manipulation and beyond,

    R. Shao, T. Wu, J. Wu, L. Nie, and Z. Liu, “Detecting and grounding multi-modal media manipulation and beyond,” IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 46, no. 8, pp. 5556–5574, 2024

    2024

  20. [28]

    Deter: Detecting edited regions for deterring generative manipulations,

    S. Wang, Y . Zhu, R. Wang, A. Dharmasiri, O. Russakovsky, and Y . Wu, “Deter: Detecting edited regions for deterring generative manipulations,” 2023

    2023

  21. [29]

    Linguistic profiling of deepfakes: An open database for next-generation deepfake detection,

    Y . Wang, Z. Huang, Z. Ma, and X. Hong, “Linguistic profiling of deepfakes: An open database for next-generation deepfake detection,” ArXiv, vol. abs/2401.02335, 2024

    2024 arXiv

  22. [30]

    Model attribution of face-swap deepfake videos,

    S. Jia, X. Li, and S. Lyu, “Model attribution of face-swap deepfake videos,” in 2022 IEEE International Conference on Image Processing (ICIP) , pp. 2356–2360, 2022

    2022

  23. [31]

    Au- tosplice: A text-prompt manipulated image dataset for media forensics,

    S. Jia, M. Huang, Z. Zhou, Y . Ju, J. Cai, and S. Lyu, “Au- tosplice: A text-prompt manipulated image dataset for media forensics,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pp. 893–903, 2023

    2023

  24. [32]

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

    P. Korshunov and S. Marcel, “Deepfakes: a new threat to face recognition? assessment and detection,” ArXiv, vol. abs/1812.08685, 2018

    2018 arXiv

  25. [33]

    On the detection of digital face manipulation,

    H. Dang, F. Liu, J. Stehouwer, X. Liu, and A. K. Jain, “On the detection of digital face manipulation,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), (Los Alamitos, CA, USA), pp. 5780–5789, IEEE Computer Society, jun 2020

    2020

  26. [34]

    Ganprintr: Improved fakes and evaluation of the state of the art in face manipulation detec- tion,

    J. C. Neves, R. Tolosana, R. Vera-Rodriguez, V . Lopes, H. Proenc ¸a, and J. Fierrez, “Ganprintr: Improved fakes and evaluation of the state of the art in face manipulation detec- tion,” IEEE Journal of Selected Topics in Signal Processing , vol. 14, no. 5, pp. 1038–1048, 2020

    2020

  27. [35]

    Celeb-df: A large-scale challenging dataset for deepfake forensics,

    Y . Li, X. Yang, P. Sun, H. Qi, and S. Lyu, “Celeb-df: A large-scale challenging dataset for deepfake forensics,” 2020 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR) , pp. 3204–3213, 2019

    2020

  28. [36]

    Faceforensics++: Learning to detect ma- nipulated facial images,

    A. Rossler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Niessner, “Faceforensics++: Learning to detect ma- nipulated facial images,” in 2019 IEEE/CVF International Conference on Computer Vision (ICCV) , (Los Alamitos, CA, USA), pp. 1–11, IEEE Computer Society, nov 2019

    2019

  29. [37]

    Face forensics in the wild,

    T. Zhou, W. Wang, Z. Liang, and J. Shen, “Face forensics in the wild,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 5774–5784, 2021

    2021

  30. [38]

    Dfgc 2021: A deepfake game competition,

    B. Peng, H. Fan, W. Wang, J. Dong, Y . Li, S. Lyu, Q. Li, Z. Sun, H. Chen, B. Chen, Y . Hu, S. Luo, J. Huang, Y . Yao, B. Liu, H. Ling, G. Zhang, Z. Xu, C. Miao, C. Lu, S. He, X. Wu, and W. Zhuang, “Dfgc 2021: A deepfake game competition,” in 2021 IEEE International Joint Conf...

    2021

  31. [39]

    Wilddeep- fake: A challenging real-world dataset for deepfake detection,

    B. Zi, M. Chang, J. Chen, X. Ma, and Y .-G. Jiang, “Wilddeep- fake: A challenging real-world dataset for deepfake detection,” Proceedings of the 28th ACM International Conference on Multimedia, 2020

    2020

  32. [40]

    Exposing deep fakes using inconsistent head poses,

    X. Yang, Y . Li, and S. Lyu, “Exposing deep fakes using inconsistent head poses,” in ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8261–8265, 2019

    2019

  33. [41]

    Deepfake generation, detection and datasets: a rapid-review,

    A. KoC ¸ ak and M. Alkan, “Deepfake generation, detection and datasets: a rapid-review,” in 2022 15th International Conference on Information Security and Cryptography (ISC- TURKEY), pp. 86–91, 2022

    2022

  34. [42]

    Ai vs. human vision: A comparative analysis for distinguishing ai-generated and natural images,

    R. Purohit, Y . Sane, D. Vaishampayan, S. Vedantam, and M. Singh, “Ai vs. human vision: A comparative analysis for distinguishing ai-generated and natural images,” in 2024 F ourth International Conference on Advances in Electri- cal, Computing, Communication and Sustainable Te...

    2024

  35. [43]

    Bobulski and M

    J. Bobulski and M. Kubanek, Detection of Fake Facial Images and Changes in Real Facial Images , pp. 110–122. Computa- tional Collective Intelligence, 08 2024

    2024

  36. [44]

    Deepfake on face and expression swap: A review,

    S. Waseem, S. A. R. S. Abu Bakar, B. A. Ahmed, Z. Omar, T. A. E. Eisa, and M. E. E. Dalam, “Deepfake on face and expression swap: A review,”IEEE Access, vol. 11, pp. 117865– 117906, 2023

    2023

  37. [45]

    Deepfake detection for human face images and videos: A survey,

    A. Malik, M. Kuribayashi, S. M. Abdullahi, and A. N. Khan, “Deepfake detection for human face images and videos: A survey,” IEEE Access , vol. 10, pp. 18757–18775, 2022

    2022

  38. [46]

    A systematic review on fake image creation techniques,

    R. Chauhan, R. Popli, and I. Kansal, “A systematic review on fake image creation techniques,” in 2023 10th International Conference on Computing for Sustainable Global Develop- ment (INDIACom), pp. 779–783, 2023

    2023

  39. [47]

    Open-set deepfake detection to fight the unknown,

    M. M. Diniz and A. Rocha, “Open-set deepfake detection to fight the unknown,” in ICASSP 2024 - 2024 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 13091–13095, 2024

    2024

  40. [48]

    Advshadow: Evading deepfake detection via adversarial shadow attack,

    J. Liu, M. Zhang, J. Ke, and L. Wang, “Advshadow: Evading deepfake detection via adversarial shadow attack,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 4640–4644, 2024

    2024

  41. [49]

    Evading deepfake-image detectors with white- and black-box attacks,

    N. Carlini and H. Farid, “Evading deepfake-image detectors with white- and black-box attacks,” in IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 2804–2813, 06 2020

    2020

  42. [50]

    An analysis of recent advances in deepfake image detection in an evolving threat landscape,

    S. Abdullah, A. Cheruvu, S. Kanchi, T. Chung, P. Gao, M. Jadliwala, and B. Viswanath, “An analysis of recent advances in deepfake image detection in an evolving threat landscape,” in 2024 IEEE Symposium on Security and Privacy (SP), (Los Alamitos, CA, USA), pp. 91–109, IEEE Co...

    2024

  43. [51]

    Deepfake attacks: Generation, detection, datasets, challenges, and research directions,

    A. Naitali, M. Ridouani, F. Salahdine, and N. Kaabouch, “Deepfake attacks: Generation, detection, datasets, challenges, and research directions,” Comput., vol. 12, p. 216, 2023

    2023

  44. [52]

    Detecting and grounding multi- modal media manipulation,

    R. Shao, T. Wu, and Z. Liu, “Detecting and grounding multi- modal media manipulation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2023

    2023

  45. [53]

    GANprintR: Improved Fakes and Evaluation of the State-of-the-Art in Face Manipulation Detection,

    J. C. Neves, R. Tolosana, R. Vera-Rodriguez, V . Lopes, H. Proenc ¸a, and J. Fierrez, “GANprintR: Improved Fakes and Evaluation of the State-of-the-Art in Face Manipulation Detection,” 2019

    2019

  46. [54]

    Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization,

    F. Guillaro, D. Cozzolino, A. Sud, N. Dufour, and L. Ver- doliva, “Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 20606–20615, June 2023

    2023

  47. [55]

    Deep- fakebench: A comprehensive benchmark of deepfake detec- tion,

    Z. Yan, Y . Zhang, X. Yuan, S. Lyu, and B. Wu, “Deep- fakebench: A comprehensive benchmark of deepfake detec- tion,” in Advances in Neural Information Processing Systems (A. Oh, T. Neumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, eds.), vol. 36, pp. 4534–4565, Curran...

    2023

  48. [56]

    Openforensics: Large-scale challenging dataset for multi-face forgery detection and segmentation in-the-wild,

    T.-N. Le, H. H. Nguyen, J. Yamagishi, and I. Echizen, “Openforensics: Large-scale challenging dataset for multi-face forgery detection and segmentation in-the-wild,” in Interna- tional Conference on Computer Vision , 2021

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.