Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Common face-enhancement filters can be repurposed as anti-forensic attacks on deepfake detectors, with GAN-based beautification reaching 75% attack success.

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 ↗

T0 review

2026-08-04 22:42 UTC pith:HIVVIWNS

load-bearing objection A useful, internally consistent empirical matrix of face-enhancement attacks, but the missing no-enhancement ASR baseline undercuts the headline numbers and needs fixing before I'd trust the magnitudes. the 2 major comments →

arxiv 2509.07178 v1 pith:HIVVIWNS submitted 2025-09-08 cs.CV

Realism to Deception: Investigating Deepfake Detectors Against Face Enhancement

Muhammad Saad Saeed , Ijaz Ul Haq , Khalid Malik This is my paper
classification cs.CV
keywords face enhancementdeepfake detectionanti-forensic attackattack success rateGAN-based restorationadversarial trainingfrequency artifactsFaceForensics++
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.

This paper argues that common face-enhancement techniques—basic smoothing filters and GAN-based beautification/restoration—can function as anti-forensic attacks against deepfake detectors. On FaceForensics++, DeepFakeDetection, and CelebDF-v2, the authors apply six enhancement methods to fake video frames and measure how often six detectors (naive, spatial, and frequency-based) misclassify the enhanced fakes as real. They report that even simple Gaussian and bilateral smoothing push attack success rates into the 36–64% range, while GAN-based methods, especially FabSoften+GFPGAN, reach average ASRs around 73–75% with individual combinations as high as 87.39%. The finding matters because these are everyday, model-agnostic operations available in consumer editing apps, so detection systems trained on clean forgeries may fail under realistic post-processing. The paper also shows adversarial training on one enhancement type does not reliably transfer to others, and closes by urging forensic models that are resilient to distributional shifts rather than just adversarial noise.

Core claim

The paper's central claim is that face enhancement is a viable anti-forensic pathway: transformations designed to improve appearance suppress the low-level artifacts—texture irregularities, upsampling traces, and high-frequency anomalies—that deepfake detectors key on. The evidence is a systematic cross-product evaluation: six enhancement methods (Gaussian smoothing, bilateral filtering, spectral smoothing, FabSoften, GFPGAN, and FabSoften+GFPGAN) are applied to frames from three benchmark datasets, and each enhanced set is scored against six detectors spanning naive, spatial, and frequency approaches. Attack success rate (the proportion of fake frames classified as real) increases monotonic

What carries the argument

The central object is the face enhancement pipeline used as an attack: six transformations M1–M6 ranging from local smoothing and frequency-domain low-pass filtering to wavelet-based fabric softening and GAN-based blind restoration. The mechanism that carries the argument is attack success rate (ASR), the proportion of fake frames reclassified as real after enhancement; enhancements that preserve high structural similarity (SSIM) can still shift predictions because they suppress the forensic cues—texture irregularities, upsampling traces, high-frequency anomalies—that naive, spatial, and frequency detectors respectively key on. The paper's key comparison is ASR against perceptual quality (SS

Load-bearing premise

The load-bearing premise is that the elevated attack success rates are caused by the enhancement transformations; the paper does not tabulate each detector's ASR on unenhanced fake frames, so if those baselines were already high, the anti-forensic effect would not be established.

What would settle it

Run the same six detectors on the same fake frames before and after each enhancement, using identical sampling and thresholds. If unenhanced ASR is already at or near the enhanced ASR, or the per-frame before-after delta is near zero, the central claim of an anti-forensic effect fails.

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

Share X LinkedIn Reddit HN

If this is right

  • Everyday, publicly available enhancement tools can be repurposed as model-agnostic attacks, meaning no detector access or optimization is needed to degrade forensic accuracy.
  • All three detector families—naive, spatial, and frequency—degrade under enhancement, so the weakness is not an artifact of one architecture but of reliance on editable low-level cues.
  • GAN-based enhancement trades structural similarity for evasion; FabSoften+GFPGAN reaches the highest ASR while producing visibly altered faces, so attack strength can be chosen by a perceptual-quality budget.
  • Adversarial training on a single enhancement does not generalize; only training on the strongest GAN enhancement (M6) improves consistency, and even then transfer to other transformations is partial.
  • Social-media pipelines that routinely beautify or restore faces can unintentionally blind forensic systems, creating a realistic, non-adversarial failure mode.

Where Pith is reading between the lines

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

  • I would expect the same enhancement operations to threaten commercial and API-based detectors that are not fine-tuned on enhancement-augmented data, since the attack requires no model knowledge; this is a testable extension the paper does not run.
  • The reported trade-off between SSIM and ASR suggests a practical benchmark: detectors should be evaluated on an enhancement robustness curve, not a single clean-test AUC, to expose how quickly accuracy degrades as allowed distortion increases.
  • Because frequency detectors fall to spectral smoothing and GAN restoration, future defenses should avoid relying on high-frequency cues alone and should consider cross-domain consistency or temporal cues that enhancement pipelines are less likely to preserve.

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

2 major / 5 minor

Summary. The paper investigates whether common face enhancement operations (Gaussian/Bilateral/Spectral smoothing, FabSoften, GFPGAN, and FabSoften+GFPGAN) can act as anti-forensic tools against deepfake detectors. Six detectors from three categories (Naive, Spatial, Frequency) are evaluated on FaceForensics++, DeepFakeDetection, and CelebDF-v2. The paper reports attack success rates (ASR), AUC, perceptual quality metrics, a parameter ablation, and an adversarial-training experiment. The central claim is that face enhancement methods substantially degrade detection accuracy, with GAN-based enhancement being the most effective.

Significance. The question is practically important: if the effect is real, ordinary face-enhancement filters used in consumer apps and social media can inadvertently or deliberately bypass forensic detectors. The study's breadth is a strength: six detectors across three architectural families, three datasets, six enhancement methods, and an ablation linking ASR to perceptual quality. The adversarial-training experiment is a useful first step. However, the quantitative claim currently rests on ASR values that are not compared with a no-enhancement control, so the magnitude of the anti-forensic effect is not yet established from the reported numbers.

major comments (2)
  1. [Section 3.1 / Section 4.1, Table 1, Figure 2] ASR is defined as the proportion of fake samples misclassified as real, which is a threshold-dependent false-negative rate. The paper never reports the corresponding ASR on unenhanced fake frames for the same detectors, datasets, and decision threshold. The grey bars in Figure 2 are baseline AUC, a threshold-independent metric, and cannot serve as this control. Without a no-enhancement ASR baseline, the reported values (e.g., 87.39% for EffNetB4 on DFD in Table 1) cannot be attributed to the enhancement operations; they may partly reflect poor calibration or inappropriate default thresholds of the detectors. Please report baseline ASR at the same operating point used for the attack ASR, and specify how the threshold is selected for each detector.
  2. [Section 4.4, Table 4] The adversarial-training analysis does not include a control condition on unenhanced data. After fine-tuning on M2/M4/M6, the model's AUC on the original, unenhanced FF++ test set is not reported, so one cannot distinguish genuine robustness improvement from catastrophic forgetting of clean-fake detection or simple domain adaptation to the enhancement artifact. The table also reports a single run with no confidence intervals. Please add a clean-test column and a control fine-tuned on unenhanced data, and state the number of runs.
minor comments (5)
  1. [Abstract / Section 4.1] The abstract states 'ASR up to 64.63%' for basic filters and 'up to 75.12%' for GAN-based techniques, but Table 1 contains per-cell values up to 69.92% for FabSoften and 87.39% for FabSoften+GFPGAN. The numbers 74.19/75.12/73.09 in Section 4.1 are FF++-only averages yet Table 1 spans three datasets. Please clarify the aggregation and correct the abstract to match Table 1.
  2. [Section 3.2] M3 is described as using the Discrete Fourier Transform, but reference [1] is the DCT paper (Ahmed et al.). Please clarify which transform is actually used and cite appropriately.
  3. [Figure 2] The grey bars labelled 'baseline AUC' have no numerical values or legend. Please add the exact AUC values or a table, and state whether these are the same detectors/configurations as in Table 1.
  4. [Section 3.1] The evaluation protocol is underspecified: it says 'every fifth frame' and 'standard train-test splits' but does not state whether the detectors were pre-trained on FF++ and tested cross-dataset, or fine-tuned on each dataset, nor how the ASR decision threshold is chosen. Please specify for reproducibility.
  5. [Throughout] Typos: conclusion 'This study highlight' should be 'highlights'; 'reliability ad robustness' should be 'reliability and robustness'.

Circularity Check

0 steps flagged

No significant circularity: all load-bearing results are external measurements against standard benchmarks; the only author-overlap citation is non-load-bearing related work.

full rationale

The paper's central assertion is that face enhancement operations (M1-M6) can degrade deepfake detectors. This is supported by directly measured ASR values in Table 1 and AUC in Figure 2 across FF++, DFD, and CelebDF-v2, using standard detector models (EffNetB4, Xception, CORE, UCF, F3Net, SPSL). The measurements are external to the paper's argument in the sense that no detector parameter or enhancement parameter is defined in terms of the target ASR; ASR is computed after applying fixed, published enhancement algorithms to test frames. There is no derivation chain in which an input quantity equals an output quantity by construction. The only overlapping-author reference is [2] (Ain et al., with coauthor K. Malik), cited in the literature review as prior work on perceptual facial-mole attacks; the present experiments neither invoke nor rely on [2]'s results, so this is normal self-citation, not load-bearing circularity. The paper does not present a prediction from a fitted model; the ablation in Section 4.3 tunes enhancement hyperparameters on Xception/FF++, then transfers those settings to other detectors/datasets. That is a generalization/measurement issue, not circularity. The absence of an explicit no-enhancement ASR control next to Table 1 is a legitimate comparison gap for interpreting whether degradation is attributable to enhancement, but it does not make the reported ASR equal to the paper's input definitions. Therefore no circular step is exhibited.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

All results are empirical; the central claim rests on the representativeness of datasets and detectors, the chosen enhancement parameters, and the assumption that ASR measures enhancement-induced evasion rather than baseline detector weakness. No new entities are introduced.

free parameters (5)
  • M1 Gaussian smoothing kernel = 7x7 kernel, sigma=1.5
    Selected in ablation on Xception FF++ as best ASR/SSIM trade-off; main results in Table 1 use this setting.
  • M2 Bilateral filter parameters = d=17, sigma_color=60, sigma_space=100
    Selected from Table 3 as best trade-off; used for all main results.
  • M3 Spectral smoothing cutoff radius = r=40
    Selected in ablation; smaller radii blur more and raise ASR, larger radii preserve detail but lower ASR.
  • M4 FabSoften alpha parameters = alpha_r=10, alpha_eps=5
    Selected as trade-off; alpha_r=10, alpha_eps=10 gave ASR 65.78 but lower SSIM.
  • M6 combination strategy = FabSoften + GFPGAN
    Chosen from combination ablation as highest overall ASR (76.25 on FF++ Xception) and used for main results.
axioms (4)
  • domain assumption ASR after enhancement is attributable to the enhancement, not to baseline detector failures.
    ASR is reported without a no-enhancement baseline in Table 1; if baseline ASR is already high, the anti-forensic interpretation would be unsupported. Section 4.1, Table 1 and Figure 2.
  • domain assumption The six detectors are representative of deepfake detection categories.
    The paper selects two detectors per category; generalization beyond these models is assumed but not proven. Section 3.3.
  • domain assumption Sampled frames preserve the detection behavior of the videos.
    Every fifth frame is used; temporal detectors and video-level aggregation are not considered. Section 3.1.
  • domain assumption Hand-tuned enhancement parameters are appropriate for all detectors and datasets.
    Parameters selected once on Xception FF++ are reused in all Table 1 cells; other parameter choices may yield different ASR values. Section 4.3.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Realism to Deception: Investigating Deepfake Detectors Against Face Enhancement." pith.science (2026). https://pith.science/paper/HIVVIWNS

@misc{pith2026250907178,
  author       = {Pith},
  title        = {Pith review of: Realism to Deception: Investigating Deepfake Detectors Against Face Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HIVVIWNS}},
  note         = {Machine review of arXiv:2509.07178}
}
Share X LinkedIn Reddit HN
read the original abstract

Face enhancement techniques are widely used to enhance facial appearance. However, they can inadvertently distort biometric features, leading to significant decrease in the accuracy of deepfake detectors. This study hypothesizes that these techniques, while improving perceptual quality, can degrade the performance of deepfake detectors. To investigate this, we systematically evaluate whether commonly used face enhancement methods can serve an anti-forensic role by reducing detection accuracy. We use both traditional image processing methods and advanced GAN-based enhancements to evaluate the robustness of deepfake detectors. We provide a comprehensive analysis of the effectiveness of these enhancement techniques, focusing on their impact on Na\"ive, Spatial, and Frequency-based detection methods. Furthermore, we conduct adversarial training experiments to assess whether exposure to face enhancement transformations improves model robustness. Experiments conducted on the FaceForensics++, DeepFakeDetection, and CelebDF-v2 datasets indicate that even basic enhancement filters can significantly reduce detection accuracy achieving ASR up to 64.63\%. In contrast, GAN-based techniques further exploit these vulnerabilities, achieving ASR up to 75.12\%. Our results demonstrate that face enhancement methods can effectively function as anti-forensic tools, emphasizing the need for more resilient and adaptive forensic methods.

Figures

Figures reproduced from arXiv: 2509.07178 by Ijaz Ul Haq, Khalid Malik, Muhammad Saad Saeed.

Figure 1
Figure 1. Figure 1: Using face enhancement as anti-forensic in deepfake detectors (left).Naïve, Spa [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Deepfake detector AUC across FaceForensics++ (FF++), DeepfakeDetection [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Grad-CAM visualization of detector attention. Ground truth: Deepfake. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of face enhancement. Bottom: impact on prediction shift [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative effects of different parameters on the visual output of face enhancement [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (a) Some severe examples where the GAN-based face enhancement have dras [PITH_FULL_IMAGE:figures/full_fig_p009_6.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. Uncertainty-Aware Deepfake Detection via Multi-View Structural Learning

    cs.CV 2026-07 conditional novelty 6.0

    A three-stream detector with disagreement-calibrated evidential uncertainty cuts calibration error roughly in half on four of five cross-dataset deepfake benchmarks.

Reference graph

Works this paper leans on

26 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Discrete cosine transform.IEEE transactions on Computers, 100:90–93, 2006

    Nasir Ahmed, T_ Natarajan, and Kamisetty R Rao. Discrete cosine transform.IEEE transactions on Computers, 100:90–93, 2006

    work page 2006

  2. [2]

    Exposing the limits of deepfake detection using novel facial mole attack: A perceptual black-box adversar- ial attack study

    Qurat Ul Ain, Ali Javed, Khalid Mahmood Malik, and Aun Irtaza. Exposing the limits of deepfake detection using novel facial mole attack: A perceptual black-box adversar- ial attack study. In2024 IEEE International Conference on Image Processing (ICIP), pages 3820–3826. IEEE, 2024

    work page 2024

  3. [3]

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

    Nicholas Carlini and Hany Farid. Evading deepfake-image detectors with white-and black-box attacks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 658–659, 2020

    work page 2020

  4. [4]

    Towards evaluating the robustness of neural net- works

    Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural net- works. In2017 ieee symposium on security and privacy (sp), pages 39–57. Ieee, 2017

    2017

  5. [5]

    Restricted black-box adversarial attack against deepfake face swapping.IEEE Transactions on Information F orensics and Security, 18:2596–2608, 2023

    Junhao Dong, Yuan Wang, Jianhuang Lai, and Xiaohua Xie. Restricted black-box adversarial attack against deepfake face swapping.IEEE Transactions on Information F orensics and Security, 18:2596–2608, 2023

    work page 2023

  6. [6]

    An adversarial attack approach for explainable ai evaluation on deepfake detection models.Computers & Security, 139: 103684, 2024

    Balachandar Gowrisankar and Vrizlynn LL Thing. An adversarial attack approach for explainable ai evaluation on deepfake detection models.Computers & Security, 139: 103684, 2024

    work page 2024

  7. [7]

    Evad- ing deepfake detectors via adversarial statistical consistency

    Yang Hou, Qing Guo, Yihao Huang, Xiaofei Xie, Lei Ma, and Jianjun Zhao. Evad- ing deepfake detectors via adversarial statistical consistency. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12271– 12280, 2023

    work page 2023

  8. [8]

    Fakepolisher: Making deepfakes more detection-evasive by shallow reconstruction

    Yihao Huang, Felix Juefei-Xu, Run Wang, Qing Guo, Lei Ma, Xiaofei Xie, Jianwen Li, Weikai Miao, Yang Liu, and Geguang Pu. Fakepolisher: Making deepfakes more detection-evasive by shallow reconstruction. InProceedings of the 28th ACM interna- tional conference on multimedia, pages 1217–1226, 2020

    work page 2020

  9. [9]

    Adversarial deepfakes: Evaluating vulnerability of deepfake detectors to adversarial examples

    Shehzeen Hussain, Paarth Neekhara, Malhar Jere, Farinaz Koushanfar, and Julian McAuley. Adversarial deepfakes: Evaluating vulnerability of deepfake detectors to adversarial examples. InProceedings of the IEEE/CVF winter conference on applica- tions of computer vision, pages 3348–3357, 2021

    work page 2021

  10. [10]

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

    Marija Ivanovska and Vitomir Struc. On the vulnerability of deepfake detectors to attacks generated by denoising diffusion models. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1051–1060, 2024

    work page 2024

  11. [11]

    Deepfake: a social construction of technology perspective.Current Issues in Tourism, 24(13):1798–1802, 2021

    Andrei OJ Kwok and Sharon GM Koh. Deepfake: a social construction of technology perspective.Current Issues in Tourism, 24(13):1798–1802, 2021

    work page 2021

  12. [12]

    Frequency domain regular- ization for iterative adversarial attacks.Pattern Recognition, 134:109075, 2023

    Tengjiao Li, Maosen Li, Yanhua Yang, and Cheng Deng. Frequency domain regular- ization for iterative adversarial attacks.Pattern Recognition, 134:109075, 2023

    work page 2023

  13. [13]

    Generalizing face forgery de- tection with high-frequency features

    Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery de- tection with high-frequency features. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16317–16326, 2021. SAEED, HAQ, MALIK: REALISM TO DECEPTION11

    work page 2021

  14. [14]

    Ava: Incon- spicuous attribute variation-based adversarial attack bypassing deepfake detection

    Xiangtao Meng, Li Wang, Shanqing Guo, Lei Ju, and Qingchuan Zhao. Ava: Incon- spicuous attribute variation-based adversarial attack bypassing deepfake detection. In 2024 IEEE Symposium on Security and Privacy (SP), pages 74–90. IEEE, 2024

    work page 2024

  15. [15]

    Core: Consistent representation learning for face forgery detection

    Yunsheng Ni, Depu Meng, Changqian Yu, Chengbin Quan, Dongchun Ren, and You- jian Zhao. Core: Consistent representation learning for face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12–21, 2022

    work page 2022

  16. [16]

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

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in fre- quency: Face forgery detection by mining frequency-aware clues. InEuropean confer- ence on computer vision, pages 86–103. Springer, 2020

    work page 2020

  17. [17]

    Faceforensics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. InProceedings of the IEEE/CVF international conference on computer vision, pages 1–11, 2019

    2019

  18. [18]

    Deep person generation: A survey from the perspective of face, pose, and cloth synthesis.ACM Computing Surveys, 55(12):1–37, 2023

    Tong Sha, Wei Zhang, Tong Shen, Zhoujun Li, and Tao Mei. Deep person generation: A survey from the perspective of face, pose, and cloth synthesis.ACM Computing Surveys, 55(12):1–37, 2023

    work page 2023

  19. [19]

    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. PMLR, 2019

    2019

  20. [20]

    Curve and surface smoothing without shrinkage

    Gabriel Taubin. Curve and surface smoothing without shrinkage. InProceedings of IEEE international conference on computer vision, pages 852–857. IEEE, 1995

    work page 1995

  21. [21]

    Bilateral filtering for gray and color images

    Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In Sixth international conference on computer vision (IEEE Cat. No. 98CH36271), pages 839–846. IEEE, 1998

    work page 1998

  22. [22]

    A robust open- set multi-instance learning for defending adversarial attacks in digital image.IEEE Transactions on Information F orensics and Security, 2023

    Kutub Uddin, Yoonmo Yang, Tae Hyun Jeong, and Byung Tae Oh. A robust open- set multi-instance learning for defending adversarial attacks in digital image.IEEE Transactions on Information F orensics and Security, 2023

    work page 2023

  23. [23]

    Deep learning-based counter anti- forensic of gan-based attack in hevc compressed domain using coding pattern analysis

    Kutub Uddin, Yoonmo Yang, and Byung Tae Oh. Deep learning-based counter anti- forensic of gan-based attack in hevc compressed domain using coding pattern analysis. Expert Systems with Applications, 233:120912, 2023

    work page 2023

  24. [24]

    Fabsoften: Face beautification via dynamic skin smoothing, guided feathering, and texture restoration

    Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. Fabsoften: Face beautification via dynamic skin smoothing, guided feathering, and texture restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition Workshops, pages 530–531, 2020

    work page 2020

  25. [25]

    Towards real-world blind face restoration with generative facial prior

    Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9168–9178, 2021

    2021

  26. [26]

    Ucf: Uncovering common features for generalizable deepfake detection

    Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 22412–22423, 2023

    work page 2023

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.