Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Beyond Spatial Frequency: Pixel-wise Temporal Frequency-based Deepfake Video Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A pixel-wise temporal frequency axis, not stacked spatial spectra, is what makes deepfake video detection generalize across unseen datasets.

desk verdict Per-pixel temporal frequency is a genuinely new and well-ablated feature for deepfake detection, but the headline SOTA claim rests on baseline numbers from heterogeneous protocols and the margins are thin. read the letter →

arxiv 2507.02398 v2 pith:DBEYQ62J submitted 2025-07-03 cs.CV cs.AI

Taehoon Kim , Jongwook Choi , Yonghyun Jeong , Haeun Noh , Jaejun Yoo , Seungryul Baek , Jongwon Choi This is my paper
classification cs.CVcs.AI
keywords deepfakevideodetectiontemporalfrequencypixel-wiseFouriertransformmedianfilterattentionproposalmodulecross-datasetgeneralizationfaceforgerytransformer
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

This paper argues that deepfake videos can be detected by looking at the time axis of each pixel: applying a 1D Fourier transform to each pixel's intensity over 32 frames, after subtracting a median-filtered version of each frame, exposes localized flicker-like artifacts that spatial-frequency and RGB-based detectors miss. The authors build a detector that locates the facial regions with the strongest temporal artifacts and combines those with spatio-temporal context, training only on FaceForensics++. They report an average video-level AUC of 92.2% across five unseen datasets, outperforming the strongest baseline they compare with, and improved generalization across unseen synthesis methods and a racially distinct dataset. If this holds, it would give deepfake video detection a signal that transfers better than existing appearance-based and spatial-frequency methods.

What carries the argument

The load-bearing object is the pixel-wise temporal frequency spectrum. From a 32-frame clip, each frame is grayed and has its median-filtered version subtracted, then a 1D FFT along the time axis per pixel yields magnitude values $F_{x,y}$. These frequency maps feed a 2D ResNet to produce a global frequency feature, while an Attention Proposal Module (APM) with differentiable rectangular masks regresses five part coordinates to crop artifact-rich patches; a feature blender merges part and global frequency features into a 3D CNN stream, and two transformer encoders, one spatial and one temporal, produce the final real/fake prediction. The median filtering is crucial: without it the temporal frequency features perform near chance in the supplementary cross-synthesis ablation (62.25% vs. 92.34% average AUC).

What would settle it

Remove the median-filter preprocessing from the pipeline, retrain on FF++, and test on CDF and DFDC: the paper's supplementary Table A predicts a collapse from about 92% average cross-synthesis AUC to 62%, so a small drop would falsify the claim that the median-filtered temporal frequency does the work. A second check is to take fake videos and smooth their temporal axis with temporal averaging or frame interpolation before evaluation; if detection accuracy stays high, the detector is reading static spatial cues rather than temporal frequency.

Watch

Extended reading notes

Core claim

The central discovery is that temporal inconsistency in deepfake videos lives in pixel-wise temporal frequency spectra, not in the stacking of spatial frequency spectra. For each pixel, the 1D Fourier magnitude of the median-filtered gray intensity trace is a strong, localizable forgery clue; the magnitude, not the phase, carries the signal. The paper shows this with a simple ResNet trained on the temporal spectrum alone outperforming image and spatial-frequency detectors in cross-synthesis tests, and with the full model outperforming baselines across CDF, DFDC, FSh, DFo, DFD, and KoDF. The authors would state the finding as: pixel-wise temporal frequency is the missing axis for generalized deepfake video detection.

Load-bearing premise

The load-bearing premise is that subtracting a median-filtered gray frame before the temporal Fourier transform exposes the same kind of local flicker in unseen synthesis methods, datasets, and compression levels; the paper's own ablation shows that without this preprocessing the temporal frequency features drop to near chance, and the filter's kernel size is not analyzed in the main text.

Editorial extensions

If this is right

  • If the central claim holds, deepfake video detectors should stop representing temporal information as a stack of per-frame spatial spectra and instead analyze the pixel-wise time axis directly.
  • Training on FF++ alone should transfer to unseen datasets and synthesis methods, with the reported 92.2% average video-level AUC across CDF, DFDC, FSh, DFo, and DFD, and improved robustness to saturation, blur, and resizing.
  • Spatial-frequency-only methods should continue to lag on temporal artifacts, and detectors using temporal frequency should focus on localized facial regions rather than full frames.
  • The signal is genuinely temporal: randomly shuffling the 32 input frames drops the method's performance by 43.8 percentage points on FF++, confirming that static appearance is not doing the work.
  • Severe video and image compression suppresses the high-frequency part of the temporal spectrum, and the paper identifies this as a real limitation for heavy H.264, JPEG, and WebP compression.

Reading between the lines

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

  • An implicit extension is that the same pixel-wise temporal frequency recipe could be adapted to other video forensics tasks, such as detecting GAN-generated video, screen-capture re-encoding, or temporal splicing, wherever local flicker is the cue.
  • The median filter likely isolates high temporal frequencies by removing a spatially smoothed static component; if so, its kernel size should interact with video resolution and compression level, and varying it across datasets would be a cheap robustness test the paper does not run.
  • Combining temporal-frequency cues with spatial-frequency cues could improve generalization further, since the two signals appear complementary: spatial frequency catches per-frame texture artifacts, while temporal frequency catches local flicker.
  • A testable extension the paper does not run is to temporally smooth fake videos before evaluation, e.g. with temporal averaging or frame interpolation; the resulting AUC drop would quantify how much of the signal is genuinely temporal-frequency-based.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a deepfake video detection method based on pixel-wise temporal frequency. Each pixel's intensity trace over T=32 frames is preprocessed by subtracting a median-filtered version of each frame (Eq. 1), converted to a 1D Fourier magnitude spectrum along the time axis, and processed by a shared 2D ResNet. An Attention Proposal Module (APM) crops five part-level frequency patches, and a joint transformer module (STE/TTE) integrates global and part-level frequency features with spatio-temporal features from a frozen pre-trained FTCN 3D ResNet. The model is trained on FaceForensics++ and evaluated on unseen datasets (CDF, DFDC, FSh, DFo, DFD, KoDF), on cross-synthesis splits, and under perturbations; the paper reports 92.2% average video-level AUC versus 91.1% for AltFreezing in Table 2, and includes ablations showing that each component contributes.

Significance. If the comparative results hold, the paper makes a useful contribution: pixel-wise temporal frequency is a genuinely different artifact representation from stacked spatial frequency, the controlled preliminary experiment (Table 1) and the ablations (Tables 7–9) support the proposed mechanism, and the evaluation spans a wider set of unseen datasets than is typical. The supplementary frame-shuffle test (Table B) and APM heatmaps are useful falsifiable checks, and the cross-synthesis comparisons in Tables 3–5 are reproduced under the paper's own pipeline, which is a strength. However, the headline state-of-the-art claim rests on Table 2, whose baseline numbers are drawn from heterogeneous prior protocols with no error bars; the 1.1-point average margin over AltFreezing is too small to support that claim without a unified re-run. The centrality of the median-filter preprocessing is also demonstrated by the collapse from 92.34% to 62.25% in Table A, but the filter is underspecified.

major comments (4)
  1. [Table 2, §5.1] The central claim of state-of-the-art cross-dataset generalization is not established by Table 2 as presented. The caption states that results for other methods were obtained from [7, 19] and only a few entries are marked as reproduced, while the paper's own pipeline uses RetinaFace+SORT tracking, a fixed 32-frame window, and its own video-level aggregation. Because the source papers use different face detection/tracking, frame sampling, and aggregation protocols, mixing those numbers can shift video-level AUC by more than the reported 1.1-point margin over AltFreezing (91.1 vs 92.2) and 1.4 points over StyleFlow (90.8). On DFDC and FSh, RealForensics actually reports 75.9 and 99.7, above Ours at 75.2 and 99.3. To support the headline claim, the authors should re-run at least the principal baselines (AltFreezing, FTCN, RealForensics, StyleFlow) under the exact evaluation pipeline used for Ours, or restrict the SOTA claim to the reproduced comparisons in Tables 3–5.
  2. [Eq. (1), §4.1.1; Table A (supp.)] The median-filter preprocessing in Eq. (1) is load-bearing rather than cosmetic: Table A shows that without it the global temporal-frequency model collapses from 92.34% to 62.25% average AUC in cross-synthesis. However, the kernel size and shape of Median(·) are never specified in the main text or the supplementary material, and no sensitivity analysis is given with respect to kernel size. Since the filter determines what 'dominant components' are removed before the 1D FFT, and since the final method inherits this preprocessing, the paper should state the kernel parameters and report how performance varies with them (e.g., kernel size 3/5/7 and median vs mean).
  3. [Fig. C (supp.) and §F.3] There is a discrepancy between the frame-interval analysis and the method implementation. Supplementary Fig. C reports that performance increases with frame interval and saturates at interval 8 for most synthesis types, while §F.3 states that 'these consecutive 32 frames' are fed to the model (interval 1) and §4.1.1 sets T=32 with no interval. The main text and supplementary should state explicitly which interval is used for all reported results, and, if interval 8 is better, explain why the default setting uses consecutive frames. This is needed for exact reproduction of Tables 2–9.
  4. [Tables 2–9] No error bars, confidence intervals, or multiple-seed results are reported anywhere in the paper. Several comparisons are within 0.6–1.1 AUC points of the strongest baseline (e.g., KoDF 91.3 vs 90.7 for StyleFlow in Table 3; DFDC 75.2 vs 74.7 for AltFreezing in Table 2), which is likely within run-to-run variation for training from scratch. The paper should report mean±std over at least three seeds for its own method and for the reproduced baselines, and should state whether the margins in Table 2 are stable across seeds.
minor comments (5)
  1. [§4.2.2] 'patial analysis' should be 'spatial analysis'.
  2. [§5] 'NeuralTexutres' and 'Forenscis' are typos; they should be 'NeuralTextures' and 'Forensics'.
  3. [Table 6, §5.4] The claim of 'comparable computational complexity' is overstated for latency: Ours is 50.05 ms per clip vs 14.33 ms for FTCN, about 3.5× slower, even though FLOPs are comparable; rephrase to describe the accuracy-latency trade-off.
  4. [§F.2] The description of which official weights were used and which models were retrained is helpful, but the paper should also state whether the code and trained models will be released; no code link is given.
  5. [Table 1 caption] The caption says 'Above the double line is the result of the current method, while below is the result of training only with ResNet-18...', but the rows above the double line are existing methods, not 'the current method'; clarify the layout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pixel-wise temporal-frequency method is trained only on FF++ and benchmarked on unseen datasets; baseline provenance and reuse of FTCN weights are comparability concerns, not circular derivation.

full rationale

The paper's derivation chain is self-contained. The core representation is defined in Eqs. (1)-(2): median-filtered frames are Fourier transformed along the temporal axis per pixel, and the resulting magnitude spectrum is classified. This is not defined in terms of the target output, and the claim that such spectra generalize is tested against external datasets (CDF, DFDC, FSh, DFo, DFD, KoDF) after training only on FF++, so the benchmark results are not fitted or constructed from the method's own outputs. The preliminary cross-synthesis study (Table 1) and component ablations (Tables 7-9 and supplementary Tables A-E) compare fixed architectural variants under the same training protocol rather than re-introducing a fitted quantity as a prediction. The median-filter preprocessing is an empirically selected choice (supplementary Table A), not a fitted parameter later called a prediction; the reported sensitivity to it is an honest ablation. Reusing FTCN's pre-trained TTE/3D ResNet and comparing against the authors' own StyleFlow are self-citations and weight reuse, but they do not supply the load-bearing evidence for the temporal-frequency contribution: the method's gain is measured on unseen-domain external data, and the central contribution does not depend on a self-cited result being true. The Table 2 note that baseline numbers come from [7,19] raises a comparability/error-bar concern about the SOTA margin, but that is a correctness risk, not circularity. No uniqueness theorem, ansatz-smuggling citation, or renaming of a known result is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper is an empirical deep-learning paper with no derivation from first principles. It relies on standard pretrained backbones and several hand-set hyperparameters. The most consequential assumption is that median-filtered pixel-wise temporal Fourier magnitude generalizes across unseen domains; the supplementary shows performance drops dramatically without this filter.

free parameters (4)
  • theta (patch half-size) = 44
    APM rectangular region half-size, empirically set in Section 4.1.2. It controls how much local context each part sees.
  • T (clip length in frames) = 32
    Each clip contains 32 frames for the temporal Fourier transform, as stated in Section 4.1.1.
  • logistic scale factor = 10
    Scale of the logistic function in Eq. 5, used to make APM masks differentiable.
  • number of APM parts = 5
    Number of cropped part regions; ablations in Table D show 5 is the best of {0, 1, 3, 5, 7}.
assumptions (4)
  • domain assumption FF++ training data is representative enough that features learned on it transfer to unseen datasets such as CDF, DFDC, FSh, DFo, DFD, and KoDF.
    The central evaluation protocol trains only on FF++ and evaluates on unseen datasets; if FF++ artifacts were not representative, the cross-dataset claim would fail.
  • ad hoc to paper Median-filtered pixel-wise temporal Fourier magnitude captures the relevant artifacts; without this filter, detection collapses (Table A: 62.25% vs 92.34% average).
    The preprocessing in Eq. 1 is introduced by the paper and is load-bearing for performance, but its kernel size is not reported in the main text.
  • domain assumption APM converges to artifact-rich regions from classification gradients alone, without localization labels.
    Section 4.1.2 and Supplementary B describe gradient-based coordinate updates; no theoretical guarantee is provided, and ablation shows APM matters (Table C).
  • domain assumption Pretrained 3D ResNet and TTE from FTCN provide a transferable spatio-temporal backbone.
    The method reuses FTCN's pre-trained components, as stated in Section 5, and freezes the 3D CNN in the feature blender, so the method's success partly inherits FTCN's transferability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Spatial Frequency: Pixel-wise Temporal Frequency-based Deepfake Video Detection." pith.science (2026). https://pith.science/paper/DBEYQ62J

@misc{pith2026250702398,
  author       = {Pith},
  title        = {Pith review of: Beyond Spatial Frequency: Pixel-wise Temporal Frequency-based Deepfake Video Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DBEYQ62J}},
  note         = {Machine review of arXiv:2507.02398}
}
read the original abstract

We introduce a deepfake video detection approach that exploits pixel-wise temporal inconsistencies, which traditional spatial frequency-based detectors often overlook. Traditional detectors represent temporal information merely by stacking spatial frequency spectra across frames, resulting in the failure to detect temporal artifacts in the pixel plane. Our approach performs a 1D Fourier transform on the time axis for each pixel, extracting features highly sensitive to temporal inconsistencies, especially in areas prone to unnatural movements. To precisely locate regions containing the temporal artifacts, we introduce an attention proposal module trained in an end-to-end manner. Additionally, our joint transformer module effectively integrates pixel-wise temporal frequency features with spatio-temporal context features, expanding the range of detectable forgery artifacts. Our framework represents a significant advancement in deepfake video detection, providing robust performance across diverse and challenging detection scenarios.

Figures

Figures reproduced from arXiv: 2507.02398 by the authors.

Figure 1
Figure 1. Comparison with state-of-the-art methods. Our approach leverages pixel-wise temporal frequency, which was not utilized in previous work. Ours outperforms the state-of-the-art methods across various unseen datasets, such as CDF [31]. For comparison, we trained only on FF++ [40] and evaluated video-level AUC for unseen datasets. contrast, frequency-based detectors transform pixel values into frequency-aware features, … view at source ↗
Figure 2
Figure 2. Visualization of temporal inconsistency and a mechanism of our pixel-wise temporal frequency. In (a), we depict the motion of vertical slices along the time axis in the real and fake video, as suggested by the visualization in [47]. (b) presents the difference in frequency spectrum extraction between previous frequency-based methods and our pixel-wise temporal frequency method. While previous research has primarily … view at source ↗
Figure 3
Figure 3. The pipeline of our method. For extracting temporal frequency F 0 , the video clip V is decomposed into temporal frequency components using the Fourier Transform. The frequency feature extractor obtains a part-based frequency feature Z p and a global frequency feature Z 0 using 2D ResNet and an attention proposal module. The part-based and global frequency features enter the feature blender to get a blended feature … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Evaluation of the robustness against perturbation. We compare performance across five degradation levels for four perturbation scenarios. The vertical axis shows video-level AUC (%), and the horizontal axis represents perturbation intensity, with higher values indicati…
Figure 5
Figure 5. Figure 5: Visualization of part proposed by APM on extended frame. APM propose where the temporal incoherence occurred. we demonstrate the generalization performance of our method by showing the strong performance for all unseen synthesis methods. Our method shows remarkable eff…
Figure 6
Figure 6. Figure 6: Visualization of extended frames activation map in 3D CNN. We compare activation maps from the last 3D CNN layer of FTCN (first row) and our feature blender (second row) when processing a fake clip. Frequency Feature Extractor Testing Set Part Proposal Global Part CDF …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Detecting AI-Generated Video: A Vision-Language Dual-View Survey

    cs.CV 2026-07 conditional novelty 6.0 of 10

    AIGC-V detection should be treated as factual fidelity verification and organized by a four-layer vision-language dual-view taxonomy spanning cues, motion, cross-modal consistency, and world-level reasoning.

Reference graph

Works this paper leans on

56 extracted references · 51 canonical work pages · cited by 1 Pith paper

  1. [1]

    Mesonet: a compact facial video forgery detection network

    Darius Afchar, Vincent Nozick, Junichi Y amagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1–7. IEEE, 2018. 2

    work page 2018

  2. [2]

    Simple online and realtime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE International Conference on Image Processing (ICIP), pages 3464–3468, 2016. 5

    work page 2016

  3. [3]

    Quo vadis, action recogni- tion? a new model and the kinetics dataset

    Joao Carreira and Andrew Zisserman. Quo vadis, action recogni- tion? a new model and the kinetics dataset. Inproceedings of the IEEE Conference on Computer V ision and P attern Recognition, pages 6299–6308, 2017. 4

    work page 2017

  4. [4]

    What makes fake images detectable? understanding properties that generalize

    Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. InEuropean Conference on Computer V ision, 2020. 6

    work page 2020

  5. [5]

    Self-supervised learning of adversarial example: T owards good generalizations for deepfake detection

    Liang Chen, Y ong Zhang, Yibing Song, Lingqiao Liu, and Jue W ang. Self-supervised learning of adversarial example: T owards good generalizations for deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18710–18719, 2022. 2

    work page 2022

  6. [6]

    Local relation learning for face forgery de- tection

    Shen Chen, T aiping Y ao, Y ang Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Local relation learning for face forgery de- tection. In Proceedings of the AAAI conference on artificial intelligence, pages 1081–1088, 2021. 1

    work page 2021

  7. [7]

    Exploiting style latent flows for general- izing deepfake video detection

    Jongwook Choi, T aehoon Kim, Y onghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for general- izing deepfake video detection. InProceedings of the IEEE/CVF Conference on Computer V ision and P attern Recognition (CVPR), pages 1133–1143, 2024. 2, 6, 5

    work page 2024

  8. [8]

    Xception: Deep learning with depthwise sep- arable convolutions

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

    work page 2017

Show all 56 references
  1. [9]

    Empirical evaluation of gated recurrent neural networks on sequence modeling

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Y oshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2

    2014 arXiv

  2. [10]

    Lip reading in the wild

    Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Computer V ision–ACCV 2016: 13th Asian Conference on Computer V ision, T aipei, T aiwan, November 20-24, 2016, Revised Selected P apers, P art II 13, pages 87–103. Springer,

    2016

  3. [11]

    Retinaface: Single-shot multi-level face local- isation in the wild

    Jiankang Deng, Jia Guo, Evangelos V erveras, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face local- isation in the wild. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203–5212,

  4. [12]

    The deepfake detection challenge (dfdc) dataset

    Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin W ang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 5, 4

    2006 arXiv

  5. [13]

    Implicit identity leakage: The stumbling block 5 to improving deepfake detection generalization

    Shichao Dong, Jin W ang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stumbling block 5 to improving deepfake detection generalization. InProceedings of the IEEE/CVF Conference on Computer V ision and P attern Recognition, pages 3994–4004,...

    2023

  6. [14]

    Contributing Data to Deep- fake Detection Research — ai.googleblog.com.https://ai

    Nick Dufour and Andrew Gully. Contributing Data to Deep- fake Detection Research — ai.googleblog.com.https://ai. googleblog.com/2019/09/contributing- data- to-deepfake-detection.html, 2019. [Accessed 30- 07-2023]. 5, 4

    2019

  7. [15]

    Slowfast networks for video recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202–6211, 2019. 4

    2019

  8. [16]

    Look closer to see better: Recurrent attention convolutional neural network for fine- grained image recognition

    Jianlong Fu, Heliang Zheng, and T ao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine- grained image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4438–4446,

  9. [17]

    Exploiting fine-grained face forgery clues via pro- gressive enhancement learning

    Qiqi Gu, Shen Chen, T aiping Y ao, Y ang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via pro- gressive enhancement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 735–743, 2022. 2, 6

    2022

  10. [18]

    Lips don’t lie: A generalisable and robust ap- proach to face forgery detection

    Alexandros Haliassos, Konstantinos V ougioukas, Stavros Petridis, and Maja Pantic. Lips don’t lie: A generalisable and robust ap- proach to face forgery detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5039–5049, 2021. 2, 5, 6, 4

    2021

  11. [19]

    Leveraging real talking faces via self-supervision for robust forgery detection

    Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. InProceedings of the IEEE/CVF Con- ference on Computer V ision and P attern Recognition (CVPR), pages 14950–14962, 2022. 2, 6, 5

    2022

  12. [20]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 3, 5, 2

    2016

  13. [21]

    Learning to detect fake face images in the wild

    Chih-Chung Hsu, Chia-Y en Lee, and Yi-Xiu Zhuang. Learning to detect fake face images in the wild. In 2018 international symposium on computer , consumer and control (IS3C), pages 388–391. IEEE, 2018. 2

    2018

  14. [22]

    Bihpf: Bilateral high- pass filters for robust deepfake detection

    Y onghyun Jeong, Doyeon Kim, Seungjai Min, Seongho Joe, Y oungjune Gwon, and Jongwon Choi. Bihpf: Bilateral high- pass filters for robust deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer V ision, pages 48–57, 2022. 2

    2022

  15. [23]

    Frepgan: robust deepfake detection using frequency-level perturbations

    Y onghyun Jeong, Doyeon Kim, Y oungmin Ro, and Jongwon Choi. Frepgan: robust deepfake detection using frequency-level perturbations. InProceedings of the AAAI Conference on Artifi- cial Intelligence, pages 1060–1068, 2022. 1, 2

    2022

  16. [24]

    DeeperForensics-1.0: A large-scale dataset for real-world face forgery detection

    Liming Jiang, Ren Li, W ayne Wu, Chen Qian, and Chen Change Loy. DeeperForensics-1.0: A large-scale dataset for real-world face forgery detection. InCVPR, 2020. 5, 7, 4

    2020

  17. [25]

    A style-based gen- erator architecture for generative adversarial networks.CoRR, abs/1812.04948, 2018

    Tero Karras, Samuli Laine, and Timo Aila. A style-based gen- erator architecture for generative adversarial networks.CoRR, abs/1812.04948, 2018. 5

    2018 arXiv

  18. [26]

    Image feature detectors for deepfake video detection

    Faten F Kharbat, T arik Elamsy, Ahmed Mahmoud, and Rami Abdullah. Image feature detectors for deepfake video detection. In 2019 IEEE/ACS 16th International Conference on Computer Systems and Applications (AICCSA), pages 1–4. IEEE, 2019. 2

    2019

  19. [27]

    Kodf: A large-scale korean deepfake de- tection dataset

    Patrick Kwon, Jaeseong Y ou, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. Kodf: A large-scale korean deepfake de- tection dataset. InProceedings of the IEEE/CVF International Conference on Computer V ision (ICCV), pages 10744–10753,

  20. [28]

    Faceshifter: T owards high fidelity and occlusion aware face swapping

    Lingzhi Li, Jianmin Bao, Hao Y ang, Dong Chen, and Fang W en. Faceshifter: T owards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 5, 4

    1912 arXiv

  21. [29]

    Face x-ray for more general face forgery detection

    Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Y ang, Dong Chen, Fang W en, and Baining Guo. Face x-ray for more general face forgery detection. InProceedings of the IEEE/CVF Conference on Computer V ision and P attern Recognition (CVPR), 2020. 1, 2, 6

    2020

  22. [30]

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

    Y uezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International workshop on information forensics and security (WIFS), pages 1–7. IEEE, 2018. 2

    2018

  23. [31]

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

    Y uezun Li, Xin Y ang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3207–3216, 2020. 1, 5, 4

    2020

  24. [32]

    Gener- alizing face forgery detection with high-frequency features

    Y uchen Luo, Y ong Zhang, Junchi Y an, and W ei Liu. Gener- alizing face forgery detection with high-frequency features. In Proceedings of the IEEE/CVF Conference on Computer V ision and P attern Recognition (CVPR), pages 16317–16326, 2021. 2, 3, 6, 5

    2021

  25. [33]

    T wo-branch recurrent network for isolating deepfakes in videos

    Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Gurudatt, and W ael AbdAlmageed. T wo-branch recurrent network for isolating deepfakes in videos. InComputer V ision–ECCV 2020: 16th European Conference, Glasgow , UK, August 23–28, 2020, Proceedings, P ar...

    2020

  26. [34]

    Deepfakes generation and detection: State-of-the-art, open challenges, countermeasures, and way forward

    Momina Masood, Mariam Nawaz, Khalid Mahmood Malik, Ali Javed, Aun Irtaza, and Hafiz Malik. Deepfakes generation and detection: State-of-the-art, open challenges, countermeasures, and way forward. Applied intelligence, 53(4):3974–4026, 2023. 1

    2023

  27. [35]

    F 2 trans: High-frequency fine-grained transformer for face forgery detection

    Changtao Miao, Zichang T an, Qi Chu, Huan Liu, Honggang Hu, and Nenghai Y u. F 2 trans: High-frequency fine-grained transformer for face forgery detection. IEEE Transactions on Information F orensics and Security, 18:1039–1051, 2023. 2

    2023

  28. [36]

    Generaliz- able deepfake detection with phase-based motion analysis.IEEE Transactions on Image Processing, 2024

    Ekta Prashnani, Michael Goebel, and BS Manjunath. Generaliz- able deepfake detection with phase-based motion analysis.IEEE Transactions on Image Processing, 2024. 2

    2024

  29. [37]

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

    Y uyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86–103. Springer, 2020. 2, 6

    2020

  30. [38]

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

    Y uyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86–103. Springer, 2020. 2

    2020

  31. [39]

    Encoding in style: a stylegan encoder for image-to-image translation

    Elad Richardson, Y uval Alaluf, Or Patashnik, Y otam Nitzan, Y aniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer V ision and P attern Recogni- tion (CVPR), 2021. 5 6

    2021

  32. [40]

    Faceforensics: A large-scale video dataset for forgery detection in human faces

    Andreas R¨ossler, Davide Cozzolino, Luisa V erdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1, 3, 5, 6, 2, 4

    2018 arXiv

  33. [41]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge.Interna- tional Journal of Computer V ision (I...

    2015

  34. [42]

    Recurrent convolutional strate- gies for face manipulation detection in videos.Interfaces (GUI), 3(1):80–87, 2019

    Ekraam Sabir, Jiaxin Cheng, A yush Jaiswal, W ael AbdAlmageed, Iacopo Masi, and Prem Natarajan. Recurrent convolutional strate- gies for face manipulation detection in videos.Interfaces (GUI), 3(1):80–87, 2019. 1, 2, 3, 6

    2019

  35. [43]

    Detecting deepfakes with self-blended images

    Kaede Shiohara and T oshihiko Y amasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer V ision and P attern Recognition, pages 18720–18729, 2022. 2

    2022

  36. [44]

    Face forgery detection via symmet- ric transformer

    Luchuan Song, Xiaodan Li, Zheng Fang, Zhenchao Jin, Y ueFeng Chen, and Chenliang Xu. Face forgery detection via symmet- ric transformer. InProceedings of the 30th ACM International Conference on Multimedia, pages 4102–4111, 2022. 2

    2022

  37. [45]

    Improving the efficiency and robustness of deepfakes detec- tion through precise geometric features

    Zekun Sun, Y ujie Han, Zeyu Hua, Na Ruan, and W eijia Jia. Improving the efficiency and robustness of deepfakes detec- tion through precise geometric features. InProceedings of the IEEE/CVF Conference on Computer V ision and P attern Recogni- tion, pages 3609–3618, 2021. 2

    2021

  38. [46]

    Cnn-generated images are surprisingly easy to spot

    Sheng-Y u W ang, Oliver W ang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695–8704,

  39. [47]

    Spatial-temporal frequency forgery clue for video forgery detection in vis and nir scenario.IEEE Transactions on Circuits and Systems for V ideo T echnology, 2023

    Y ukai W ang, Chunlei Peng, Decheng Liu, Nannan W ang, and Xinbo Gao. Spatial-temporal frequency forgery clue for video forgery detection in vis and nir scenario.IEEE Transactions on Circuits and Systems for V ideo T echnology, 2023. 1, 2, 6

    2023

  40. [48]

    Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection

    Y uan W ang, Kun Y u, Chen Chen, Xiyuan Hu, and Silong Peng. Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection. InProceedings of the IEEE/CVF Conference on Computer V ision and P attern Recogni- tion, pages 7278–7287, 2023. 1, 2, 6

    2023

  41. [49]

    Altfreezing for more general video face forgery detection

    Zhendong W ang, Jianmin Bao, W engang Zhou, W eilun W ang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF Conference on Com- puter V ision and P attern Recognition, pages 4129–4138, 2023. 2, 6, 5

    2023

  42. [50]

    Multimodal graph learning for deepfake detection, 2023

    Zhiyuan Y an, Peng Sun, Y ubo Lang, Shuo Du, Shanzhuo Zhang, W ei W ang, and Lei Liu. Multimodal graph learning for deepfake detection, 2023. 2, 6

    2023

  43. [51]

    Exposing deep fakes using inconsistent head poses

    Xin Y ang, Y uezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261–8265. IEEE, 2019. 2

    2019

  44. [52]

    Attributing fake images to gans: Learning and analyzing gan fingerprints

    Ning Y u, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. InProceedings of the IEEE/CVF international conference on computer vision, pages 7556–7566, 2019. 2

    2019

  45. [53]

    Istvt: Interpretable spatial-temporal video transformer for deepfake detection.IEEE Transactions on Information F orensics and Security, 18:1335–1348, 2023

    Cairong Zhao, Chutian W ang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui T ang. Istvt: Interpretable spatial-temporal video transformer for deepfake detection.IEEE Transactions on Information F orensics and Security, 18:1335–1348, 2023. 6

    2023

  46. [54]

    Multi-attentional deepfake de- tection

    Hanqing Zhao, W enbo Zhou, Dongdong Chen, Tianyi W ei, W eiming Zhang, and Nenghai Y u. Multi-attentional deepfake de- tection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2

    2021

  47. [55]

    Exploring temporal coherence for more general video face forgery detection

    Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang W en. Exploring temporal coherence for more general video face forgery detection. InProceedings of the IEEE/CVF International Conference on Computer V ision (ICCV), pages 15044–15054,

  48. [56]

    Multiple recurrent attention convolutional neural network for fine-grained image recognition

    Xiaotong Zhu and Hengwei Bian. Multiple recurrent attention convolutional neural network for fine-grained image recognition. In 2022 International Conference on Image Processing, Com- puter V ision and Machine Learning (ICICML), pages 44–48. IEEE, 2022. 4 7

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.