Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

The paper argues that SFMFNet, a 6.64-million-parameter network, reaches a 0.8682 average cross-dataset AUC by fusing wavelet frequency cues with spatial attention and selective cross-scale attention, outperforming heavier detectors like UC

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 ↗

SFMFNet uses wavelet-frequency gating, token-selective cross-attention, and blur pooling to reach 0.8682 average cross-dataset AUC with only 1.27 GFLOPs and 6.64M parameters.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible lightweight detector, but the cross-dataset numbers are internally inconsistent and the headline claim isn't yet trustworthy. the 3 major comments →

arxiv 2508.20449 v1 pith:2LQFIM5B submitted 2025-08-28 cs.CV

A Spatial-Frequency Aware Multi-Scale Fusion Network for Real-Time Deepfake Detection

Libo Lv , Tianyi Wang , Mengxiao Huang , Ruixia Liu , Yinglong Wang This is my paper
classification cs.CV
keywords spatial-frequency fusiondeepfake detectionlightweight networkreal-time detectionmulti-scale fusioncross attentionblur poolingwavelet transform
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper is trying to establish that a lightweight deepfake detector can close the generalization gap with much heavier models by explicitly coupling spatial textures with frequency artifacts. SFMFNet combines a gated spatial-frequency attention module, a token-selective cross-attention mechanism, and a blur-pooling downsampler with residual compensation. On FaceForensics++ it reaches a 0.9867 average AUC, and cross-dataset it averages 0.8682 AUC across five unseen benchmarks, using 1.27 GFLOPs and 6.64M parameters. That beats the lightweight baselines UCF (0.8267 avg AUC, 10.53 GFLOPs) and FFD (0.8164 avg AUC, 6.01 GFLOPs). A sympathetic reader would care because it suggests real-time deployment need not sacrifice generalization.

Core claim

On its own terms, the paper claims that SFMFNet detects manipulated faces more accurately than every lightweight detector it compares against, at lower computational cost. The core evidence is the AUC tables: an in-domain FF++ average AUC of 0.9867 and a cross-dataset average AUC of 0.8682 on DFD, Celeb-DF v2, DFDCP, DFDC, and UADFV, at only 1.27 GFLOPs and 6.64M parameters. The authors attribute this to three modules: a spatial-frequency hybrid aware (SFHA) module that gates features using wavelet sub-bands and coordinate-attention maps; a token-selective cross-attention (TSCA) module that prunes tokens before cross-attending fused low-level features with projected high-level features; and

What carries the argument

The load-bearing object is the spatial-frequency hybrid gating map, a sigmoid-activated fusion of a wavelet-band attention map (four Haar sub-bands processed by convolutions) and a coordinate-enhanced spatial attention map, applied to the backbone features through a residual multiplication. Around it, the token-selective cross-attention reduces the query, key, and value maps to t×t tokens before scaled dot-product attention, letting high-level semantics enrich low and mid-level features at negligible cost; and the residual blur-pooling downsampler suppresses aliasing while keeping edge cues. These three mechanisms are what let a small CNN backbone behave like a much larger detector.

Load-bearing premise

The reported edge over UCF and FFD rests on single-run, best-validation checkpoint AUCs compared across methods under one protocol; without repeated runs or variance estimates, the 4–6 point average-AUC gaps could be within training noise.

What would settle it

Retrain SFMFNet, UCF, and FFD at least five times on the FF++ training split, keeping the same benchmark protocol, and compare the mean and standard deviation of cross-dataset AUC. If the gap between SFMFNet and UCF (0.8682 vs 0.8267) falls within one standard deviation under any seed variation, the claimed advantage is not established. A complementary check is to evaluate the released model on a dataset absent from the paper's table, such as WildDeepfake, and compare with the same baselines.

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

If this is right

  • Real-time live-stream moderation becomes feasible: at 1.27 GFLOPs a 256×256 frame can be classified on edge GPUs without specialized pruning or quantization.
  • The average cross-dataset AUC of 0.8682, if stable, positions SFMFNet as the strongest accuracy-efficiency point among the lightweight detectors compared in the paper.
  • The gated regularization and consistency losses tie per-scale attention maps to the global label, so the model's gating responses could serve as coarse manipulation localization in addition to classification.
  • Removing SFHA drops Celeb-DF AUC from 0.8111 to 0.7427 in the ablation, so the spatial-frequency fusion carries much of the cross-dataset generalization.

Where Pith is reading between the lines

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

  • If the single-run results are reliable, the underlying design—cheap fusion of wavelet artifacts and spatial attention with token pruning—may transfer to other artifact-detection tasks such as detecting diffusion-generated images or GAN-generated audio spectrograms.
  • The reported weakness on UADFV (0.9420 vs UCF's 0.9528) suggests no single lightweight recipe dominates all domains; a dataset-adaptive gating weight might close that gap.
  • A direct test of the efficiency claim would be to compare SFMFNet against UCF pruned or quantized to the same 1.27 GFLOPs budget, something the paper does not report.
  • The paper's protocol fixes input at 256×256; real-time streams often need lower resolutions and stronger video compression, so evaluating SFMFNet under H.264 or H.265 recompression would show whether the frequency cues survive.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes SFMFNet, a lightweight CNN-based deepfake detector that combines a spatial-frequency hybrid aware gating module (SFHA), a blur-pooling residual downsampling module (DBP), and a token-selective cross-attention module (TSCA), trained with a weighted combination of binary cross-entropy, gated regularization, and consistency losses. On FaceForensics++ the method reports an average AUC of 0.9867; on six benchmark datasets under the DeepfakeBench protocol it reports an average cross-dataset AUC of 0.8682 with 1.27 GFLOPs and 6.64M parameters, outperforming the lightweight baselines UCF (0.8267 avg AUC) and FFD (0.8164 avg AUC). Ablations on CDF2 and DFDC attribute large gains to each proposed module. The paper is an empirical architecture study; no derivation is circular, and the reported results are not trivially implied by the loss definitions.

Significance. Real-time deepfake detection with a favorable accuracy/efficiency trade-off is practically important, and the reported numbers—0.8682 average cross-dataset AUC at 1.27 GFLOPs and 6.64M parameters—would be attractive if confirmed. The paper evaluates on six datasets, provides FLOPs/parameter counts, and includes ablations for the main modules. The architectural components are mostly established ingredients (wavelet transforms, coordinate attention, blur pooling, cross-attention) combined in a new way, so the contribution is incremental but potentially useful. However, the numerical evidence is not yet fully trustworthy: there is a direct inconsistency between Tables 1 and 2 for GoogLeNet, the backbone is never specified, and no repeated runs or significance tests are reported. These issues are load-bearing because the central claim is the cross-dataset comparison against lightweight baselines.

major comments (3)
  1. [Section 4.2, Tables 1 and 2] GoogLeNet is reported with an FF++ AUC of 0.9810 in Table 1 but 0.8321 in Table 2, while all other shared baseline models have identical values across the two tables. The paper states that all experiments follow the same DeepfakeBench preprocessing and the same best-validation-checkpoint protocol, so a single protocol cannot yield both values. This indicates that at least some Table 2 entries may have been taken from a different evaluation setting. Please correct the incorrect value or explain the discrepancy. If Table 2 was not generated under the stated unified protocol, the cross-dataset comparisons with UCF, FFD, and other baselines are not yet established as apples-to-apples.
  2. [Sections 3 and 4.1] The backbone is never named or specified. The text says only 'a CNN backbone' and Figure 1 shows generic blocks; there is no architecture name, channel configuration, depth, pretraining strategy, or input stem. The reported FLOPs and parameters depend directly on this unspecified backbone, and a reader cannot reconstruct the model from the paper. Since the central claim is a specific accuracy/efficiency trade-off, this is a reproducibility gap. Please specify the backbone architecture in detail or release code/configurations.
  3. [Sections 4.2 and 4.3, Tables 2–4] All reported AUC values are single runs using the best-validation checkpoint, with no standard deviations, confidence intervals, repeated seeds, or significance tests. The claimed gains over UCF and FFD are large in the aggregate, but some numbers are close (e.g., DFDC 0.7307 vs 0.7179 for UCF; UADFV is 0.9420 vs 0.9528), and without variance estimates the comparison is not statistically grounded. At minimum, report mean and std over multiple runs for the main Table 2 comparisons and for the ablations, or provide a significance test.
minor comments (5)
  1. [Section 4 title] 'Experients' should be 'Experiments'.
  2. [Eq. (12) and Eq. (11)] The parentheses in Eq. (12) are unbalanced (an opening parenthesis before Softmax has no matching close). In Eq. (11), the notation R((1/C) sum_{c=1}^C (Q1,K2,V2)) is ambiguous about whether the pooling is applied jointly or separately to Q, K, and V, and the channel dimension C is not defined at that point.
  3. [Figure 3 caption] The axis label 'FL OPs (G)' contains a typo; it should be 'FLOPs (G)'.
  4. [Section 2.2] GoogLeNet is described as an attention mechanism; while it uses inception modules, this characterization is imprecise and may confuse readers comparing with modern attention modules.
  5. [Table 2] No source or measurement protocol is given for the FLOPs and parameter counts of baseline models. Please state how these complexity numbers were computed (input resolution, framework, whether all tested at the same settings).

Circularity Check

0 steps flagged

No circularity: SFMFNet is an empirical architecture evaluated against external benchmarks; no prediction reduces to its training inputs by construction.

full rationale

The paper's central claims are empirical: SFMFNet's AUCs on FF++ and cross-dataset benchmarks (Tables 1 and 2) are obtained by training a network and testing on held-out data. The proposed modules (SFHA, DBP, TSCA) and loss functions are standard supervised-learning components. The gated regularization loss uses labels to supervise attention maps, but this is a training objective, not a fitted parameter renamed as a prediction; the final test AUC is not equal to the loss weights or the attention maps by construction. The paper cites several prior works by the same authors (e.g., refs. [3], [4], [11], [12]), but these are background citations for face swapping, surveys, and previous detectors; they are not used to justify the architecture's uniqueness or to import a result that forces the reported numbers. No uniqueness theorem is invoked, no fitted input is called a prediction, and no known result is merely renamed. The empirical comparison is self-contained in the sense that results are evaluated against external benchmark datasets and baseline methods. A separate concern—the apparent internal inconsistency of GoogLeNet's FF++ AUC between Table 1 (0.9810) and Table 2 (0.8321)—is a validity/reproducibility issue, not a circularity issue, and does not change the circularity verdict.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method is an empirical architecture paper. The central claim depends on hand-chosen loss weights, an unreported token pooling size, an unspecified backbone, and standard domain assumptions about the utility of frequency/spatial cues and benchmark validity. No new physical or mathematical entities are introduced.

free parameters (4)
  • loss weights lambda_CE, lambda_gate, lambda_cons = 1.0, 0.3, 0.2
    Chosen by hand in Section 4.1; no sensitivity analysis or grid search is reported.
  • token pooling size t in TSCA
    Defined in Eq. 11 but never given a numerical value in the paper, leaving the model underspecified.
  • input resolution and training hyperparameters = 256x256, batch 32, lr 0.0002
    Standard settings from Section 4.1; they affect the reported AUC and are not swept.
  • backbone architecture (unnamed)
    Section 3 refers only to 'a CNN backbone'; the specific network is never named, yet it determines most of the 6.64M parameters and 1.27 GFLOPs.
axioms (4)
  • domain assumption Wavelet-domain features and spatial attention maps capture complementary forgery clues.
    Motivates the SFHA module (Section 3.1, Eqs. 1-3) but is not proven; the ablation study supports it empirically but only on CDF2 and DFDC.
  • domain assumption Blur pooling reduces aliasing and preserves edge information during downsampling.
    Adopted from Zhang [17] and used in Eqs. 5-6; no new derivation is provided in this paper.
  • domain assumption AUC on FF++, DFD, CDF2, DFDCP, DFDC, and UADFV is a valid measure of deepfake detection generalization.
    The standard benchmark protocol following DeepfakeBench [37] is adopted in Section 4.1 without questioning the validity of the metric or protocol.
  • standard math Standard mathematical operations: Haar wavelet transform, scaled dot-product attention, backpropagation.
    Used throughout Section 3 without proof; these are well-established operations.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Spatial-Frequency Aware Multi-Scale Fusion Network for Real-Time Deepfake Detection." pith.science (2026). https://pith.science/paper/2LQFIM5B

@misc{pith2026250820449,
  author       = {Pith},
  title        = {Pith review of: A Spatial-Frequency Aware Multi-Scale Fusion Network for Real-Time Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2LQFIM5B}},
  note         = {Machine review of arXiv:2508.20449}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

With the rapid advancement of real-time deepfake generation techniques, forged content is becoming increasingly realistic and widespread across applications like video conferencing and social media. Although state-of-the-art detectors achieve high accuracy on standard benchmarks, their heavy computational cost hinders real-time deployment in practical applications. To address this, we propose the Spatial-Frequency Aware Multi-Scale Fusion Network (SFMFNet), a lightweight yet effective architecture for real-time deepfake detection. We design a spatial-frequency hybrid aware module that jointly leverages spatial textures and frequency artifacts through a gated mechanism, enhancing sensitivity to subtle manipulations. A token-selective cross attention mechanism enables efficient multi-level feature interaction, while a residual-enhanced blur pooling structure helps retain key semantic cues during downsampling. Experiments on several benchmark datasets show that SFMFNet achieves a favorable balance between accuracy and efficiency, with strong generalization and practical value for real-time applications.

Figures

Figures reproduced from arXiv: 2508.20449 by Libo Lv, Mengxiao Huang, Ruixia Liu, Tianyi Wang, Yinglong Wang.

Figure 1
Figure 1. Figure 1: Overall architecture of the proposed lightweight forgery detection model. CNN backbone extracts multi-level features, enhanced by spatial-frequency hybrid aware module to highlight forgery regions. Low-level features are downsampled and fused with mid-level features, then refined through token-selective cross-attention with high￾level features. Finally, global pooling and a multilayer perceptron classify f… view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE visualization for each detector. These detectors are trained and tested on FF++. Given the real-time requirements of our task, we focus on detectors with low inference costs. Large-scale vision-language and diffusion-based detectors, such as ASAP [23], often exceed 100 GFLOPs and billions of parameters, making them impractical for latency-sensitive applications. In contrast, spatial- and frequency￾ba… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of AUC and FLOPs, highlighting our model’s superior performance and efficiency. To provide a more intuitive comparison, [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of average pooling and our DBP module. DBP preserves clearer edges and structural details. To assess the efficiency and effectiveness of the TSCA module, we conducted two ablation studies shown in [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. CAM-VFD: Cross-Attention Multimodal Video Forgery Detection

    cs.CV 2026-05 unverdicted novelty 6.0

    CAM-VFD detects video forgeries by using cross-attention to identify contradictions between CLIP appearance, VideoMAE motion, and MiDaS depth features.

Reference graph

Works this paper leans on

36 extracted references · 15 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    In: Proceedings of the 28th ACM Interna- tional Conference on Multimedia (ACM MM), pp

    Chen, R., Chen, X., Ni, B., Ge, Y.: SimSwap: An Efficient Fram ework for High Fidelity Face Swapping. In: Proceedings of the 28th ACM Interna- tional Conference on Multimedia (ACM MM), pp. 2003–2011. AC M (2020). https://doi.org/10.1145/3394171.3413630

    arXiv 2003

  2. [2]

    StarGAN v2: Diverse Image Synthesis for Multiple Domains

    Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: StarGAN v2: Diverse Ima ge Synthesis for Multiple Domains. arXiv preprint arXiv:1912 .01865 (2020). https://arxiv.org/abs/1912.01865

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  3. [3]

    IEEE Trans

    Wang, T., Li, Z., Liu, R., Wang, Y., Nie, L.: An Efficient Attr ibute-Preserving Framework for Face Swapping. IEEE Trans. Multimedia 26, 655 4–6565 (2024). https://doi.org/10.1109/TMM.2024.3354573

    work page arXiv 2024

  4. [4]

    ACM Co mput

    Wang, T., Liao, X., Chow, K.P., Lin, X., Wang, Y.: Deepfake Detection: A Com- prehensive Survey from the Reliability Perspective. ACM Co mput. Surv. 57(3), Article 58, 35 pages (2024). https://doi.org/10.1145/369 9710

    doi:10.1145/369 2024

  5. [5]

    In: IEEE 9th Europe an Sym- posium on Security and Privacy (EuroS&P), pp

    Mittal, G., Hegde, C., Memon, N.: Gotcha: Real-Time Video Deep- fake Detection via Challenge-Response. In: IEEE 9th Europe an Sym- posium on Security and Privacy (EuroS&P), pp. 1–20. IEEE (20 24). https://doi.org/10.1109/EuroSP60621.2024.00009

    arXiv 2024

  6. [6]

    R., Cinque, L.: Faster Than Lies: Real-time Deepfake Detection using Binary Neural Networks

    Lanzino, R., Fontana, F., Diko, A., Marini, M. R., Cinque, L.: Faster Than Lies: Real-time Deepfake Detection using Binary Neural Networks . In: IEEE/CVF Conf. Real-Time Deepfake Detection with Spatial-Frequency Fusi on 13 on Computer Vision and Pattern Recognition Workshops (CVPR W), pp. 3771–

    work page

  7. [7]

    In: arXiv preprint arXiv:2406 .09398 (2024).https://arxiv.org/abs/2406.09398

    Cavia, B., Horwitz, E., Reiss, T., Hoshen, Y.: Real-Time D eepfake Detection in the Real-World. In: arXiv preprint arXiv:2406 .09398 (2024).https://arxiv.org/abs/2406.09398

    Pith/arXiv arXiv 2024

  8. [8]

    In: IEEE Internation al Work- shop on Information Forensics and Security (WIFS), pp

    Li, Y., Chang, M.-C., Lyu, S.: In Ictu Oculi: Exposing AI Cr eated Fake Videos by Detecting Eye Blinking. In: IEEE Internation al Work- shop on Information Forensics and Security (WIFS), pp. 1–7. IEEE (2018). https://doi.org/10.1109/WIFS.2018.8630787

    work page arXiv 2018

  9. [9]

    In: IEEE Conference on Computer Vision and Pattern Recognit ion Workshops (CVPR W), pp

    Li, Y., Lyu, S.: Exposing Deepfake Videos by Detecting Fac e Warping Artifacts. In: IEEE Conference on Computer Vision and Pattern Recognit ion Workshops (CVPR W), pp. 1–5. IEEE (2019). https://arxiv.org/abs/181 1.00656

    work page 2019

  10. [10]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision ( ICCV), pp

    R¨ ossler, A., Cozzolino, D., Verdoliva, L., Riess, C., T hies, J., Nießner, M.: Face- Forensics++: Learning to Detect Manipulated Facial Images . In: Proceedings of the IEEE/CVF International Conference on Computer Vision ( ICCV), pp. 1–11. IEEE (2019). https://doi.org/10.1109/ICCV.2019.00873

    work page arXiv 2019

  11. [11]

    P.: Noise Based Deepfake Detection via Multi-Head Relative-Interaction

    Wang, T., Chow, K. P.: Noise Based Deepfake Detection via Multi-Head Relative-Interaction. In: Proceedings of the AAAI Confere nce on Artifi- cial Intelligence, vol. 37, no. 12, pp. 14548–14556. AAAI Pr ess (2023). https://doi.org/10.1609/aaai.v37i12.26701

    work page doi:10.1609/aaai.v37i12.26701 2023

  12. [12]

    P., Nie, L.: Deep Convolutio nal Pooling Trans- former for Deepfake Detection

    Wang, T., Cheng, H., Chow, K. P., Nie, L.: Deep Convolutio nal Pooling Trans- former for Deepfake Detection. In: ACM Transactions on Mult imedia Computing, Communications, and Applications, vol. 19, no. 6, pp. 1–12. ACM Press (2023). https://doi.org/10.1145/3588574

    work page doi:10.1145/3588574 2023

  13. [13]

    In: IEEE International Conference on Au- tomatic Face and Gesture Recognition (FG), pp

    Das, A., Das, S., Dantcheva, A.: Demystifying Attention Mechanisms for Deepfake Detection. In: IEEE International Conference on Au- tomatic Face and Gesture Recognition (FG), pp. 1–7. IEEE (20 21). https://doi.org/10.1109/FG52635.2021.9667026

    arXiv 2021

  14. [14]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path Aggregatio n Net- work for Instance Segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8759–8768. IEEE (2018). https://doi.org/10.1109/CVPR.2018.00913

    arXiv 2018

  15. [15]

    In: Proceedings of the European Conference on Compu ter Vision (ECCV), pp

    Woo, S., Park, J., Lee, J.Y., Kweon, I.S.: CBAM: Convolut ional Block Attention Module. In: Proceedings of the European Conference on Compu ter Vision (ECCV), pp. 3–19. Springer (2018). https://doi.org/10.1007/978- 3-030-01234-2 1

    doi:10.1007/978- 2018

  16. [16]

    De Silva, D. D. N., Vithanage, H. W. M. K., Xavier, S. A., Pi yatilake, I. T. S. and Fernando, S.: Parameterized Wavelets for Convolut ional Neural Net- works. In: IEEE International Conference on Autonomous Rob ot Systems and Competitions (ICARSC), Ponta Delgada, Portugal, pp. 170–1 76. IEEE (2020). https://doi.org/10.1109/ICARSC49921.2020.9096125

    arXiv 2020

  17. [17]

    In: Proceed- ings of the International Conference on Machine Learning (I CML), pp

    Zhang, R.: Making Convolutional Networks Shift-Invari ant Again. In: Proceed- ings of the International Conference on Machine Learning (I CML), pp. 7324–7334. PMLR (2019). https://arxiv.org/abs/1904.11486

    Pith/arXiv arXiv 2019

  18. [19]

    arXiv preprint arXiv:2006.07397 (2020)

    Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., W ang, M., Canton Ferrer, C.: The DeepFake Detection Challenge (DFDC) Datase t. arXiv preprint arXiv:2006.07397 (2020). https://arxiv.org/abs/2006.0 7397 14 Lv et al

    Pith/arXiv arXiv 2006

  19. [20]

    https://ai.googleblog.com/2019/09/contributing-data-to-deepfake-detection.html

    Google AI: Contributing Data to Deepfake Detection Rese arch. https://ai.googleblog.com/2019/09/contributing-data-to-deepfake-detection.html. Last accessed 9 Jun 2025

    work page 2019

  20. [21]

    arXiv preprin t arXiv:1910.08854 (2019)

    Dolhansky, B., Howes, R., Pflaum, B., Baram, N., Canton Fe rrer, C.: The Deepfake Detection Challenge (DFDC) Preview Dataset. arXiv preprin t arXiv:1910.08854 (2019). https://arxiv.org/abs/1910.08854

    Pith/arXiv arXiv 1910

  21. [22]

    arXiv preprint arXiv :1806.02877 (2018)

    Li, Y., Chang, M.-C., Lyu, S.: In Ictu Oculi: Exposing AI G enerated Fake Face Videos by Detecting Eye Blinking. arXiv preprint arXiv :1806.02877 (2018). https://arxiv.org/abs/1806.02877

    Pith/arXiv arXiv 2018

  22. [23]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Patte rn Recognition (CVPR), pp

    Zhao, Y., Lu, Z., Gong, Y., Song, J., Yang, Y.: ASAP: Advan cing Semantic Align- ment Promotes Multi-Modal Manipulation Detecting and Grou nding. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Patte rn Recognition (CVPR), pp. 1–11. IEEE (2025). https://arxiv.org/abs/241 2.12718

    work page 2025

  23. [24]

    In: Advances in Neural Info rmation Processing Systems (NeurIPS), vol

    Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet C lassification with Deep Convolutional Neural Networks. In: Advances in Neural Info rmation Processing Systems (NeurIPS), vol. 25, pp. 1097–1105. Curran Associat es, Red Hook (2012). https://doi.org/10.1145/3065386

    doi:10.1145/3065386 2012

  24. [25]

    In: Proceedings of the IEEE Conference on Comp uter Vision and Pattern Recognition (CVPR), pp

    He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learnin g for Image Recognition. In: Proceedings of the IEEE Conference on Comp uter Vision and Pattern Recognition (CVPR), pp. 770–778. IEEE, Las Vega s (2016). https://doi.org/10.1109/CVPR.2016.90

    doi:10.1109/cvpr.2016.90 2016

  25. [26]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Chollet, F.: Xception: Deep Learning with Depthwise Sep arable Convo- lutions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1251–1258. IEEE, Honol ulu (2017). https://doi.org/10.1109/CVPR.2017.195

    doi:10.1109/cvpr.2017.195 2017

  26. [27]

    In: Proceedings of the European Con ference on Com- puter Vision (ECCV), LNCS, vol

    Qian, Y., et al.: Thinking in Frequency: Face Forgery Det ection by Mining Frequency-Aware Clues. In: Proceedings of the European Con ference on Com- puter Vision (ECCV), LNCS, vol. 12357, pp. 86–103. Springer , Glasgow (2020). https://doi.org/10.1007/978-3-030-58610-2 6

    work page doi:10.1007/978-3-030-58610-2 2020

  27. [29]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recogn ition (CVPR), pp

    Li, L., et al.: Face X-ray for More General Face Forgery De tection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recogn ition (CVPR), pp. 5000–5009. IEEE, Seattle (2020). https://doi.org/10.1109/CVPR42600.2020.00505

    arXiv 2020

  28. [30]

    On the Detection of Digital Face Manipulation

    Dang, H., et al.: On the Detection of Digital Face Manipul ation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recogn ition (CVPR), pp. 5781–5790. IEEE, Seattle (2020). https://arxiv.org/a bs/1910.01717

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  29. [31]

    In: Proceedings of the IEEE Co nference on Computer Vision and Pattern Recognition (CVPR), pp

    Yan, Z., Zhang, Y., Fan, Y., Wu, B.: UCF: Uncovering Commo n Features for Gener- alizable Deepfake Detection. In: Proceedings of the IEEE Co nference on Computer Vision and Pattern Recognition (CVPR), pp. 12454–12463. IE EE, Virtual (2023). https://doi.org/10.1109/CVPR52729.2023.01210

    arXiv 2023

  30. [32]

    IEEE Transactions on Neural N etworks and Learning Systems, 34(11), 9562–9567 (2023)

    Xu, J., Pan, Y., Pan, X., Hoi, S., Zhang, Y., Xu, Z.: RegNet : Self-Regulated Net- work for Image Classification. IEEE Transactions on Neural N etworks and Learning Systems, 34(11), 9562–9567 (2023). https://doi.org/10.1 109/TNNLS.2022.3158966

    arXiv 2023

  31. [33]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recogn ition (CVPR), Boston, MA, USA, pp

    Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Ang uelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going Deeper with Convoluti ons. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recogn ition (CVPR), Boston, MA, USA, pp. 1–9 (2015). https://doi.org/10.1109/ CVPR.2015.7298594 Real-Time Deepfake Detection with Spatial-Fre...

    work page arXiv 2015

  32. [34]

    arXiv preprint arXiv:1811.00656 (2018)

    Li, Y., Lyu, S.: Exposing DeepFake Videos by Detecting Fa ce Warping Artifacts. arXiv preprint arXiv:1811.00656 (2018). https://arxiv.o rg/abs/1811.00656

    Pith/arXiv arXiv 2018

  33. [35]

    In: ICASSP 2019 - 20 19 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processi ng (ICASSP), Brighton, UK, pp

    Nguyen, H.H., Yamagishi, J., Echizen, I.: Capsule-Fore nsics: Using Capsule Net- works to Detect Forged Images and Videos. In: ICASSP 2019 - 20 19 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processi ng (ICASSP), Brighton, UK, pp. 2307–2311. IEEE (2019). https://doi.org/10.1109/ ICASSP.2019.8682602

    work page arXiv 2019

  34. [36]

    In: Proceedings of the IEEE/CVF Conference on Compu ter Vision and Pattern Recognition (CVPR), Nashville, TN, USA, pp

    Liu, H., Li, X., Zhou, W., Chen, Y., He, Y., Xue, H., Zhang, W., Yu, N.: Spatial-Phase Shallow Learning: Rethinking Face Forgery D etection in Frequency Domain. In: Proceedings of the IEEE/CVF Conference on Compu ter Vision and Pattern Recognition (CVPR), Nashville, TN, USA, pp. 772 –781 (2021). https://doi.org/10.1109/CVPR46437.2021.00083

    work page arXiv 2021

  35. [37]

    In: Advances in Ne ural Information Processing Systems (NeurIPS) Datasets and Benchmarks Trac k, vol

    Yan, Z., Zhang, Y., Yuan, X., Lyu, S., Wu, B.: DeepfakeBen ch: A Compre- hensive Benchmark of Deepfake Detection. In: Advances in Ne ural Information Processing Systems (NeurIPS) Datasets and Benchmarks Trac k, vol. 36 (2023). https://arxiv.org/abs/2307.01426

    Pith/arXiv arXiv 2023

  36. [3780]

    https://doi.org/10.1109/CVPR W63382.2024.00381

    IEEE (2024). https://doi.org/10.1109/CVPR W63382.2024.00381

    arXiv 2024

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