WaveConViT: Wavelet-Based Convolutional Vision Transformer for Cross-Manipulation Deepfake Video Detection
Open accessThe DOI matches its Crossref record; 37 of the 37 references it cites were found in a registry or an index; 8 of the 12 sources its claims cite were checked for support.
Overview
CoChat Summary
TL;DR
WaveConViT is a spatio-temporal deepfake video detector that combines an ImageNet-initialized ConvNeXt-T, a ViT, and a DWT-based high-frequency branch with cross-attention. On FaceForensics++, it reaches 97.12% accuracy and 96.46 AUC with 10-frame sequences and a frame skip of 2, and it also outperforms baselines in cross-dataset generalization.
Key findings
- In the known-manipulation setting on FF++, WaveConViT 10f, s = 2 achieves 97.12% accuracy and 96.46 AUC, beating the best listed baseline accuracy (ConvNeXt-T at 96.79%) and AUC (Convolutional Cross ViT at 94.23).
- In cross-manipulation testing after training on FF++, WaveConViT reaches 71.33% accuracy / 73.29 AUC on Celeb-DF with 10f, s = 1, and 84.40% accuracy / 79.37 AUC on DF-1.0 with 20f, s = 2.
- The paper reports that adding the cross-attention module improves generalization accuracy by 0.73% on Celeb-DF and 1.6% on DF-1.0.
- The DWT-based high-frequency branch is reported to outperform the SRM-based alternative for cross-dataset performance, and the RGB branch alone outperforms the high-frequency branch when each is used by itself.
Methods
- The model has two spatio-temporal branches: an RGB branch using an ImageNet-1K-initialized ConvNeXt-T followed by a 2-layer, 4-head ViT, and a high-frequency branch that applies a single-level 2D Haar DWT before the same kind of feature extractor.
- A cross-attention module fuses the RGB and high-frequency streams after the ConvNeXt-T feature extraction stage; the two ViT outputs are concatenated and passed to a 2-layer MLP classification head for binary classification.
- Experiments use FaceForensics++ (1,000 real / 5,000 fake), Celeb-DF (590 real / 5,639 fake), and DeeperForensics-1.0 std (1,000 fake); videos are preprocessed by taking the first 120 frames, detecting/cropping faces with MTCNN, resizing to 224×224, and sampling sequences of 10, 20, or 60 frames with optional frame skipping.
Limitations
- The paper does not report a separate external validation on datasets beyond FF++, Celeb-DF, and DF-1.0, so broader real-world generalization remains untested in the text.
- The authors note that performance in cross-manipulation depends heavily on the target dataset and sequence configuration; this makes the best setting dataset-specific rather than universal.
- The paper itself notes that longer sequences may contain richer temporal information but yield fewer training/test data points; this is a trade-off that can limit performance, especially for the ViT-based approach.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
8 Valid · 1 Not valid · 5 Not checked · 40 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“In order to enhance interaction between the two main branches, a cross-attention module similar to the one proposed in [25] is applied after the first stage of the feature extraction process (i.e., to the output of the two ConvNeXt-T extractors).”
What the sources say
- Supportedcites [25] Qian, Y., Yin, G., Sheng, L., Chen, Z., Shao, J.: Thinking in frequency: face forgery detection by mining frequency-aware clues. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.-M. (eds.) ECCV 2020. LNCS, vol. 12357, pp. 86–103. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-58610-2_6 source
Section 3.3 (p. 8) describes a cross-attention fusion module for feature interaction between the paper’s two branches. The Introduction (p. 3) likewise says MixBlock encourages interactions between them. The citing paper’s use of ConvNeXt-T extractors and its placement of the module describe its own implementation, not a feature attributed to this source. (LLM, conf=0.94)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“This reinforces the findings in [3], where high-frequency filtering was shown to suppress meaningful temporal manipulation artifacts such as color inconsistency between successive frames in a deepfake video or flickering artifacts.”
What the sources say
- Supportedcites [3] Atamna, M., Tkachenko, I., Miguet, S.: Improving generalization in facial manipulation detection using image noise residuals and temporal features. In: 2023 IEEE International Conference on Image Processing (ICIP), pp. 3424–3428 (2023) source
Section 4.2 (Generalization performance) states that pre-processing with SRM filters 'removes various temporal manipulation artifacts such as abrupt illumination changes and color inconsistency between successive frames, which are very valuable in temporal analysis', and that high-frequency information is therefore insufficient by itself — matching the claim that high-frequency filtering suppresses meaningful temporal artifacts, including color inconsistency between successive frames. The claim's second example, 'flickering artifacts', corresponds to the paper's 'abrupt illumination changes' (a close but not literal paraphrase). The Conclusion (p. 5) repeats the point: 'high-frequency information alone is insufficient for image sequence classification as other important clues, found in low-frequency image content, are suppressed.' (LLM, conf=0.85)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“This trade-off may help explain why WaveConViT performs better with relatively shorter sequences, as Vision Transformers are notorious for performing best when trained on larger datasets [12].”
What the sources say
- Supportedcites [12] Dosovitskiy, A., et al.: An image is worth 16x16 words: transformers for image recognition at scale. In: International Conference on Learning Representations (2021) source
The citation supports the sentence’s claim about Vision Transformers and dataset size. Section 4.3 (p. 7) reports that ViTs overfit more than comparable ResNets on smaller pre-training datasets but perform better with larger ones. (LLM, conf=0.93)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“As shown in [3], frequency-aware deepfake detection works better when learning occurs on a diverse set of tampering artifacts.”
What the sources say
- Partly supportedcites [3] Atamna, M., Tkachenko, I., Miguet, S.: Improving generalization in facial manipulation detection using image noise residuals and temporal features. In: 2023 IEEE International Conference on Image Processing (ICIP), pp. 3424–3428 (2023) sourceExperimental — accuracy not yet measured
Section 4.2 (Generalization performance) states that 'training XceptionNet in the noise residual domain generally outperforms the RGB domain, in particular when numerous manipulation techniques are included in the training set', which supports the claim's core: frequency/noise-residual-aware detection gains most when trained on a diverse set of tampering artifacts. However, the paper restricts this to the spatial (image-classification) architecture and immediately qualifies it: 'The results for the spatio-temporal architecture, however, are mixed', and the Conclusion (Section 5) adds that 'high-frequency information alone is insufficient for image sequence classification'. The unqualified generalization to 'deepfake detection' in the citing sentence therefore overstates what this source shows. Note also that the paper's 'diverse set' means multiple manipulation techniques/datasets (FF++, DF-1.0, Celeb-DF) rather than a variety of artifact types per se. (LLM, conf=0.75)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“An ablation study demonstrates the soundness of our approach and validates the usefulness of each component of our architecture.”
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
21 more claims
Create a free account to see all 54 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 17
- The quote only partly states this claim. 2
- It may state two separate claims as one. 12
Citations
Cites
ImageNet: A large-scale hierarchical image database
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · 2009 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2009.5206848 · Cited by 65,333
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 47,472
Xception: Deep Learning with Depthwise Separable Convolutions
Chollet, F. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.195 · Cited by 20,380
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
Tan, M., & Le, Q. V. (2019) · arXiv · doi:10.48550/arxiv.1905.11946 · Cited by 17,072 · cited 1×
“For both Efficient ViT and Convolutional Cross ViT, we use an EfficientNet-B0 [29] that is pre-trained on ImageNet as a feature extractor, as per the original paper.”
Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., & Xie, S. (2022) · arXiv · Cited by 8,924
Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., & Xie, S. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01167 · Cited by 8,494
The wavelet transform, time-frequency localization and signal analysis
Daubechies, I. (1990) · IEEE Transactions on Information Theory · doi:10.1109/18.57199 · Cited by 6,483 · cited 1×
“Two of the simpler wavelets from the Daubechies family of wavelets [8], db1 and db2, were also evaluated in the early stages of experimentation; however, we settled on the Haar wavelet as no appreciable difference in performance was observed.”
Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks
Zhang, K., Zhang, Z., Li, Z., & Qiao, Y. (2016) · IEEE Signal Processing Letters · doi:10.1109/lsp.2016.2603342 · Cited by 6,214
FaceForensics++: Learning to Detect Manipulated Facial Images
Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., & Niessner, M. (2019) · 2019 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00009 · Cited by 3,314
ImageNet: A large-scale hierarchical image database
Deng, J. H., Dong, W. F., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvprw.2009.5206848 · Cited by 2,714
Research works printed but not linked: 2
- Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)
- Luo, Y., Zhang, Y., Yan, J., Liu, W.: Generalizing face forgery detection with high-frequency features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 16317–16326, June 2021 · DOI · Paper page
Cited by
HFVideoSwin: High-Frequency Spatio-Temporal Features for More Generalizable Deepfake Video Detection
Atamna, M., Tkachenko, I., & Miguet, S. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-31452-9_11
Self-citation
Metadata
- Authors
- Atamna, M., Tkachenko, I., & Miguet, S.
- Affiliations
- Lyon 1 Université (FR)
- Corresponding author
- Atamna, Mehdi
- Venue
- Lecture Notes in Computer Science
- Publisher
- Springer Nature Switzerland
- Published
- 2025
- Pages
- 211-225
- ISSN
- 0302-9743 (ISSN-L)
- 1611-3349
- OpenAlex
- W4406779112
- Semantic Scholar
- 495501b9e8d3927c13b0710f1b06ee52aeeb1155
- Cited by
- 1
- References
- 31
- Citations by year
- 2026: 1
- FWCI
- 0.78
- Type
- conference paper
- Language
- en
- Open access
- green
- hal.science
- Licence
- other-oa
- www.springernature.com/gp/researchers/text-and-data-mining (tdm)
- www.springernature.com/gp/researchers/text-and-data-mining (vor)
- Locations
- Lecture notes in computer science (publishedVersion)
- HAL (Le Centre pour la Communication Scientifique Directe) (submittedVersion, other-oa)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Face recognition and analysis
- Keywords
- deepfake detection, face manipulation, vision transformer, discrete wavelet transform, frame skipping, temporal sampling, high-frequency characteristics
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: WaveConViT: Wavelet-Based Convolutional Vision Transformer for Cross-Manipulation Deepfake Video Detection. https://app.cochat.ai/papers/10.1007/978-3-031-88223-4_16
- The paper
- Atamna, M., Tkachenko, I., & Miguet, S. (2025). WaveConViT: Wavelet-Based Convolutional Vision Transformer for Cross-Manipulation Deepfake Video Detection. Lecture Notes in Computer Science. https://doi.org/10.1007/978-3-031-88223-4_16
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.