The DOI matches its Crossref record; 82 of the 82 references it cites were found in a registry or an index; 6 of the 7 sources its claims cite were checked for support.
Generalization of Forgery Detection With Meta Deepfake Detection Model
Open accessAbstract
Face forgery generating algorithms that produce a range of manipulated videos/images have developed quickly. Consequently, this causes an increase in the production of fake information, making it difficult to identify. Because facial manipulation technologies raise severe concerns, face forgery detection is gaining increasing attention in the area of computer vision. In real-world applications, face forgery detection systems frequently encounter and perform poorly in unseen domains, due to poor generalization. In this paper, we propose a deepfake detection method based on meta-learning called Meta Deepfake Detection (MDD). The goal of the model is to develop a generalized model capable of directly solving new unseen domains without the need for model updates. The MDD algorithm establishes various weights for facial images from various domains. Specifically, MDD uses meta-weight learning to shift information from the source domains to the target domains with meta-optimization steps, which aims for the model to generate effective representations of the source and target domains. We build multi-domain sets using meta splitting strategy to create a meta-train set and meta-test set. Based on these sets, the model determines the gradient descent and obtains backpropagation. The inner and outer loop gradients were aggregated to update the model to enhance generalization. By introducing pair-attention loss and average-center alignment loss, the detection capabilities of the system were substantially enhanced. In addition, we used some evaluation benchmarks established from several popular deepfake datasets to compare the generalization of our proposal in several baselines and assess its effectiveness.
Overview
CoChat Summary
TL;DR
The paper proposes Meta Deepfake Detection (MDD), a meta-learning method for generalizing face forgery detection to unseen domains without model updates. Across several CID and CVD benchmarks, MDD improves AUC over fine-tuned baselines and related methods, with gains attributed to meta-splitting, pair-attention loss, average-center alignment loss, and block shuffling transformation.
Key findings
- On CVD benchmarks, MDD improved AUC over FT-Base from 0.582 to 0.708 on CVD-CV23-1, from 0.672 to 0.788 on CVD-CV23-2, and from 0.717 to 0.821 on CVD-CV23-3.
- On CID benchmarks, MDD improved AUC over FT-Base from 0.903 to 0.931 on CID-DF23, 0.742 to 0.777 on CID-DF40, 0.792 to 0.821 on CID-FF23, 0.669 to 0.691 on CID-FF40, 0.579 to 0.658 on CID-FS23, 0.609 to 0.681 on CID-FS40, 0.764 to 0.791 on CID-NT23, and 0.618 to 0.621 on CID-NT40.
- Ablation on CID-DF23: full MDD reached AUC 0.931, ACC 0.861, LOSS 0.781; without PAL it reached 0.918/0.838/0.819; without ACA 0.921/0.842/0.802; with only softmax loss 0.907/0.829/0.846.
- Block shuffling transformation improved CID-DF23 performance from AUC 0.919 to 0.931, ACC 0.832 to 0.861, and LOSS 0.847 to 0.781. MDD also improved different backbones on CID-DF23: EfficientNet-B0 0.903→0.931, EfficientNet-B4 0.921→0.938, EfficientNet-B7 0.956→0.967, ResNet-50 0.911→0.929.
Methods
- MDD uses meta-learning with meta-splitting: source domains are split into meta-train and meta-test sets, inner-loop and outer-loop gradients are aggregated, and SGD is used for meta-optimization to generalize to unseen target domains without model updates.
- Data come from DFDC, Celeb-DF-v2, and FaceForensics CC; videos are sampled to images, faces are extracted with MTCNN and resized to 224×224 RGB, and only 10 frames of facial extraction are used for training and testing. No data augmentation is used, but block pixel shuffling transformation is applied to part of the extracted face images.
- The backbone is EfficientNet-B0 with 5.3M parameters; hyperparameters include meta-train step-size 0.0005, meta optimization learning rate 0.0005, balancing hyperparameter 0.5, PAL/SOF/ACA balancing hyperparameter 0.01, and batch size 32. Evaluation uses AUC, ACC, and log loss across seven CID/CVD benchmarks, comparing against FT-Base, Multi-task, MLDG, LTW, Multi-Att, and MAML.
Limitations
- The paper states future work will “find a new strategy to develop MDD and experiment with more benchmarks,” indicating the current evaluation is limited to the selected benchmarks.
- The paper states that “the model is less effective for complex models than simples model,” and notes that gains are harder when performance is already high.
- Evident from the text: only 10 frames per video are used for training and testing, and no data augmentation is used except block shuffling, which may limit temporal modeling and robustness; performance on some low-quality settings remains modest, e.g., CID-NT40 AUC improves only from 0.618 to 0.621.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
3 Valid · 1 Not valid · 0 Not checked · 37 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Additionally, several researches in [57] and [58] have demonstrated that creating patches by using characteristics gathered in an image also increases the quality of the training process.”
What the sources say
- Partly supportedcites [57] sourceExperimental — accuracy not yet measured
The Results (p. 7) report better deepfake-detection scores when EfficientNet B0 is used as a patch extractor, and the Conclusions (p. 9) say it extracts visual features. This supports the patch-and-feature aspect of the claim, but the paper measures detection performance, not the quality of the training process itself. (LLM, conf=0.84)
- Partly supportedcites [58] Wodajo, 2021, DeepFake video detection using convolutional vision transformer sourceExperimental — accuracy not yet measured
The Detection section (pp. 4–5) says the model extracts features from face images and splits the resulting feature maps into patches. It does not show that creating those patches improves the quality of training; the Evaluation section (pp. 5–6) reports model performance without isolating the effect of patch creation. (LLM, conf=0.85)
boostedstates a 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.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The local spatial structure of the local regions might be destroyed by the shuf ing of the pixels in an image, which prevents the network from extracting valuable features.”
What the sources say
- Partly supportedcites [51] sourceExperimental — accuracy not yet measured
“A. Block Scrambling-Based Image Encryption” (p. 2) describes randomly permuting image blocks, and the same section (p. 3) notes low correlation between blocks after encryption. This supports disruption of spatial relationships between blocks, but not the claim that individual pixels are shuffled or that a network is prevented from extracting features; the paper discusses JPEG compression and jigsaw-puzzle attacks instead. (LLM, conf=0.83)
- Partly supportedcites [52] sourceExperimental — accuracy not yet measured
The “Color-based image encryption” section describes randomly permuting image blocks, which supports the claim’s general premise that scrambling changes the image’s spatial arrangement. It does not say that pixels are shuffled within local regions or that a network is prevented from extracting features; the paper instead evaluates resistance to jigsaw-puzzle reconstruction. (LLM, conf=0.74)
- Partly supportedcites [53] sourceExperimental — accuracy not yet measured
The Brute-Force Attack section (page unavailable) says Tanaka’s method shuffles pixels within blocks, and the Use of Plain Test Images results (page unavailable) say models trained with that scheme could not classify plain test images because of encryption-related damage. The paper does not say that shuffling destroys local spatial structure or specifically prevents the network from extracting useful features; those mechanisms go beyond what our copy reports. (LLM, conf=0.78)
hedgedstates a possible 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
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, if the blocks of the image are shuf ed in a proper way, it can lead to preserving essential characteristics while also enhancing the quality of the model [55], [56].”
What the sources say
- Not checked — no full text of the sourcecites [55] source
- Partly supportedcites [56] sourceExperimental — accuracy not yet measured
The Related Work section (p. 2) says PatchShuffle shuffles pixels within local patches while largely preserving the image’s global structure. It does not say that shuffling the blocks themselves improves model quality; the performance improvements reported in the Introduction (p. 2) concern Random Erasing, a different operation. (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.
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
“The quality of the model degrades the greatest when the pair-attention loss and average-center alignment loss are not employed.”
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.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“By introducing pair-attention loss and average-center alignment loss, the detection capabilities of the system were substantially enhanced.”
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.
22 more claims
Create a free account to see all 41 claims and every source verdict.
- It may state two separate claims as one. 11
- Probably not a substantial claim. 6
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ B. SETTINGS
“In CVD-CV23-2 and CVD-CV23-3, we have used DFDC test set for target domains.”
§ Meta-train:
“| CVD-CV23-2 | NeuralTextures FaceSwap Face2Face DeepFakes | NeuralTextures FaceSwap Face2Face DeepFakes Celeb-DF-v2 | C23 |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 1) CID COMPARISONS
“from 0.609 to 0.681 in CID-FS40”
§ C. IMPLEMENTATION DETAILS
“| CID-FS40 | MDD (Ours) | 0.649 | 0.611 | 1.268 |”
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474
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 · cited 1×
“To evaluate the performance of the model, our comparisons include (i) Base: The model was pre-trained on ImageNet [60] without being fine-tuned on our benchmarks.”
MethodAuto-Encoding Variational Bayes
Kingma, D. P., & Welling, M. (2013) · arXiv · doi:10.48550/arxiv.1312.6114 · Cited by 20,067
Generative adversarial networks
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2020) · Communications of the ACM · doi:10.1145/3422622 · Cited by 18,384
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
Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
Finn, C., Abbeel, P., & Levine, S. (2017) · arXiv · doi:10.48550/arxiv.1703.03400 · Cited by 8,512
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 · cited 3×
“caded convolutional network (MTCNN) [50] library for facial extraction.”
BackgroundMethodRandom Erasing Data Augmentation
Zhong, Z., Zheng, L., Kang, G., Li, S., & Yang, Y. (2020) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v34i07.7000 · Cited by 3,617 · cited 1×
“However, if the blocks of the image are shuffled in a proper way, it can lead to preserving essential characteristics while also enhancing the quality of the model [55], [56].”
BackgroundRandom Erasing Data Augmentation
Zhong, Z., Zheng, L., Kang, G., Li, S., & Yang, Y. (2017) · arXiv · doi:10.48550/arxiv.1708.04896 · Cited by 3,617
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 · cited 6×
“for deepfake detection have been proposed in [7], [8], [9], [10], and [11].”
BackgroundMethodInfluential
Cited by
Kosarkar, U. P., & Sakarkar, G. (2024) · Multimedia Tools and Applications · doi:10.1007/s11042-024-19220-w · Cited by 206
Exploring autonomous methods for deepfake detection: A detailed survey on techniques and evaluation
Sunil, R., Mer, P., Diwan, A., Mahadeva, R., & Sharma, A. (2025) · Heliyon · doi:10.1016/j.heliyon.2025.e42273 · Cited by 54
Samesh Enathe, V. P., Chandra Mouli, S., & Dheepthi, R. (2024) · International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS) · doi:10.1109/adics58448.2024.10533511 · Cited by 12
Raza, A., Basit, A., Amin, A., Arfeen, Z. A., Masud, M. I., Fayyaz, U., & Jumani, T. A. (2026) · AI · doi:10.3390/ai7020068 · Cited by 20
Detecting facial manipulated images via one-class domain generalization
Xu, P., Ma, Z., Mei, X., & Shen, J. (2024) · Multimedia Systems · doi:10.1007/s00530-023-01214-7 · Cited by 7
“To further demonstrate the advantage of our method, we compare the accuracy of our method with that of recent works [2, 22, 28, 34, 36, 38, 39, 42] specifically targeting the generalization of facial manipulation detection on four manipulation techniques of FaceForensics++ with three qualities.”
Metadata
- Authors
- Tran, V.-N., Kwon, S.-G., Lee, S.-H., Le, H.-S., & Kwon, K.-R.
- Affiliations
- Pukyong National University (KR)
- Kyungil University (KR)
- Dong-A University (KR)
- Vietnam National University Ho Chi Minh City (VN)
- Venue
- IEEE Access
- Published
- 2023
- Volume
- 11
- Pages
- 535-546
- ISSN
- 2169-3536 (ISSN-L)
- OpenAlex
- W4312909749
- Semantic Scholar
- f03d0fbdb6ec7eca98c7064eb7a33d15e379b2ca
- Cited by
- 47
- References
- 71
- Citations by year
- 2026: 17 · 2025: 12 · 2024: 16 · 2023: 2
- FWCI
- 2.45
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- IEEE Access (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Basic Science Research Program through the National Research Foundation of Korea
- Ministry of Education (2020R1I1A306659411, 2020R1F1A1069124)
- Ministry of Science and ICT (MSIT), South Korea, through the Information Technology Research Center (ITRC) Support Program
- Institute for Information & Communications Technology Planning & Evaluation (IITP-2022-2020-0-01797)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Face recognition and analysis
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
- Keywords
- deepfake detection, face forgery detection, meta-learning, domain generalization, unseen domains, face manipulation, fake information
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 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
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: Generalization of Forgery Detection With Meta Deepfake Detection Model. https://app.cochat.ai/papers/10.1109/access.2022.3232290
- The paper
- Tran, V.-N., Kwon, S.-G., Lee, S.-H., Le, H.-S., & Kwon, K.-R. (2023). Generalization of Forgery Detection With Meta Deepfake Detection Model. IEEE Access. https://doi.org/10.1109/access.2022.3232290
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.