Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

De-Fake: Style based Anomaly Deepfake Detection

T0 review · 5 major / 6 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 This paper claims that face-swapped deepfakes can be reliably detected by comparing style features of a suspicious image against a genuine reference photo, reporting up to 98% accuracy on its custom dataset.

desk verdict A reference-based identity-consistency deepfake detector whose headline 'privacy-preserving' claim is contradicted by its own dual-input design, with a load-bearing threshold error that undermines the reported numbers. read the letter →

arxiv 2507.03334 v2 pith:PU6SZJU5 submitted 2025-07-04 cs.CV cs.AI

Sudev Kumar Padhi , Harshit Kumar , Umesh Kashyap , Sk. Subidh Ali This is my paper
classification cs.CVcs.AI
keywords deepfakedetectionface-swapstylefeaturesanomalyprivacy-preservingStackedIdentityLossArcFaceautoencoder
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 claims that face-swapped deepfakes can be caught by comparing style features of a suspicious image with a genuine reference photo of the source person, instead of hunting pixel-level artifacts. It proposes SafeVision, which extracts multi-layer style features with pre-trained face-recognition models and either classifies the real/fake pair or scores it as an anomaly via reconstruction error. On its custom dataset built from 18 face-swap techniques, the classifier reaches 98% accuracy and the anomaly detector 96%, with 97% and 93% on eight unseen techniques. The authors present the method as privacy-preserving and as the first deepfake detector built on style features. The method's input, however, includes a real reference image of the source, which Section III.B states explicitly.

What carries the argument

The load-bearing object is the style-feature vector: activations from selected intermediate layers of pre-trained models (VGG-19, ResNet-101, U-Net, LinkNet, ArcFace, VGGFace2, CosFace, FaceNet) stacked per image; ArcFace's embedding layer performs best. For classification, the Stacked Identity Loss (SIL) enforces cosine-similarity structure between real-real and fake-real feature pairs. For anomaly detection, the Anomaly Scoring Mechanism uses a dual-encoder autoencoder whose latent vectors are fused by Hadamard product and decoded into a single reconstruction; the L2 distance between original and reconstructed features forms the anomaly score, with a threshold set at the validation mean plus two standard deviations. The stated rationale comes from neural style transfer: identity plays the role of style, while expression, pose, and lighting play the role of content.

What would settle it

Build a face-swap generator that explicitly matches the source person's style statistics—for example, by penalizing the Gram-matrix distance between the swapped face and the source—and test SafeVision on its output; if these style-preserving swaps are classified as real, the central style-discrepancy assumption is falsified.

Watch

Extended reading notes

Core claim

SafeVision's central claim is that every person's facial images share a distinctive style—stable low- and high-level visual features across lighting, pose, expression, age, and background—and that a face swap breaks this style consistency even when the blend is seamless. The classifier branch (SFE + SFC) concatenates layer-wise features from a pre-trained model for a real-real pair and a fake-real pair, and is trained with binary cross-entropy plus a Stacked Identity Loss that pulls real-real feature pairs together and pushes fake-real pairs apart in cosine similarity. The anomaly branch (AEN + ASM) reconstructs fused style features through an attention-based dual-encoder autoencoder and flags a suspicious image when its L2 reconstruction error exceeds a threshold learned from real images. The paper reports 98% classification and 96% anomaly-detection accuracy on its own dataset, cross-dataset accuracy of 97%/93% on unseen swap techniques, and higher accuracy on its dataset than existing detectors trained on FaceForensics++, Celeb-DF, and DFDC.

Load-bearing premise

The method assumes a genuine, clean reference photo of the source person is available for every suspicious image; when the reference is itself a deepfake, the paper reports accuracy falls to 42%, and in cross-dataset tests the authors substitute celebrity photos rather than true reference images.

Editorial extensions

If this is right

  • If a reference photo of the source person is available, face-swap detection becomes a pairwise feature-comparison task that does not need training data from every new swap technique.
  • The anomaly variant can flag face swaps using only real images for training, which suits settings where labeled fake examples are scarce.
  • The reported 93–97% accuracy on eight unseen swap techniques implies the style-discrepancy signal transfers across generators rather than memorizing one generator's artifacts.
  • Because ArcFace embeddings drive the best results, improvements in face-recognition embeddings should translate directly into better face-swap detection.
  • The method's reliance on a trusted reference makes the reference itself a target: if an attacker supplies a fake reference, detection falls to random-level accuracy (42% in the paper's test).

Reading between the lines

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

  • The paper calls the method privacy-preserving, but its inputs include a reference photograph of the source person; a real deployment would need to store, transmit, or fetch that reference, so the privacy guarantee is only as strong as the handling of the reference images, a point the paper leaves unexamined.
  • An adversarial extension follows immediately from the authors' own failure mode: a face-swap generator trained with a style-consistency loss that matches the source's style features would likely defeat SafeVision, since the detector's signal is exactly the style mismatch.
  • The same style-comparison machinery could be applied to video by using an earlier frame of the same person as the reference, turning each video into a self-referential consistency check without any external photo.
  • The 42% accuracy when the reference is fake suggests the method could be inverted into a reference-verification tool: the same network can tell whether a candidate reference photo is consistent with the person's style distribution.
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

5 major / 6 minor

Summary. The paper proposes SafeVision (also titled De-Fake), a face-swap deepfake detection framework that extracts multi-layer style features from a pre-trained face recognition model for pairs of images. Two detection variants are presented: a supervised classifier (SFC) trained with binary cross-entropy plus a Stacked Identity Loss, and an unsupervised anomaly detector (AEN+ASM) that reconstructs fused style features and thresholds the reconstruction loss. The authors introduce a custom face-swap dataset built from 18 face-swapping methods, evaluate both variants on their own dataset and on FaceForensics++, WildDeepfake, DF-TIMIT, Celeb-DF, DF-Platter, and DF-40, and compare against XceptionNet, SRM, RECCE, SPSL, RFM, MesoNet, EfficientNet, and Face X-Ray. The abstract and introduction claim the method detects face-swaps "without accessing the real facial image" and is privacy-preserving, and the paper reports up to 98% classification and 96% anomaly-detection accuracy.

Significance. If the central claim were supported, a style-based face-swap detector that operates without a reference real image would be a notable contribution, as would a dataset spanning 18 face-swap techniques and the systematic comparison against eight existing detectors. The paper also attempts to validate generalization across multiple public datasets. However, the key claims are not established: the method as formulated requires a genuine reference image, the anomaly threshold is numerically inconsistent, the anomaly score is not formally defined, and the reported generalization numbers rely on a non-independent feature-extractor selection and on substituting arbitrary celebrity images for missing references. The dataset and the comparison table have potential value, but the method's core contribution and its reported performance are not reproducible from the manuscript as written.

major comments (5)
  1. [Abstract; Section III.B; Section V.E] The abstract claims detection works "without accessing the real facial image," but the method requires exactly such an image. Section III.B states that the second input is "a real image of the source individual," and Section III.D.1 reiterates that both the suspicious image and the corresponding real image are passed through the AEN. Section V.E further acknowledges that when the supplied real image is itself a deepfake, accuracy drops to 42%. The privacy-preserving, no-reference claim is therefore contradicted by the method's own formulation and by the paper's stated limitation.
  2. [Section V.B] The anomaly threshold computation is numerically wrong. The text reports validation L2 losses with a mean of 0.5 and standard deviation of 0.05, then states that the threshold is set at the mean plus two standard deviations, "resulting in 0.2." Mean + 2*std = 0.5 + 0.1 = 0.6, not 0.2. Because the reported 96% anomaly-detection accuracy depends on this threshold, the operating point used to produce the headline result is undefined or incorrect, and the experimental results must be re-derived with a correctly computed threshold and an explicit statement of how the threshold was chosen.
  3. [Section V.C] The cross-dataset evaluation does not exercise the method as described. For datasets without a genuine reference image for the source identity, the authors state they "still managed to find the real images by just taking images of celebrities for such datasets." This substitutes an arbitrary celebrity photo for the required real image of the source individual, so the input pair is not the intended real-reference/suspicious-image pair. The reported accuracies on FaceForensics++, WildDeepfake, and Celeb-DF therefore do not validate the method's claimed generalization; they validate a different, much easier or harder, task depending on the identity of the substituted celebrities.
  4. [Section V.A; Table III; Section V.B] The feature extractor is selected using the test set. Table III reports accuracy for eight candidate extractors, and the text concludes "Therefore, we will be using ArcFace as our style feature extractor module." This selection is made on the same test data used for the headline results, so the reported 98% and 96% accuracies are optimistically biased and are not estimates of performance on unseen data. The extractor should be chosen using the validation set, or a nested cross-validation should be reported, before any test-set numbers are claimed.
  5. [Section III.D.2; Eq. (5); Section IV.F] The anomaly detection method is not fully specified. Eq. (5) defines a reconstruction loss, but the actual anomaly score used for decision-making is not given: Section V.B refers to "L2 loss values" without a formula, and Section IV.F mentions an "anomaly score loss" in the training configuration that is never defined. The text also says the decision is based on "the sum of L2 loss between the real-reconstructed and suspicious-reconstructed features," which is vague and not tied to Eq. (5). Without a precise definition of the anomaly score and threshold, the reported anomaly-detection numbers cannot be reproduced or independently verified.
minor comments (6)
  1. [Section IV.C] The dataset split sentence is self-contradictory: "we have used 10 techniques ... for testing and rest of the techniques are used for testing our trained model." Given the next sentence says the test set uses the remaining 8 techniques, the first occurrence of "testing" should presumably be "training." Please correct.
  2. [Section V.B, Section V.D] The reported accuracies are inconsistent across sections: Section V.B reports 98% classification / 96% anomaly detection, Section V.D reports 97.3% / 95.1% for "Ours (Classification)" and "Ours (Anomaly)" on a 10,000-image set, and Table IV reports 98% and 96% for the custom dataset. The manuscript should clarify which evaluation subset, protocol, and table each number refers to.
  3. [Section III.D.2] Eq. (5) is followed by an orphaned fragment: "where freconstructed(xi) is the feature vector reconstructed by the ASM ..." with no corresponding equation. Either restore the missing equation or move this sentence to a place where it follows the defined notation.
  4. [Table I; Section II.B] The reference list and citations contain placeholders and likely errors. Table I lists "REface [8]" but reference [8] is a diffusion-based face-swapping paper (Baliah et al.), not REface. Section II.B cites "deep learning models [28], [ ?]" with an empty placeholder, and Table I has "Custom [?]" entries. These should be resolved before submission.
  5. [Section IV.F.1; Section V.C] There are typographical errors in section headings and text: "Anamoly Feature Extraction" should be "Anomaly Feature Extraction," and "FaceForenisc++" should be "FaceForensics++." Similar typos and inconsistent spacing appear throughout (e.g., "SF E," "F aceswap," "PRNU").
  6. [Section V.A, Table III] The claim that Gram matrices are computed from layer activations is not reflected in the method description: Section III.C.1 and III.D.1 describe concatenation of layer-wise features, but no Gram-matrix operation is defined. Either connect the Gram-matrix statement to the actual implementation or remove it to avoid confusion.

Circularity Check

3 steps flagged · score 6.0 of 10

SafeVision's central 'without accessing the real facial image' claim is contradicted by its own input definition, and its headline accuracy is selected from test-set model comparison.

  1. self definitional [Abstract; Section III.B 'Problem Formulation'; Section V.E 'Discussion and Limitations']
    "Our key idea is to identify these style discrepancies to detect face-swapped images effectively without accessing the real facial image. ... In our proposed method, we use two images as input for our model to identify the face-swapping. The first image is the suspicious image which is doubted to be face-swapped and the second image is a real image of the source individual."

    The advertised property, 'without accessing the real facial image,' is defined in terms of not using the very input that the method's problem formulation requires as its second input. The paper's own limitation statement concedes the dependence: when the supplied reference is itself a deepfake, accuracy drops to 42%. Thus the privacy-preservation claim is not derived from the formulation; it is contradicted by the formulation's input definition.

  2. fitted input called prediction [Section V.A 'Feature Extractor Module' and Table III]
    "Table III shows the performance of both the classification and anomaly detection methods on the extraction of style features using different models. ... As we can see, the feature extracted from the models trained for facial recognition models outperforms the generic image recognition models. Therefore, we will be using ArcFace as our style feature extractor module."

    The feature extractor is selected after inspecting Table III, which reports test-set accuracy for each candidate extractor, and the method's reported accuracy is then taken from the same test set for the selected extractor. Model selection on the test set means the reported 98% classification and 96% anomaly figures are the best of eight candidates rather than unbiased predictions of a single pre-specified pipeline; the 'prediction' is partly constructed by the selection procedure.

1 more flagged steps
  1. other [Section V.C 'Evaluation of Our Models on Other Datasets']
    "Although every dataset didn't have a corresponding real image associated with it, we still managed to find the real images by just taking images of celebrities for such datasets."

    For datasets without a genuine source reference, arbitrary celebrity photos are substituted as the 'real' input. This replaces the specified same-identity comparison with a different, unspecified comparison, so the high cross-dataset accuracies do not evaluate the claimed method; they evaluate a variant whose reference image is not the source identity. The external validation is therefore not independent evidence for the stated style-discrepancy claim.

full rationale

The paper contains no self-citation chain or imported uniqueness theorem; the core classifier training on a custom dataset is a legitimate empirical exercise rather than a circular derivation. However, the central privacy claim reduces by definition: Section III.B requires a real image of the source individual as the second input, while the abstract promises detection 'without accessing the real facial image,' and Section V.E confirms that the method fails when that reference is itself fake (42% accuracy). The headline accuracy is also partially forced by construction: Section V.A and Table III show that ArcFace was chosen as the feature extractor after inspecting its test-set accuracy, so the reported numbers are a post-hoc selection over eight candidates rather than an unbiased estimate for a fixed architecture. The cross-dataset evaluation in Section V.C substitutes arbitrary celebrity photos when no genuine reference exists, so the claimed generalization to established datasets does not exercise the method as specified. The validation-set threshold selection is standard practice, but the stated numbers are internally inconsistent (a mean of 0.5 plus two standard deviations of 0.05 is 0.6, not 0.2), which is a correctness concern rather than circularity. Overall, the evaluation pipeline contains multiple places where the reported 'predictions' are shaped by test-set selection or by redefining the required input, giving a partially circular validation while leaving the underlying supervised training procedure non-circular.

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

The central claim rests on the existence of a stable, extractable per-person style signal and on the availability of a genuine reference image. Neither is independently established outside the paper, and the reference image requirement is contradicted by the abstract's privacy claim.

free parameters (4)
  • Anomaly threshold = 0.2 (inconsistent with stated mean+2*std=0.6)
    Set from the validation loss distribution in Section V.B; used to classify suspicious images as face-swapped.
  • SIL weight alpha = not reported
    Scales Stacked Identity Loss in L_final = L_BCE + alpha * L_SIL, Section III.C.5.
  • Feature extractor layer selection = specific layers per model in Table III
    Chosen based on accuracy on the evaluation set, effectively fitting to test data.
  • Feature extractor model = ArcFace
    Selected because it achieved the highest accuracy in Table III; model selection is performed on the evaluation set.
assumptions (4)
  • domain assumption Each person has a distinctive visual style that remains stable across lighting, pose, expression, age, and background.
    Sections I and III.B; foundational to using style features for identity consistency.
  • domain assumption Face-swapping introduces measurable style discrepancies between the swapped image and the source person's real style.
    Section III.C; the core hypothesis, not derived from theory.
  • domain assumption Intermediate features of pre-trained face recognition models encode these style attributes.
    Section III.C.1; the paper assumes this without verification beyond classification accuracy.
  • ad hoc to paper For evaluation on external datasets, a real image of the source identity is available; when it is not, celebrity photos are substituted.
    Section V.C; this substitution breaks the method's stated input requirement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of De-Fake: Style based Anomaly Deepfake Detection." pith.science (2026). https://pith.science/paper/PU6SZJU5

@misc{pith2026250703334,
  author       = {Pith},
  title        = {Pith review of: De-Fake: Style based Anomaly Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PU6SZJU5}},
  note         = {Machine review of arXiv:2507.03334}
}
read the original abstract

Detecting deepfakes involving face-swaps presents a significant challenge, particularly in real-world scenarios where anyone can perform face-swapping with freely available tools and apps without any technical knowledge. Existing deepfake detection methods rely on facial landmarks or inconsistencies in pixel-level features and often struggle with face-swap deepfakes, where the source face is seamlessly blended into the target image or video. The prevalence of face-swap is evident in everyday life, where it is used to spread false information, damage reputations, manipulate political opinions, create non-consensual intimate deepfakes (NCID), and exploit children by enabling the creation of child sexual abuse material (CSAM). Even prominent public figures are not immune to its impact, with numerous deepfakes of them circulating widely across social media platforms. Another challenge faced by deepfake detection methods is the creation of datasets that encompass a wide range of variations, as training models require substantial amounts of data. This raises privacy concerns, particularly regarding the processing and storage of personal facial data, which could lead to unauthorized access or misuse. Our key idea is to identify these style discrepancies to detect face-swapped images effectively without accessing the real facial image. We perform comprehensive evaluations using multiple datasets and face-swapping methods, which showcases the effectiveness of SafeVision in detecting face-swap deepfakes across diverse scenarios. SafeVision offers a reliable and scalable solution for detecting face-swaps in a privacy preserving manner, making it particularly effective in challenging real-world applications. To the best of our knowledge, SafeVision is the first deepfake detection using style features while providing inherent privacy protection.

Figures

Figures reproduced from arXiv: 2507.03334 by the authors.

Figure 1
Figure 1. Overview of the Proposed Method incongruities [33]. Early methods [6], [21], [61], [66] ap￾proached this problem as a binary classification task, using deep convolutional neural networks (CNNs) to distinguish real images from fake ones. While these methods performed well on controlled datasets, they were heavily dependent on the distribution of training data and lacked a deep understanding of forgery principles [72]… view at source ↗
Figure 2
Figure 2. Overview of the Proposed Method Cross Entropy (BCE) loss and the proposed Stacked Identity Loss (SIL). The final loss function used during training is the weighted sum of these two losses, with the SIL scaled by a coefficient α. 3) Stacked Identity Loss: The Stacked Identity Loss (SIL) is specifically designed to operate across the multi-layer fea￾ture representations extracted by the SF E. It encourages the classif… view at source ↗
Figure 3
Figure 3. Overview of the Proposed Method appearance that are present in real images. The AEN utilizes a pre-trained image recognition model, similar to SF E used in the classifier-based approach, to extract features from various layers of the network. These features are then processed through an unsupervised learning framework, enabling the AEN to learn a representation of what constitutes a “normal” image and to detect devi… view at source ↗

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

84 extracted references · 64 canonical work pages · cited by 1 Pith paper

  1. [1]

    https://research.google/blog/contributing-data-to-deepfake-detection-research/, accessed: 2023-09-30

    Dfd. https://research.google/blog/contributing-data-to-deepfake-detection-research/, accessed: 2023-09-30

    2023

  2. [2]

    https://github.com/deepfakes/faceswap?tab=readme-ov-file, accessed: 2022-01-04

    Faceswap. https://github.com/deepfakes/faceswap?tab=readme-ov-file, accessed: 2022-01-04

    2022

  3. [3]

    https://github.com/haofanwang/inswapper, accessed: 2022- 01-04

    Inswapper. https://github.com/haofanwang/inswapper, accessed: 2022- 01-04

    2022

  4. [4]

    https://github.com/s0md3v/roop, accessed: 2022-01-04

    Roop. https://github.com/s0md3v/roop, accessed: 2022-01-04

    2022

  5. [5]

    https://github.com/C0untFloyd/roop-unleashed?tab= readme-ov-file#usage, accessed: 2022-01-04

    Roop unleashed. https://github.com/C0untFloyd/roop-unleashed?tab= readme-ov-file#usage, accessed: 2022-01-04

    work page 2022

  6. [7]

    In: 2018 IEEE international workshop on information forensics and security (WIFS)

    Afchar, D., Nozick, V ., Yamagishi, J., Echizen, I.: Mesonet: a compact facial video forgery detection network. In: 2018 IEEE international workshop on information forensics and security (WIFS). pp. 1–7. IEEE (2018)

    work page 2018

  7. [8]

    arXiv preprint arXiv:2409.07269 (2024)

    Baliah, S., Lin, Q., Liao, S., Liang, X., Khan, M.H.: Realistic and efficient face swapping: A unified approach with diffusion models. arXiv preprint arXiv:2409.07269 (2024)

    arXiv 2024

  8. [9]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Bao, J., Chen, D., Wen, F., Li, H., Hua, G.: Towards open-set identity preserving face synthesis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6713–6722 (2018)

    work page 2018

Show all 84 references
  1. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cao, J., Ma, C., Yao, T., Chen, S., Ding, S., Yang, X.: End-to- end reconstruction-classification learning for face forgery detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4113–4122 (2022)

    2022

  2. [11]

    In: Proceedings of the International Conference on Automatic Face and Gesture Recognition (FG)

    Cao, Q., Shen, L., Xie, W., Zu, X.: Vggface2: A dataset for recognising faces with large pose variations. In: Proceedings of the International Conference on Automatic Face and Gesture Recognition (FG). pp. 534– 542 (2018)

    2018

  3. [12]

    arXiv preprint arXiv:1707.03718 (2017)

    Chaurasia, A., Culurciello, E.: Linknet: Exploiting encoder rep- resentations for efficient semantic segmentation. arXiv preprint arXiv:1707.03718 (2017)

    2017 arXiv

  4. [13]

    In: 2021 IEEE International conference on Multimedia and Expo (ICME)

    Chen, H.S., Rouhsedaghat, M., Ghani, H., Hu, S., You, S., Kuo, C.C.J.: Defakehop: A light-weight high-performance deepfake detector. In: 2021 IEEE International conference on Multimedia and Expo (ICME). pp. 1–

    2021

  5. [14]

    In: Proceedings of the 28th ACM international conference on multimedia

    Chen, R., Chen, X., Ni, B., Ge, Y .: Simswap: An efficient framework for high fidelity face swapping. In: Proceedings of the 28th ACM international conference on multimedia. pp. 2003–2011 (2020)

    2020

  6. [15]

    In: MM ’20: The 28th ACM International Conference on Multimedia (2020)

    Chen, R., Chen, X., Ni, B., Ge, Y .: Simswap: An efficient framework for high fidelity face swapping. In: MM ’20: The 28th ACM International Conference on Multimedia (2020)

    2020

  7. [16]

    In: Proceedings of the AAAI conference on artificial intelligence

    Chen, S., Yao, T., Chen, Y ., Ding, S., Li, J., Ji, R.: Local relation learning for face forgery detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1081–1088 (2021)

    2021

  8. [17]

    Proceedings of CVPR (2022)

    Chen, Z., Others: Mmnet: Multi-collaboration framework for robust deepfake detection. Proceedings of CVPR (2022)

    2022

  9. [18]

    IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR) (2018)

    Cheng, X., Others: Xceptionnet: Deep learning for facial manipulation detection. IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR) (2018)

    2018

  10. [19]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Choi, Y ., Uh, Y ., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188–8197 (2020)

    2020

  11. [20]

    arXiv preprint arXiv:1806.05622 (2018)

    Chung, J.S., Nagrani, A., Zisserman, A.: V oxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622 (2018)

    2018 arXiv

  12. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition

    Dang, H., Liu, F., Stehouwer, J., Liu, X., Jain, A.K.: On the detection of digital face manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition. pp. 5781–5790 (2020)

    2020

  13. [23]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Das, S., Seferbekov, S., Datta, A., Islam, M.S., Amin, M.R.: Towards solving the deepfake problem: An analysis on improving deepfake detection using dynamic face augmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3776– 3785 (2021)

    2021

  14. [24]

    https://github.com/iperov/ DeepFaceLab (2020)

    DeepFaceLab: Deepfacelab github repository. https://github.com/iperov/ DeepFaceLab (2020)

    2020

  15. [25]

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

    Deng, J., Guo, J., Niannan, X., Dong, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4690–4699 (2019)

    2019

  16. [26]

    arXiv preprint arXiv:2006.07397 (2020)

    Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., Ferrer, C.C.: The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397 (2020)

    2020 arXiv

  17. [27]

    arXiv preprint arXiv:2210.14457 10 (2022)

    Dong, S., Wang, J., Ji, R., Liang, J., Fan, H., Ge, Z.: Towards a robust deepfake detector: Common artifact deepfake detection model. arXiv preprint arXiv:2210.14457 10 (2022)

    2022 arXiv

  18. [28]

    International Conference on Learning Representations (ICLR) (2021)

    Dosovitskiy, A., Others: An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR) (2021)

    2021

  19. [29]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Gao, G., Huang, H., Fu, C., Li, Z., He, R.: Information bottleneck disentanglement for identity swapping. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3404–3413 (2021)

    2021

  20. [30]

    arXiv preprint arXiv:1508.06576 (2015)

    Gatys, L.A.: A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576 (2015)

    2015 arXiv

  21. [31]

    IEEE Access 10, 83452–83462 (2022)

    Groshev, A., Maltseva, A., Chesakov, D., Kuznetsov, A., Dimitrov, D.: Ghost—a new face swap approach for image and video domains. IEEE Access 10, 83452–83462 (2022). https://doi.org/10.1109/ACCESS.2022.3196668 12

    2022

  22. [32]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Gu, Q., Chen, S., Yao, T., Chen, Y ., Ding, S., Yi, R.: Exploiting fine- grained face forgery clues via progressive enhancement learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 735–743 (2022)

    2022

  23. [33]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Haliassos, A., V ougioukas, K., Petridis, S., Pantic, M.: Lips don’t lie: A generalisable and robust approach to face forgery detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5039–5049 (2021)

    2021

  24. [34]

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

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–778 (2016)

    2016

  25. [35]

    IEEE Transactions on Pattern Analysis and Machine Intelli- gence (2022)

    Hou, Y ., Others: Mstatattack: Statistical attack for deepfake detection systems. IEEE Transactions on Pattern Analysis and Machine Intelli- gence (2022)

    2022

  26. [36]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Huang, B., Wang, Z., Yang, J., Ai, J., Zou, Q., Wang, Q., Ye, D.: Implicit identity driven deepfake face swapping detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4490–4499 (2023)

    2023

  27. [37]

    In: Proceedings of the IEEE international conference on computer vision

    Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adap- tive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501–1510 (2017)

    2017

  28. [38]

    IEEE Transactions on Information Forensics and Security (2020)

    Ismail, A., Others: Deepfake detection using xgboost to exploit visual artifacts. IEEE Transactions on Information Forensics and Security (2020)

    2020

  29. [39]

    In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14

    Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694–711. Springer (2016)

    2016

  30. [40]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110–8119 (2020)

    2020

  31. [41]

    Pattern Recognition Letters (2019)

    Kharbat, F., Others: Svm-based deepfake detection using hog, orb, and surf features. Pattern Recognition Letters (2019)

    2019

  32. [42]

    Proceedings of ACM MM (2020)

    Koopman, J., Others: Exploiting prnu for deepfake video detection. Proceedings of ACM MM (2020)

    2020

  33. [43]

    arXiv preprint arXiv:1812.08685 (2018)

    Korshunov, P., Marcel, S.: Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685 (2018)

    2018 arXiv

  34. [44]

    arXiv preprint arXiv:1912.13457 (2019)

    Li, L., Bao, J., Yang, H., Chen, D., Wen, F.: Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457 (2019)

    2019 arXiv

  35. [46]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, L., Bao, J., Zhang, T., Yang, H., Chen, D., Wen, F., Guo, B.: Face x-ray for more general face forgery detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5001–5010 (2020)

    2020

  36. [47]

    Proceedings of AAAI Conference on Artificial Intelligence (2022)

    Li, M., Others: Deepfake catcher: A simple fusion model for robust detection. Proceedings of AAAI Conference on Artificial Intelligence (2022)

    2022

  37. [48]

    IEEE Transactions on Image Processing (2023)

    Li, W., Others: Style latent flow modeling for generalized deepfake detection. IEEE Transactions on Image Processing (2023)

    2023

  38. [49]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, Y ., Yang, X., Sun, P., Qi, H., Lyu, S.: Celeb-df: A large-scale chal- lenging dataset for deepfake forensics. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3207–3216 (2020)

    2020

  39. [51]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, H., Li, X., Zhou, W., Chen, Y ., He, Y ., Xue, H., Zhang, W., Yu, N.: Spatial-phase shallow learning: rethinking face forgery detection in frequency domain. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 772–781 (2021)

    2021

  40. [52]

    IEEE Transactions on Multimedia (2021)

    Liu, K., Others: Variational autoencoder multi-layer perceptron (va-mlp) for deepfake detection. IEEE Transactions on Multimedia (2021)

    2021

  41. [53]

    arXiv preprint arXiv:2211.14068 (2022)

    Liu, Z., Li, M., Zhang, Y ., Wang, C., Zhang, Q., Wang, J., Nie, Y .: Fine-grained face swapping via regional gan inversion. arXiv preprint arXiv:2211.14068 (2022)

    2022 arXiv

  42. [54]

    In: Proceedings of the Asian Conference on Computer Vision

    Liu, Z., Wang, H., Wang, S.: Cross-domain local characteristic enhanced deepfake video detection. In: Proceedings of the Asian Conference on Computer Vision. pp. 3412–3429 (2022)

    2022

  43. [55]

    In: Proceedings of International Conference on Computer Vision (ICCV) (December 2015)

    Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of International Conference on Computer Vision (ICCV) (December 2015)

    2015

  44. [56]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Luo, Y ., Zhang, Y ., Yan, J., Liu, W.: Generalizing face forgery detection with high-frequency features. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 16317–16326 (2021)

    2021

  45. [57]

    Proceedings of CVPR Workshops (2020)

    McCloskey, S., Albright, J.: Detecting gan images using saturation cues. Proceedings of CVPR Workshops (2020)

    2020

  46. [58]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Narayan, K., Agarwal, H., Thakral, K., Mittal, S., Vatsa, M., Singh, R.: Df-platter: Multi-face heterogeneous deepfake dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 9739–9748 (2023)

    2023

  47. [59]

    arXiv preprint arXiv:1804.03447 (2018)

    Natsume, R., Yatagawa, T., Morishima, S.: Rsgan: face swapping and editing using face and hair representation in latent spaces. arXiv preprint arXiv:1804.03447 (2018)

    2018 arXiv

  48. [60]

    In: Computer Vision– ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Aus- tralia, December 2–6, 2018, Revised Selected Papers, Part VI 14

    Natsume, R., Yatagawa, T., Morishima, S.: Fsnet: An identity-aware generative model for image-based face swapping. In: Computer Vision– ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Aus- tralia, December 2–6, 2018, Revised Selected Papers, Part VI 14. pp. 117–132...

    2019

  49. [61]

    In: ICASSP 2019- 2019 IEEE international conference on acoustics, speech and signal processing (ICASSP)

    Nguyen, H.H., Yamagishi, J., Echizen, I.: Capsule-forensics: Using capsule networks to detect forged images and videos. In: ICASSP 2019- 2019 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 2307–2311. IEEE (2019)

    2019

  50. [62]

    In: Proceedings of the IEEE International Conference on Computer Vision

    Nirkin, Y ., Keller, Y ., Hassner, T.: FSGAN: Subject agnostic face swapping and reenactment. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 7184–7193 (2019)

    2019

  51. [63]

    In: European conference on computer vision

    Qian, Y ., Yin, G., Sheng, L., Chen, Z., Shao, J.: Thinking in frequency: Face forgery detection by mining frequency-aware clues. In: European conference on computer vision. pp. 86–103. Springer (2020)

    2020

  52. [64]

    In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted In- tervention (MICCAI)

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted In- tervention (MICCAI). pp. 234–241 (2015)

    2015

  53. [65]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV)

    Rosberg, F., Aksoy, E.E., Alonso-Fernandez, F., Englund, C.: Facedancer: Pose- and occlusion-aware high fidelity face swapping. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). pp. 3454–3463 (January 2023)

    2023

  54. [66]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M.: Faceforensics++: Learning to detect manipulated facial images. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1–11 (2019)

    2019

  55. [67]

    Advances in Neural Information Processing Systems (NeurIPS) (2017)

    Sabour, S., Frosst, N., Hinton, G.E.: Dynamic routing between capsules. Advances in Neural Information Processing Systems (NeurIPS) (2017)

    2017

  56. [68]

    In: Proceedings of the IEEE confer- ence on computer vision and pattern recognition

    Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE confer- ence on computer vision and pattern recognition. pp. 815–823 (2015)

    2015

  57. [69]

    In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV)

    Shiohara, K., Yang, X., Taketomi, T.: Blendface: Re-designing identity encoders for face-swapping. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV). pp. 7634–7644 (October 2023)

    2023

  58. [70]

    In: Proceedings of the International Conference on Machine Learning (ICML) (2014)

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: Proceedings of the International Conference on Machine Learning (ICML) (2014)

    2014

  59. [71]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., Nießner, M.: Face2face: Real-time face capture and reenactment of rgb videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2387–2395 (2016)

    2016

  60. [73]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, C., Deng, W.: Representative forgery mining for fake face detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14923–14932 (2021)

    2021

  61. [74]

    arXiv preprint arXiv:2403.01108 (2024)

    Wang, F.: Face swap via diffusion model. arXiv preprint arXiv:2403.01108 (2024)

    2024 arXiv

  62. [75]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Wang, H., Wang, Y ., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265–5274 (2018)

    2018

  63. [76]

    arXiv preprint arXiv:2301.06789 (2023)

    Wang, Y ., Others: Mintime: Multi-supervision for generalizable deepfake detection. arXiv preprint arXiv:2301.06789 (2023)

    2023 arXiv

  64. [77]

    CoRR abs/2106.09965 (2021)

    Wang, Y ., Chen, X., Zhu, J., Chu, W., Tai, Y ., Wang, C., Li, J., Wu, Y ., Huang, F., Ji, R.: Hififace: 3d shape and semantic prior guided high fidelity face swapping. CoRR abs/2106.09965 (2021)

    2021 arXiv

  65. [78]

    In: Proceedings of the AAAI conference on artificial intelligence

    Wei, J., Wang, S., Huang, Q.: F 3net: fusion, feedback and focus for salient object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 12321–12328 (2020) 13

    2020

  66. [79]

    In: European conference on computer vision

    Xu, C., Zhang, J., Han, Y ., Tian, G., Zeng, X., Tai, Y ., Wang, Y ., Wang, C., Liu, Y .: Designing one unified framework for high-fidelity face reenactment and swapping. In: European conference on computer vision. pp. 54–71. Springer (2022)

    2022

  67. [80]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (2022)

    Xu, Z., Hong, Z., Ding, C., Zhu, Z., Han, J., Liu, J., Ding, E.: Mobilefaceswap: A lightweight framework for video face swapping. In: Proceedings of the AAAI Conference on Artificial Intelligence (2022)

    2022

  68. [81]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)

    Xu, Z., Zhou, H., Hong, Z., Liu, Z., Liu, J., Guo, Z., Han, J., Liu, J., Ding, E., Wang, J.: Styleswap: Style-based generator empowers robust face swapping. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)

    2022

  69. [82]

    arXiv preprint arXiv:2406.13495 (2024)

    Yan, Z., Yao, T., Chen, S., Zhao, Y ., Fu, X., Zhu, J., Luo, D., Wang, C., Ding, S., Wu, Y ., et al.: Df40: Toward next-generation deepfake detection. arXiv preprint arXiv:2406.13495 (2024)

    2024 arXiv

  70. [83]

    In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Yang, X., Li, Y ., Lyu, S.: Exposing deep fakes using inconsistent head poses. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 8261–8265. IEEE (2019)

    2019

  71. [84]

    Proceedings of NeurIPS (2022)

    Zhang, H., Others: Rethinking invariant risk minimization for general- izable deepfake detection. Proceedings of NeurIPS (2022)

    2022

  72. [85]

    IEEE signal processing letters 23(10), 1499–1503 (2016)

    Zhang, K., Zhang, Z., Li, Z., Qiao, Y .: Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters 23(10), 1499–1503 (2016)

    2016

  73. [86]

    Proceedings of ICCV (2023)

    Zhao, J., Others: Deepfake video detection using facial feature points and ch-transformer. Proceedings of ICCV (2023)

    2023

  74. [87]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhao, W., Rao, Y ., Shi, W., Liu, Z., Zhou, J., Lu, J.: Diffswap: High- fidelity and controllable face swapping via 3d-aware masked diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8568–8577 (2023)

    2023

  75. [88]

    arXiv preprint arXiv:2304.02934 (2023)

    Zhou, L., Others: Latent space augmentation for robust deepfake detec- tion. arXiv preprint arXiv:2304.02934 (2023)

    2023 arXiv

  76. [89]

    In: Proceedings of the 28th ACM international conference on multimedia

    Zi, B., Chang, M., Chen, J., Ma, X., Jiang, Y .G.: Wilddeepfake: A challenging real-world dataset for deepfake detection. In: Proceedings of the 28th ACM international conference on multimedia. pp. 2382–2390 (2020)

    2020

Pith tools

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