Pith. sign in

REVIEW 4 major objections 6 minor 71 references

Suppressing Gradient Conflict for Generalizable Deepfake Detection

T0 review · 4 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 The paper attributes the '1+1<2' failure of deepfake training to gradient conflicts between original and online-synthesized forgeries, and shows its CS-DFD framework resolves them, reaching 98.88 FF++ AUC and 88.02 average cross-domain AUC.

desk verdict A real problem and a sensible UVS module, but the CGR gradient derivation doesn't minimize its stated loss, so the core mechanism is unproven. read the letter →

arxiv 2507.21530 v1 pith:FURV7EKE submitted 2025-07-29 cs.CV

Ming-Hui Liu , Harry Cheng , Xin Luo , Xin-Shun Xu This is my paper
classification cs.CV
keywords deepfakedetectiongradientconflictcross-domaingeneralizationonlinesynthesizedforgeriesupdatevectorsearchdescentlossrepresentationlearningfaceforgery
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

Deepfake detectors that train on one fixed set of forgeries tend to overfit, so researchers add online-synthesized fake images that carry more generalizable artifacts. The paper documents a surprising failure: when those synthesized forgeries are added on top of the original ones, performance drops on both the source dataset and unseen target datasets, a '1+1<2' effect. It traces the cause to gradient conflicts during backpropagation, where the two forgery types push the model's parameters in opposing directions. To fix this, the proposed CS-DFD framework replaces the conflicting gradient with a searched update vector that lowers both losses simultaneously, and adds a conflict-descent loss that realigns the feature embedding. This yields 98.88 AUC on the source-domain FF++ benchmark and an average cross-domain AUC of 88.02 on four unseen deepfake datasets.

What carries the argument

The mechanism that carries the argument is a two-level gradient correction. At the parameter-update level, UVS solves max min_j g_j^T v subject to ||v - g0|| <= c||g0||, where g_j are the gradients of the two sub-losses and g0 is their sum; a Lagrangian dual and projected-gradient iteration yield the update vector v* = g0 + (sum of mu_j g_j)/(2 $\lambda$). At the representation level, CGR adds a learnable projection layer and a Conflict Descent Loss L_phi = -e1^T e2, the negative dot product of the unit gradients of the two sub-losses; the Hessian of this loss is approximated by a diagonal Fisher-information matrix so that the projection layer can be trained with element-wise operations at O(d) cost. UVS keeps the optimizer moving in a direction that helps both data subsets, while CGR reshapes the feature space so the two subsets stop producing opposing gradients.

What would settle it

Train CS-DFD with the projection layer updated by the exact gradient of the Conflict Descent Loss, computed by automatic differentiation instead of the paper's Equation (24) diagonal approximation, and compare conflict suppression (gradient cosine similarity during training) and cross-domain AUC; if the exact-gradient version is not better, the approximation is not what carries the claims.

Watch

Extended reading notes

Core claim

The paper's central claim is that the '1+1<2' degradation in deepfake detection is an optimization pathology rather than an inherent limit of the data: the gradients of the losses on original forgeries and online-synthesized forgeries conflict, forcing the optimizer to oscillate and settle in a poor region of the loss landscape. The proposed Conflict-Suppressed Deepfake Detection (CS-DFD) framework attacks both sides of this problem. The Update Vector Search (UVS) module constructs a single update vector near the original gradient that maximizes the minimum first-order decrease of the two sub-losses, solving a constrained extremum problem with a Lagrangian dual. The Conflict Gradient Reduction (CGR) module inserts a small learnable projection layer before the classifier and trains it with a Conflict Descent Loss, which is the negative cosine similarity between the two sub-loss gradient vectors, so that the embedding space gradually aligns them. With both modules, the detector reaches 98.88 AUC on FF++ and an average cross-domain AUC of 88.02 on Celeb-DF, DFDC, DFDCp, and UADFV, surpassing methods trained on either forgery type alone.

Load-bearing premise

The conflict-reduction story rests on the assumption that the approximate gradient used to train the projection layer actually lowers the conflict loss without erasing task-specific features; if that approximation fails, the reported gains could come from the projection layer alone.

Editorial extensions

If this is right

  • The source-versus-generalization trade-off that forced prior methods to choose between original forgeries and online-synthesized forgeries is no longer necessary: CS-DFD keeps FF++ accuracy at 98.88 while improving average cross-domain AUC to 88.02.
  • The method is backbone-agnostic: applying CS-DFD to Xception, ViT-B, and ViT-L improves both source accuracy and cross-domain AUC relative to the plain 'original plus synthesized' training strategy on each backbone.
  • The training overhead is small (about 2% extra time), so the gradient-conflict fix can be adopted in settings where compute budgets are fixed.
  • Each module contributes independently: adding UVS alone lifts Celeb-DF AUC from 64.59 to 94.92, and adding CGR alone lifts it to 94.45, so the two fixes can be used separately if only one source of conflict dominates.

Reading between the lines

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

  • Beyond the paper, the gradient-conflict diagnosis should transfer to any training mix that combines a strong in-domain signal with synthetic augmentation, such as classifiers trained with generator-produced negatives, where the synthetic data is useful but its gradients fight the real data.
  • We infer that the benefit of CGR will be largest when the two sources have strongly negative gradient cosine similarity early in training; tracking e1·e2 over time could serve as a cheap predictor of which dataset pairs will benefit from CS-DFD.
  • A testable extension is to apply UVS alone as a drop-in replacement for naive gradient averaging in multi-source or multi-task learning outside forensics, since it requires only two sub-loss gradients and a norm constraint.
  • The paper's framing suggests that prior augmentation studies that reported no gain from adding synthetic data may have been limited by optimization conflicts rather than by the data itself, and re-running those experiments with conflict suppression could revise those negative results.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes CS-DFD, a deepfake-detection framework that trains on a combination of original forgeries and online-synthesized forgeries while attempting to suppress the gradient conflict between these two data sources. The method has two modules: UVS, which solves a constrained optimization problem to replace the total gradient with an update vector that maximizes the minimum first-order descent on the two sub-losses, and CGR, which inserts a learnable projection layer and adds a conflict descent loss based on the negative cosine similarity of the two sub-loss gradients, with the gradient of that loss approximated through a diagonal Fisher-information proxy. Experiments on FF++, Celeb-DF, DFDC, DFDCp, and UADFV report FF++ AUC of 98.88 and an average target-domain AUC of 88.02, which the authors claim is state of the art.

Significance. The problem addressed is important: combining original and synthesized forgery data is a natural way to improve cross-domain generalization, and the paper provides a clear empirical demonstration that naive joint training degrades performance. If the proposed method works as claimed, the UVS module in particular offers a useful and principled way to reconcile conflicting gradients from heterogeneous data within a single task. The manuscript includes a reasonable set of comparisons across multiple backbones and benchmarks, and the '1+1<2' observation is a valuable contribution. However, the significance is weakened by two issues: the CGR module's central gradient derivation is not mathematically sound as written, and the reported results are based on single runs with hyperparameters that appear to be selected using the target-domain test sets. The UVS component appears to be the main driver of the empirical gains in the ablation study, and its derivation is substantially sound, which gives the paper a solid core to build on.

major comments (4)
  1. [§IV-B, Eqs. (19)-(24)] The derivation of the CGR update is not a valid gradient of the stated conflict descent loss Lφ = -e1^T e2. The exact derivative of Lφ with respect to θp contains two groups of terms: -(H1 g2 + H2 g1)/(||g1||||g2||) and a norm-derivative term proportional to (g1^T g2)[H1 g1/||g1||^3 + H2 g2/||g2||^3], where H_i is the Hessian of the sub-loss L_i. Equation (19) keeps only the first group, and Eq. (23) replaces H_i with τ Diag(g_i g_i^T). The discarded norm-derivative terms are not negligible in general. A concrete countercheck: when g1 ≈ g2, we have Lφ ≈ -1 and the exact gradient of Lφ is approximately zero, but Eq. (24) contributes a nonzero conflict term -2γ g⊙g⊙g. Thus Eq. (24) is not a descent direction for Lφ, and the claim that CGR 'gradually maps features into a low-conflict embedding space' is not established. The authors should either derive the exact gradient and justify any approximation, show numerically or analytically that the approximate update has negative inner product with the exact gradient, or explicitly present Eq. (24) as a heuristic and validate its behavior independently. This is load-bearing because the central claim, that representation-level conflict suppression is achieved through Lφ, rests on this derivation.
  2. [§V-D4 and Fig. 7] The hyperparameters c and τ appear to be selected by monitoring AUC on the target-domain test sets. The text in §V-D4 states that 'as c increases, the AUC on the target domain initially improves' and that 'we empirically set c = 0.5 and τ = 0.01 as the default hyperparameter configuration,' but it does not state that the selection was made on a validation split independent of Celeb-DF, DFDC, DFDCp, and UADFV. If the target test sets were used for model selection, the reported cross-domain averages, including the headline 88.02 AUC, are optimistically biased. The manuscript should specify the validation protocol, select hyperparameters on a source-domain validation split or a disjoint validation set, and report results under that fixed configuration.
  3. [Tables I-III] All reported AUC values are single runs with no standard deviations, confidence intervals, or number of seeds. This is a serious limitation for comparative claims: the differences between CS-DFD and the strongest baselines in Table I are often around one AUC point (e.g., 88.02 vs. 87.04 for Effort, or 85.86 vs. 84.56 for FreqBlender on DFDCp), and without variance estimates the claim of state-of-the-art performance is not statistically supported. The authors should report mean and standard deviation over at least three runs, or otherwise justify why the differences are robust.
  4. [§IV-B, Eq. (19)] There is an internal inconsistency in the definition of the Hessian H. The text says 'H is the Hessian matrix, representing the second-order partial derivative of the corresponding loss function Lφ(·)', but the formula in Eq. (19) uses H1 and H2 in products like H1 g2 + H2 g1, which can only be correct if H_i is the Hessian of the sub-loss L_i (i.e., the derivative of g_i) rather than the Hessian of Lφ. This must be corrected, because the subsequent Fisher approximation in Eqs. (21)-(23) is applied to the sub-loss Hessians, not to the Hessian of Lφ.
minor comments (6)
  1. [§IV-A, Eq. (8)] The notation 'max_{j∈[1,2]} Δj = max_{v∈R} ...' is ambiguous; the intended objective is to maximize over v the minimum of the two descent rates, not to maximize over j. Please rewrite as max_v min_j (g_j^T v).
  2. [§IV-B, Eq. (23)] Equation (22) defines G as an outer-product matrix, but Eq. (23) writes H ≈ τ·Diag(G) = τ·g⊗g and states that ⊗ denotes element-wise product. This conflates a diagonal matrix with a vector of its diagonal entries; please clarify the notation, since the subsequent multiplication in Eq. (24) is element-wise vector multiplication.
  3. [§V-E] The efficiency claim of only ~2% additional training time is surprising, because computing g1 and g2 separately for all model parameters requires two backward passes through the backbone per step, whereas the baseline requires one. Please clarify how the individual gradients are computed (e.g., gradient accumulation, separate passes only through the projection layer, or another implementation detail).
  4. [Throughout] There are several typos and formatting issues: 'the uniquely update vector' in §IV-A, 'Harry cheng' with lower-case surname and 'the the School' in the author affiliation block, 'UA VDF' with an extra space in Table I, and 'A VG' as a column header. These should be cleaned up.
  5. [§I and §II-B] The claim of being 'the first to identify' gradient conflicts between heterogeneous forgery data is difficult to verify and is somewhat undercut by the existence of ProDet [10], which also trains on both original and online-synthesized forgeries and explicitly discusses the trade-off. Please soften the novelty claim or clarify the precise difference from [10].
  6. [Fig. 2 and Fig. 5] The loss-trajectory visualizations are described as showing oscillatory behavior and mutual suppression, but the captions and axes are not fully labeled, and the reader cannot independently assess the claimed pattern. Adding axis labels, legends, and explicit marker annotations would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the update-vector and conflict losses are defined from source-domain gradients, and all reported accuracy numbers come from external benchmarks.

full rationale

The paper's central mechanism is defined entirely in terms of gradients of the two source-domain losses L1 and L2 computed on [real, original fakes] and [real, online synthesized fakes]. The UVS update vector v* is derived from g1, g2, and g0 under a constrained first-order objective; the CGR conflict loss L_phi = -e1^T e2 is a function of those gradient vectors, with the Hessian approximated via a diagonal Fisher proxy rather than imported from the target test sets. The reported AUC values on FF++, Celeb-DF, DFDC, DFDCp, and UADFV are external benchmark results, not quantities defined by the optimization objective, so the claimed 'predictions' are not equal to the method's inputs by construction. The only self-citation, reference [44], appears in a related-work enumeration of detection methods and is not load-bearing for Equations (9)-(24) or for the experimental claims. The approximation in Eq. (23) and the dropped norm-derivative terms in Eq. (19) are potential correctness concerns about whether the CGR update actually follows the stated conflict loss, but an incorrect or approximate gradient is not the same as a circular derivation. Similarly, the empirical choice of c and tau is hyperparameter selection, not a fitted parameter renamed as a prediction. No self-definitional step, fitted-input-called-prediction step, or load-bearing self-citation chain is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The UVS module relies on standard convex duality, while the CGR module relies on two unvalidated approximations: the diagonal-FIM Hessian approximation and the omission of norm-derivative terms. The tuned hyperparameters c and tau are the only explicit free parameters, and the online synthesis algorithm is an unreported free choice that affects the central comparison.

free parameters (3)
  • c (UVS neighborhood radius) = 0.5
    Controls how far the replacement update vector v may deviate from the original gradient g0; chosen by sensitivity analysis on target-domain AUC in Fig. 7a.
  • tau (CGR Hessian scale) = 0.01
    Scales the diagonal Fisher approximation of the Hessian in the conflict descent loss; chosen by sensitivity analysis on target-domain AUC in Fig. 7b.
  • Projection layer output dimension = not reported
    The capacity of the low-conflict embedding layer is not specified, and no sensitivity analysis is given for it.
assumptions (4)
  • standard math Strong duality holds for the max-min update-vector problem in Equations 10 to 15.
    Used to convert the primal constrained maximization into the dual minimization over mu in the UVS module.
  • domain assumption Gradient conflict between original and online synthesized forgeries is the primary cause of the 1+1<2 degradation.
    Established only by loss-curve observation in Fig. 2 and not by controlled causal experiments.
  • ad hoc to paper The Hessian of each sub-loss can be approximated by a diagonal Fisher information matrix tau*Diag(g tensor g), and the norm-derivative terms in the gradient of L_phi can be neglected.
    This is the core tractability assumption in CGR, Equations 21 to 24; it is stated but not derived or validated.
  • ad hoc to paper A single learnable projection layer before the classifier can separate a low-conflict embedding space without destroying task-specific features.
    Assumed in the CGR module; no analysis of capacity or interference with the preceding layers is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Suppressing Gradient Conflict for Generalizable Deepfake Detection." pith.science (2026). https://pith.science/paper/FURV7EKE

@misc{pith2026250721530,
  author       = {Pith},
  title        = {Pith review of: Suppressing Gradient Conflict for Generalizable Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FURV7EKE}},
  note         = {Machine review of arXiv:2507.21530}
}
read the original abstract

Robust deepfake detection models must be capable of generalizing to ever-evolving manipulation techniques beyond training data. A promising strategy is to augment the training data with online synthesized fake images containing broadly generalizable artifacts. However, in the context of deepfake detection, it is surprising that jointly training on both original and online synthesized forgeries may result in degraded performance. This contradicts the common belief that incorporating more source-domain data should enhance detection accuracy. Through empirical analysis, we trace this degradation to gradient conflicts during backpropagation which force a trade-off between source domain accuracy and target domain generalization. To overcome this issue, we propose a Conflict-Suppressed Deepfake Detection (CS-DFD) framework that explicitly mitigates the gradient conflict via two synergistic modules. First, an Update Vector Search (UVS) module searches for an alternative update vector near the initial gradient vector to reconcile the disparities of the original and online synthesized forgeries. By further transforming the search process into an extremum optimization problem, UVS yields the uniquely update vector, which maximizes the simultaneous loss reductions for each data type. Second, a Conflict Gradient Reduction (CGR) module enforces a low-conflict feature embedding space through a novel Conflict Descent Loss. This loss penalizes misaligned gradient directions and guides the learning of representations with aligned, non-conflicting gradients. The synergy of UVS and CGR alleviates gradient interference in both parameter optimization and representation learning. Experiments on multiple deepfake benchmarks demonstrate that CS-DFD achieves state-of-the-art performance in both in-domain detection accuracy and cross-domain generalization.

Figures

Figures reproduced from arXiv: 2507.21530 by the authors.

Figure 1
Figure 1. Different training strategies. (a) Traditional training strategy, which [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Gradient conflicts when trained on different types of data. (a) When [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overall architecture of the CS-DFD framework, which consists of two modules: Update Vector Search (UVS) and Conflict Gradient Reduction (CGR). [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Loss trend comparison of different methods when trained simultane [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visualization of vector fields at different pixel locations for selected [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 66 canonical work pages

  1. [1]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in CVPR, 2022, pp. 10 674–10 685

    work page 2022

  2. [2]

    Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models,

    F. Bao, C. Li, J. Zhu, and B. Zhang, “Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models,” in ICLR, 2022, pp. 1–12

    work page 2022

  3. [3]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in ICML, 2015, pp. 2256–2265

    work page 2015

  4. [4]

    Dcface: Synthetic face generation with dual condition diffusion model,

    M. Kim, F. Liu, A. K. Jain, and X. Liu, “Dcface: Synthetic face generation with dual condition diffusion model,” in CVPR, 2023, pp. 12 715–12 725

    work page 2023

  5. [5]

    Diffusion facial forgery detection,

    H. Cheng, Y . Guo, T. Wang, L. Nie, and M. Kankanhalli, “Diffusion facial forgery detection,” in ACM MM, 2024, pp. 5939–5948

    work page 2024

  6. [6]

    Region-aware face swapping,

    C. Xu, J. Zhang, M. Hua, Q. He, Z. Yi, and Y . Liu, “Region-aware face swapping,” in CVPR, 2022, pp. 7622–7631

    work page 2022

  7. [7]

    Diffswap: High- fidelity and controllable face swapping via 3d-aware masked diffusion,

    W. Zhao, Y . Rao, W. Shi, Z. Liu, J. Zhou, and J. Lu, “Diffswap: High- fidelity and controllable face swapping via 3d-aware masked diffusion,” in CVPR, 2023, pp. 8568–8577

    work page 2023

  8. [8]

    Towards General Visual-Linguistic Face Forgery Detection(V2)

    K. Sun, S. Chen, T. Yao, Z. Zhou, J. Ji, X. Sun, C.-W. Lin, and R. Ji, “Towards general visual-linguistic face forgery detection,”arXiv preprint arXiv:2502.20698, pp. 1–10, 2025

    work page Pith review arXiv 2025

Show all 71 references
  1. [9]

    Diffusionface: Towards a comprehensive dataset for diffusion-based face forgery analysis,

    Z. Chen, K. Sun, Z. Zhou, X. Lin, X. Sun, L. Cao, and R. Ji, “Diffusionface: Towards a comprehensive dataset for diffusion-based face forgery analysis,” CoRR, pp. 1–10, 2024

    2024

  2. [10]

    Can we leave deepfake data behind in training deepfake detector?

    J. Cheng, Z. Yan, Y . Zhang, Y . Luo, Z. Wang, and C. Li, “Can we leave deepfake data behind in training deepfake detector?” in NeurIPS, 2024, pp. 1–12

    2024

  3. [11]

    Contrastive learning for deepfake classification and localization via multi-label ranking,

    C. Hong, Y . Hsu, and T. Liu, “Contrastive learning for deepfake classification and localization via multi-label ranking,” in CVPR, 2024, pp. 17 627–17 637

    2024

  4. [12]

    Mmnet: multi- collaboration and multi-supervision network for sequential deepfake detection,

    R. Xia, D. Liu, J. Li, L. Yuan, N. Wang, and X. Gao, “Mmnet: multi- collaboration and multi-supervision network for sequential deepfake detection,” IEEE TIFS, pp. 3409–3422, 2024

    2024

  5. [13]

    Generalizing deepfake video detection with plug-and-play: Video-level blending and spatiotemporal adapter tuning,

    Z. Yan, Y . Zhao, S. Chen, M. Guo, X. Fu, T. Yao, S. Ding, and L. Yuan, “Generalizing deepfake video detection with plug-and-play: Video-level blending and spatiotemporal adapter tuning,” in CVPR, 2025, pp. 12 615–12 625

    2025

  6. [14]

    Rethinking the up-sampling operations in cnn-based generative network for gener- alizable deepfake detection,

    C. Tan, H. Liu, Y . Zhao, S. Wei, G. Gu, P. Liu, and Y . Wei, “Rethinking the up-sampling operations in cnn-based generative network for gener- alizable deepfake detection,” in CVPR, 2024, pp. 28 130–28 139

    2024

  7. [15]

    Contrastive pseudo learning for open-world deepfake attribution,

    Z. Sun, S. Chen, T. Yao, B. Yin, R. Yi, S. Ding, and L. Ma, “Contrastive pseudo learning for open-world deepfake attribution,” inICCV, 2023, pp. 20 882–20 892

    2023

  8. [16]

    Jointly defending deepfake manipulation and adversarial attack using decoy mechanism,

    G. Chen and C. Hsu, “Jointly defending deepfake manipulation and adversarial attack using decoy mechanism,” IEEE TPAMI, pp. 9922– 9931, 2023

    2023

  9. [17]

    Supervised anomaly detection via conditional generative adversarial network and ensemble active learning,

    Z. Chen, J. Duan, L. Kang, and G. Qiu, “Supervised anomaly detection via conditional generative adversarial network and ensemble active learning,” IEEE TPAMI, pp. 7781–7798, 2023

    2023

  10. [18]

    Faceforensics++: Learning to detect manipulated facial images,

    A. R ¨ossler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Nießner, “Faceforensics++: Learning to detect manipulated facial images,” in ICCV, 2019, pp. 1–11

    2019

  11. [19]

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

    Y . Qian, G. Yin, L. Sheng, Z. Chen, and J. Shao, “Thinking in frequency: Face forgery detection by mining frequency-aware clues,” in ECCV, 2020, pp. 86–103

    2020

  12. [20]

    Multi- attentional deepfake detection,

    H. Zhao, W. Zhou, D. Chen, T. Wei, W. Zhang, and N. Yu, “Multi- attentional deepfake detection,” in CVPR, 2021, pp. 2185–2194

    2021

  13. [21]

    Face x-ray for more general face forgery detection,

    L. Li, J. Bao, T. Zhang, H. Yang, D. Chen, F. Wen, and B. Guo, “Face x-ray for more general face forgery detection,” inCVPR, 2020, pp. 5000– 5009

    2020

  14. [22]

    Generalizing face forgery detection with high-frequency features,

    Y . Luo, Y . Zhang, J. Yan, and W. Liu, “Generalizing face forgery detection with high-frequency features,” in CVPR, 2021, pp. 16 317– 16 326

    2021

  15. [23]

    Representative forgery mining for fake face detection,

    C. Wang and W. Deng, “Representative forgery mining for fake face detection,” in CVPR, 2021, pp. 14 923–14 932

    2021

  16. [24]

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

    S. Dong, J. Wang, R. Ji, J. Liang, H. Fan, and Z. Ge, “Implicit identity leakage: The stumbling block to improving deepfake detection generalization,” in CVPR, 2023, pp. 3994–4004

    2023

  17. [25]

    Detecting deepfakes with self-blended images,

    K. Shiohara and T. Yamasaki, “Detecting deepfakes with self-blended images,” in CVPR, 2022, pp. 18 699–18 708

    2022

  18. [26]

    Freqblender: Enhancing deepfake detection by blending frequency knowledge,

    J. Zhou, Y . Li, B. Wu, B. Li, J. Dong et al. , “Freqblender: Enhancing deepfake detection by blending frequency knowledge,” in NeurIPS, 2024, pp. 44 965–44 988

    2024

  19. [27]

    Learning self- consistency for deepfake detection,

    T. Zhao, X. Xu, M. Xu, H. Ding, Y . Xiong, and W. Xia, “Learning self- consistency for deepfake detection,” in ICCV, 2021, pp. 15 023–15 033

    2021

  20. [28]

    Text-guided human image manipulation via image-text shared space,

    X. Xu, Y . Chen, X. Tao, and J. Jia, “Text-guided human image manipulation via image-text shared space,” IEEE TPAMI , pp. 6486– 6500, 2022

    2022

  21. [29]

    Deepfake detection based on discrepancies between faces and their context,

    Y . Nirkin, L. Wolf, Y . Keller, and T. Hassner, “Deepfake detection based on discrepancies between faces and their context,” IEEE TPAMI, pp. 6111–6121, 2022

    2022

  22. [30]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in ICLR, 2014, pp. 1–14

    2014

  23. [31]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. C. Courville, and Y . Bengio, “Generative adversarial nets,” in NeurIPS, 2014, pp. 2672–2680

    2014

  24. [32]

    Face2face: Real-time face capture and reenactment of RGB videos,

    J. Thies, M. Zollh ¨ofer, M. Stamminger, C. Theobalt, and M. Nießner, “Face2face: Real-time face capture and reenactment of RGB videos,” in CVPR, 2016, pp. 2387–2395

    2016

  25. [33]

    A style-based generator architecture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” in CVPR, 2019, pp. 4401–4410

    2019

  26. [34]

    Towards open-set identity preserving face synthesis,

    J. Bao, D. Chen, F. Wen, H. Li, and G. Hua, “Towards open-set identity preserving face synthesis,” in CVPR, 2018, pp. 6713–6722

    2018

  27. [35]

    Exchanging faces in images,

    V . Blanz, K. Scherbaum, T. Vetter, and H. Seidel, “Exchanging faces in images,” Computer Graphics Forum, pp. 669–676, 2004

    2004

  28. [36]

    gddim: Generalized denoising diffusion implicit models,

    Q. Zhang, M. Tao, and Y . Chen, “gddim: Generalized denoising diffusion implicit models,” arXiv preprint arXiv:2206.05564 , 2022

    2022 arXiv

  29. [37]

    Improved denoising diffusion proba- bilistic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion proba- bilistic models,” in ICML, 2021, pp. 8162–8171

    2021

  30. [38]

    BDDM: bilateral denoising diffusion models for fast and high-quality speech synthesis,

    M. W. Y . Lam, J. Wang, D. Su, and D. Yu, “BDDM: bilateral denoising diffusion models for fast and high-quality speech synthesis,” in ICLR, 2022, pp. 1–12

    2022

  31. [39]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in ICLR, 2021, pp. 1–12

    2021

  32. [40]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in NeurIPS, 2020, pp. 1–12

    2020

  33. [41]

    Diffface: Diffusion-based face swapping with facial guidance,

    K. Kim, Y . Kim, S. Cho, J. Seo, J. Nam, K. Lee, S. Kim, and K. Lee, “Diffface: Diffusion-based face swapping with facial guidance,” CoRR, pp. 1–11, 2022

    2022

  34. [42]

    Transcending forgery specificity with latent space augmentation for generalizable deepfake detection,

    Z. Yan, Y . Luo, S. Lyu, Q. Liu, and B. Wu, “Transcending forgery specificity with latent space augmentation for generalizable deepfake detection,” in CVPR, 2024, pp. 8984–8994

    2024

  35. [43]

    Improving generalization of deepfake detectors by imposing gradient regularization,

    W. Guan, W. Wang, J. Dong, and B. Peng, “Improving generalization of deepfake detectors by imposing gradient regularization,” IEEE TIFS, pp. 5345–5356, 2024

    2024

  36. [44]

    Data: Multi- disentanglement based contrastive learning for open-world semi- supervised deepfake attribution,

    M.-H. Liu, X.-Q. Liu, X. Luo, and X.-S. Xu, “Data: Multi- disentanglement based contrastive learning for open-world semi- supervised deepfake attribution,” arXiv preprint arXiv:2505.04384 , 2025

    2025 arXiv

  37. [45]

    Exploring frequency adversarial attacks for face forgery detection,

    S. Jia, C. Ma, T. Yao, B. Yin, S. Ding, and X. Yang, “Exploring frequency adversarial attacks for face forgery detection,” in CVPR, 2022, pp. 4093–4102

    2022

  38. [46]

    Sstnet: Detecting manipulated faces through spatial, steganalysis and temporal features,

    X. Wu, Z. Xie, Y . Gao, and Y . Xiao, “Sstnet: Detecting manipulated faces through spatial, steganalysis and temporal features,” in ICASSP, 2020, pp. 2952–2956

    2020

  39. [47]

    Multi-factor adaptive vision selection for egocentric video question answering,

    H. Zhang, M. Liu, Z. Liu, X. Song, Y . Wang, and L. Nie, “Multi-factor adaptive vision selection for egocentric video question answering,” in ICML, 2024, pp. 59 310–59 328

    2024

  40. [48]

    Efficientnet: Rethinking model scaling for convolutional neural networks,

    M. Tan and Q. V . Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” in ICML, 2019, pp. 6105–6114. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

    2019

  41. [49]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” in ICLR, 2021, pp. 1–12

    2021

  42. [50]

    V oice-face homogeneity tells deepfake,

    H. Cheng, Y . Guo, T. Wang, Q. Li, X. Chang, and L. Nie, “V oice-face homogeneity tells deepfake,” ACM ToMM, pp. 1–22, 2023

    2023

  43. [51]

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

    A. Haliassos, R. Mira, S. Petridis, and M. Pantic, “Leveraging real talking faces via self-supervision for robust forgery detection,” in CVPR, 2022, pp. 14 930–14 942

    2022

  44. [52]

    Emotions don’t lie: An audio-visual deepfake detection method using affective cues,

    T. Mittal, U. Bhattacharya, R. Chandra, A. Bera, and D. Manocha, “Emotions don’t lie: An audio-visual deepfake detection method using affective cues,” in ACM MM, 2020, pp. 2823–2832

    2020

  45. [53]

    Dual contrastive learning for general face forgery detection,

    K. Sun, T. Yao, S. Chen, S. Ding, J. Li, and R. Ji, “Dual contrastive learning for general face forgery detection,” in AAAI, 2022, pp. 2316– 2324

    2022

  46. [54]

    Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection,

    L. Chen, Y . Zhang, Y . Song, L. Liu, and J. Wang, “Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection,” in CVPR, 2022, pp. 18 689–18 698

    2022

  47. [55]

    UCF: uncovering common features for generalizable deepfake detection,

    Z. Yan, Y . Zhang, Y . Fan, and B. Wu, “UCF: uncovering common features for generalizable deepfake detection,” in ICCV, 2023, pp. 22 355–22 366

    2023

  48. [56]

    Towards more general video-based deepfake detection through facial feature guided adaptation for foundation model,

    Y .-H. Han, T.-M. Huang, S.-T. Lo, P.-H. Huang, K.-L. Hua, and J.-C. Chen, “Towards more general video-based deepfake detection through facial feature guided adaptation for foundation model,” in CVPR, 2025, pp. 22 995–23 005

    2025

  49. [57]

    Conflict-averse gradient descent for multi-task learning,

    B. Liu, X. Liu, X. Jin, P. Stone, and Q. Liu, “Conflict-averse gradient descent for multi-task learning,” NeurIPS, pp. 18 878–18 890, 2021

    2021

  50. [58]

    Gradient surgery for multi-task learning,

    T. Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman, and C. Finn, “Gradient surgery for multi-task learning,” in NeurIPS, vol. 33, 2020, pp. 5824–5836

    2020

  51. [59]

    Gradient alignment for cross-domain face anti-spoofing,

    B. M. Le and S. S. Woo, “Gradient alignment for cross-domain face anti-spoofing,” in CVPR, 2024, pp. 188–199

    2024

  52. [60]

    Sharpness-aware gradient matching for domain generalization,

    P. Wang, Z. Zhang, Z. Lei, and L. Zhang, “Sharpness-aware gradient matching for domain generalization,” in CVPR, 2023, pp. 3769–3778

    2023

  53. [61]

    Nasvit: Neural architecture search for efficient vision transformers with gradient conflict-aware supernet training,

    C. Gong and D. Wang, “Nasvit: Neural architecture search for efficient vision transformers with gradient conflict-aware supernet training,” in ICLR, 2022

    2022

  54. [62]

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

    Y . Li, X. Yang, P. Sun, H. Qi, and S. Lyu, “Celeb-df: A large-scale challenging dataset for deepfake forensics,” in CVPR, 2020, pp. 3204– 3213

    2020

  55. [63]

    The deepfake detection challenge dataset,

    B. Dolhansky, J. Bitton, B. Pflaum, J. Lu, R. Howes, M. Wang, and C. Canton-Ferrer, “The deepfake detection challenge dataset,” CoRR, pp. 1–13, 2020

    2020

  56. [64]

    In ictu oculi: Exposing ai generated fake face videos by detecting eye blinking,

    Y . Li, M.-C. Chang, and S. Lyu, “In ictu oculi: Exposing ai generated fake face videos by detecting eye blinking,” in WIFS, 2018, pp. 1–7

    2018

  57. [65]

    CORE: consistent representation learning for face forgery detection,

    Y . Ni, D. Meng, C. Yu, C. Quan, D. Ren, and Y . Zhao, “CORE: consistent representation learning for face forgery detection,” inCVPRW, 2022, pp. 12–21

    2022

  58. [66]

    End-to- end reconstruction-classification learning for face forgery detection,

    J. Cao, C. Ma, T. Yao, S. Chen, S. Ding, and X. Yang, “End-to- end reconstruction-classification learning for face forgery detection,” in CVPR, 2022, pp. 4103–4112

    2022

  59. [67]

    Learning to discover forgery cues for face forgery detection,

    J. Tian, P. Chen, C. Yu, X. Fu, X. Wang, J. Dai, and J. Han, “Learning to discover forgery cues for face forgery detection,” IEEE TIFS , pp. 3814–3828, 2024

    2024

  60. [68]

    Fully unsupervised deepfake video detection via enhanced contrastive learning,

    T. Qiao, S. Xie, Y . Chen, F. Retraint, and X. Luo, “Fully unsupervised deepfake video detection via enhanced contrastive learning,” IEEE TPAMI, pp. 4654–4668, 2024

    2024

  61. [69]

    Exploiting style latent flows for generalizing deepfake video detection,

    J. Choi, T. Kim, Y . Jeong, S. Baek, and J. Choi, “Exploiting style latent flows for generalizing deepfake video detection,” in CVPR, 2024, pp. 1133–1143

    2024

  62. [70]

    Orthogonal subspace decomposition for generalizable ai-generated image detection,

    Z. Yan, J. Wang, P. Jin, K.-Y . Zhang, C. Liu, S. Chen, T. Yao, S. Ding, B. Wu, and L. Yuan, “Orthogonal subspace decomposition for generalizable ai-generated image detection,” in ICML, 2025

    2025

  63. [71]

    Towards general visual-linguistic face forgery detection,

    K. Sun, S. Chen, T. Yao, Z. Zhou, J. Ji, X. Sun, C.-W. Lin, and R. Ji, “Towards general visual-linguistic face forgery detection,” in CVPR, 2025, pp. 19 576–19 586

    2025

Pith tools

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