Pith. sign in

REVIEW 3 major objections 4 minor 76 references

Fair Deepfake Detectors Can Generalize

T0 review · 3 major / 4 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 making deepfake detectors fairer also makes them generalize better to unseen forgeries.

desk verdict Solid fairness-aware detector with a causal story that does not survive contact with its own implementation. read the letter →

arxiv 2507.02645 v1 pith:BI22RYBT submitted 2025-07-03 cs.LG cs.CV

Harry Cheng , Ming-Hui Liu , Yangyang Guo , Tianyi Wang , Liqiang Nie , Mohan Kankanhalli This is my paper · ORCID
classification cs.LGcs.CV
keywords deepfakedetectiondemographicfairnesscausalinferenceback-dooradjustmentcross-domaingeneralizationinversepropensityweightingfeaturenormalizationfairness-generalizationtrade-off
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

The paper claims that demographic fairness and cross-domain generalization are not competing objectives in deepfake detection: improving fairness causally improves how well a detector recognizes unseen manipulation methods. To support that claim, it builds a causal graph in which fairness is a treatment and the confounders are data distribution and model capacity, then applies the back-door adjustment from causal inference to estimate an average causal effect of 2.35 AUC percentage points. It then introduces DAID, a plug-and-play training framework that reweights training samples by demographic propensity, normalizes features within demographic subgroups, and aligns same-label features across groups. If the causal account is right, fairness interventions become a direct route to more robust detectors rather than an ethical constraint that costs accuracy.

What carries the argument

The load-bearing object is the back-door adjustment formula $P(A\mid do(F=f)) = \sum_{dd,mc} P(A\mid F=f,DD=dd,MC=mc)P(DD=dd,MC=mc)$, applied to a causal graph where fairness $F$ is the treatment and accuracy $A$ is the outcome, with data distribution $DD$ and model capacity $MC$ as confounders. The accompanying method DAID implements the controlled intervention in training: inverse-propensity weighting and subgroup-wise feature normalization neutralize $DD$, and a cosine alignment loss on a low-rank orthonormal projection suppresses demographic signals to neutralize $MC$. This machinery converts a fairness intervention into a generalization signal.

What would settle it

Train two detectors from the same backbone on identical training data with identical demographic composition, differing only in a fairness regularizer that does not resample the data; if the more fair detector does not beat the less fair one on held-out cross-domain benchmarks such as Celeb-DF, the claimed causal effect would be contradicted.

Watch

Extended reading notes

Core claim

The paper's central discovery is the claim that the observed association between fairness and generalization is confounded, and once data distribution and model capacity are controlled, higher demographic fairness causes higher cross-domain accuracy. The evidence is a stratified experiment with model capacity varied between Xception and EfficientNet and data distribution stratified by six gender-by-race subgroups; averaging the fairness effect over those strata gives an average causal effect of 2.35 percentage points with a 95% confidence interval [0.0186, 0.0280] and p < 0.001. The authors then translate the insight into a training method, DAID, whose demographic-aware rebalancing and demographic-agnostic feature aggregation realize the fairness intervention, and report simultaneous gains in fairness (lower Skew) and generalization (higher AUC) across DFDC, DFD, and Celeb-DF.

Load-bearing premise

The causal graph in the paper is correct, meaning there are no unobserved confounders and the back-door adjustment is valid; in particular, the intervention that raises fairness must not itself alter the data-distribution confounder.

Editorial extensions

If this is right

  • Fairness-aware training can be used as a deliberate generalization strategy for deepfake detectors, not just a compliance measure.
  • Plug-and-play DAID improves both Skew and AUC on DFDC, DFD, and Celeb-DF across four backbones (Xception, EfficientNet, F3-Net, CADDM) without changing the inference architecture.
  • Data-level confounder control (rebalancing and normalization) contributes the largest gains: removing it produces the biggest performance drop in the ablation table.
  • The training-time overhead of the fairness intervention is under 5% (243 vs 233 minutes on EfficientNet), so the approach is practical at scale.
  • Controlling both data distribution and model capacity is necessary: the ablation shows each module alone gives smaller gains than the combined framework.

Reading between the lines

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

  • If fairness causally drives generalization, fairness metrics could serve as a cheap proxy for estimating how well a deepfake detector will perform on an unlabeled target domain during development.
  • A direct testable extension is to apply DAID to other demographic attribute sets or to audio-only and multimodal deepfake detectors, where the same confounding structure may hold.
  • The paper estimates the causal effect under a resampling intervention that changes the training data distribution; a stronger causal test would hold the data distribution fixed and intervene only on the feature-level fairness loss.
  • The causal graph could be extended with a manipulation-type node; identifying how fairness mediates the path to unseen forgeries would clarify when the result transfers to entirely new synthesis techniques.
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

3 major / 4 minor

Summary. The manuscript claims to be the first to establish that improving demographic fairness causally improves cross-domain generalization in deepfake detection. It models fairness as a treatment F, generalization as outcome A, and data distribution DD and model capacity MC as confounders in a DAG; applies back-door adjustment (Eqs. 2-3); reports an ACE of 2.35 AUC points; and proposes DAID, which combines inverse-propensity reweighting, subgroup-wise feature normalization, and demographic-agnostic feature alignment. Experiments on FF++-trained models evaluated on DFDC, DFD, and Celeb-DF show simultaneous improvements in Skew and AUC.

Significance. If the causal claim were established, it would overturn the common fairness-generalization trade-off view and provide a principled design principle. The empirical DAID results across three benchmarks and four backbones are encouraging, and the paper provides a clear method description with extensive ablations. However, the causal identification is not sound as presented; the headline claim is therefore not supported. The empirical framework may still be useful, but the contribution would need to be reframed without the current causal language.

major comments (3)
  1. [Section 3.2, Eq. (2)-(3), Fig. 1b] The back-door adjustment is misapplied because the intervention used to define do(F=1) is resampling the training data, which changes the empirical distribution of sensitive attributes—i.e., the confounder DD itself. The formula requires observing P(A | F=f, DD=dd, MC=mc) with DD fixed across treatment levels; the two training regimes instead differ in DD, so the contrast is a descriptive difference between training recipes, not a causal effect after confounding removal. The reported ACE=2.35pp therefore cannot support the claimed causal relationship.
  2. [Section 3.2, ACE Estimation Results] The bootstrap CI and p-value do not license the causal conclusion. With B=1000 bootstrap resamples of the evaluation set, the uncertainty reflects test-set sampling for the already-trained models; it does not cover variation across training runs, random seeds, or a broader set of architectures. Since MC has only two levels and no replication is reported, the ACE estimate has no valid inferential basis for the population of deepfake detectors.
  3. [Section 3.3, Eq. (5), Table 1] DAID's demographic-aware data rebalancing uses the same inverse-propensity reweighting mechanism that defines the high-fairness intervention in the ACE estimation. Consequently, the strong empirical performance of DAID cannot serve as independent validation of the causal hypothesis; it is equally consistent with the weaker claim that reweighting the training distribution improves cross-domain AUC. The manuscript should either remove the causal validation language or provide a test whose intervention mechanism is distinct from the method being validated.
minor comments (4)
  1. [Section 3.3 and 4.4.2] "It worth noting" should read "It is worth noting."
  2. [Section 4.3] "several the state-of-the-art (SoTA) approaches" should read "several state-of-the-art approaches."
  3. [Eq. (8)-(9)] The loss Lcos is defined on h in Eq. (9) but applied to \hat h in Eq. (8); the notation should be harmonized, and the role of the epsilon term in Eq. (9) should be explained.
  4. [Section 3.2] The main text says details of the ACE estimation are in the supplementary materials, but the number of training runs, random seeds, and the exact computation of P(A | F=f, DD=dd, MC=mc) should be stated in the main text for reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

The causal ACE is not identified because do(F=1) is implemented by resampling, which changes the confounder DD by construction; DAID then reuses the same inverse-propensity weighting, so its benchmark gains do not independently validate the causal claim.

  1. self definitional [Section 3.2, 'Fairness Intervention (do(F))', and Eq. (2)-(3); DD defined in Section 3.1.]
    "DD captures the distribution of sensitive attributes (e.g., race, gender), while MC denotes the model’s architectural capacity. ... High fairness (F = 1): Cross-entropy loss with a simple resampling strategy Cheng et al. (2024a), where each sample in the cross-entropy loss is assigned a weight to suppress the over-representation of majority groups."

    In the DAG, DD is the confounder that the back-door adjustment must hold fixed. But the do(F=1) regime is implemented by resampling the training data to suppress majority over-representation, which directly changes the empirical distribution of sensitive attributes, i.e., the variable DD. The two training regimes therefore differ in DD by construction, so Eq. (2) cannot recover P(A|do(F=f)); the ACE in Eq. (3) is a weighted difference of AUCs between training sets with different DD, not the interventional contrast with confounders controlled. The claimed causal effect reduces to an observed, confounded comparison.

  2. fitted input called prediction [Section 3.3, 'Demographic-aware Data Rebalancing', Eq. (5), vs. Section 3.2 'Fairness Intervention'.]
    "Motivated by our causal findings, we conclude that, as long as confounders are properly controlled, the clear causal pathway can be leveraged to enhance generalization by intervening on more readily measurable fairness. Therefore, we introduce Demographic Attribute-Insensitive Intervention Detection (DAID) ... To equalize the influence of majority and minority groups, we compute a sample-specific importance weight: w_i = \left( \prod_{k=1}^{K} \hat{P}(s_i^{(k)}) \right)^{-1}."

    DAID is presented as validation of the causal theory, but its demographic-aware data-rebalancing module is the same inverse-propensity sample weighting used to implement do(F=1) in the ACE estimate ('each sample in the cross-entropy loss is assigned a weight to suppress the over-representation of majority groups'). Thus DAID's benchmark improvements are not an independent test of the claim that fairness improves generalization: the method embeds the very intervention from which the ACE was estimated, so the empirical support is a consistency check on the training-regime contrast, not an out-of-sample prediction.

full rationale

The paper's central causal claim rests on the back-door adjustment in Eq. (2)-(3). The high-fairness treatment is operationalized by resampling the training data, which changes the confounder DD (the distribution of sensitive attributes) by definition. Consequently the ACE estimate does not identify P(A|do(F=1)) - P(A|do(F=0)); it is a weighted comparison of models trained under different data distributions, so the causal conclusion is structurally entangled with its own intervention. The proposed DAID framework then reuses the same inverse-propensity weighting as its first module, meaning the later benchmark improvements cannot independently validate the causal hypothesis. These issues are specific and quotable, not a matter of missing consensus. The paper still contains useful empirical comparisons against external baselines and DAID may be practically effective, but the headline 'first causal relationship' is not established by the paper's own derivation chain. Score 6 reflects one or more predictions that reduce by construction while leaving substantial independent experimental content.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The causal analysis relies on an assumed DAG and a back-door adjustment that is misapplied because the intervention changes the confounder DD. The method introduces two tuned hyperparameters. No new physical or conceptual entities are postulated.

free parameters (2)
  • lambda_attr = 0.7
    Selected based on hyperparameter sensitivity analysis (Figure 3) on validation performance; this tuning is a free choice that affects the final results.
  • lambda_ortho = 0.2
    Selected together with lambda_attr based on empirical observations; controls the strength of the orthogonality regularization.
assumptions (5)
  • domain assumption Causal graph in Figure 1b is correct: fairness (F) directly causes generalization (A), with data distribution (DD) and model capacity (MC) as confounders.
    The DAG is assumed without testing alternative structures or directions; it is introduced in Section 3.1.
  • domain assumption Back-door adjustment with Z={DD, MC} is valid for estimating the causal effect of F on A.
    Requires Z to be pre-treatment and unaffected by the intervention; Section 3.2. This assumption is violated because the fairness intervention changes DD.
  • ad hoc to paper The high-fairness training regime (resampling) changes F without changing the confounder DD.
    Resampling directly alters the demographic composition of the training data, so DD is affected. This is the key invalidating assumption for the causal analysis.
  • domain assumption Empirical frequency of (DD, MC) in the held-out test set approximates P(DD, MC).
    Stated in the footnote of Section 3.2; assumes the test set is an i.i.d. sample from the deployment population.
  • ad hoc to paper The demographic-agnostic feature aggregation module controls the MC confounder.
    The link between suppressing demographic features and controlling model capacity is not formally established; Section 3.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fair Deepfake Detectors Can Generalize." pith.science (2026). https://pith.science/paper/BI22RYBT

@misc{pith2026250702645,
  author       = {Pith},
  title        = {Pith review of: Fair Deepfake Detectors Can Generalize},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BI22RYBT}},
  note         = {Machine review of arXiv:2507.02645}
}
read the original abstract

Deepfake detection models face two critical challenges: generalization to unseen manipulations and demographic fairness among population groups. However, existing approaches often demonstrate that these two objectives are inherently conflicting, revealing a trade-off between them. In this paper, we, for the first time, uncover and formally define a causal relationship between fairness and generalization. Building on the back-door adjustment, we show that controlling for confounders (data distribution and model capacity) enables improved generalization via fairness interventions. Motivated by this insight, we propose Demographic Attribute-insensitive Intervention Detection (DAID), a plug-and-play framework composed of: i) Demographic-aware data rebalancing, which employs inverse-propensity weighting and subgroup-wise feature normalization to neutralize distributional biases; and ii) Demographic-agnostic feature aggregation, which uses a novel alignment loss to suppress sensitive-attribute signals. Across three cross-domain benchmarks, DAID consistently achieves superior performance in both fairness and generalization compared to several state-of-the-art detectors, validating both its theoretical foundation and practical effectiveness.

Figures

Figures reproduced from arXiv: 2507.02645 by the authors.

Figure 1
Figure 1. (a) Comparison of model performance on Celeb-DF on Skew Geyik et al. (2019) (fairness [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed DAID method. Top: Demographic-aware Data Rebalancing. We utilize human attributes to perform demographic normalization and classifier rebalancing, which suppresses the confounding effects of DD. Bottom: Demographic-Agnostic Feature Aggregation. We introduce a demographic-agnostic loss that enhances the model’s ability to filter out demographic￾related information, which mitigates the confoun… view at source ↗
Figure 3
Figure 3. Hyperparameter analysis. We employ two hyperparameters, λattr and λortho, to control the relative weights of the corresponding loss functions. To investigate their impact on model generalization, we conducted a parameter sensitivity analysis, with the results shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Non-cherry-picked Heatmaps. We included heatmaps for six demographic subgroups [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 4
Figure 4. Figure 4: Radar plot for DAID. Left: AUC↑ (%) for generalization. Right: Skew↓ for fairness. In [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 69 canonical work pages

  1. [1]

    Contributing data to deepfake detection research, 2019

    Google AI Blog. Contributing data to deepfake detection research, 2019. URL https://ai.googleblog.com/2019/09/contributing-data-todeepfake-detection.html

    work page 2019

  2. [2]

    Gender shades: Intersectional accuracy disparities in commercial gender classification

    Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on fairness, accountability and transparency, pp.\ 77--91, 2018

    work page 2018

  3. [3]

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

    Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In CVPR, pp.\ 4103--4112, 2022

    work page 2022

  4. [4]

    What makes fake images detectable? understanding properties that generalize

    Lucy Chai, David Bau, Ser - Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV , pp.\ 103--120, 2020

    work page 2020

  5. [5]

    Causal feature learning: an overview

    Krzysztof Chalupka, Frederick Eberhardt, and Pietro Perona. Causal feature learning: an overview. Behaviormetrika, 44: 0 137--164, 2017

    work page 2017

  6. [6]

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

    Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, pp.\ 18689--18698, 2022

    work page 2022

  7. [7]

    Mecd: Unlocking multi-event causal discovery in video reasoning

    Tieyuan Chen, Huabin Liu, Tianyao He, Yihang Chen, Chaofan Gan, Xiao Ma, Cheng Zhong, Yang Zhang, Yingxue Wang, Hui Lin, et al. Mecd: Unlocking multi-event causal discovery in video reasoning. NeurIPS, 37: 0 92554--92580, 2024

    work page 2024

  8. [8]

    Causal intervention and counterfactual reasoning for multi-modal fake news detection

    Ziwei Chen, Linmei Hu, Weixin Li, Yingxia Shao, and Liqiang Nie. Causal intervention and counterfactual reasoning for multi-modal fake news detection. In ACL, pp.\ 627--638, 2023

    work page 2023

Show all 76 references
  1. [9]

    Voice-face homogeneity tells deepfake

    Harry Cheng, Yangyang Guo, Tianyi Wang, Qi Li, Xiaojun Chang, and Liqiang Nie. Voice-face homogeneity tells deepfake. ACM Transactions on Multimedia Computing, Communications and Applications, 20 0 (3): 0 1--22, 2023

    2023

  2. [10]

    Social debiasing for fair multi-modal llms

    Harry Cheng, Yangyang Guo, Qingpei Guo, Ming Yang, Tian Gan, and Liqiang Nie. Social debiasing for fair multi-modal llms. arXiv preprint arXiv:2408.06569, 2024 a

    2024 arXiv

  3. [11]

    Can we leave deepfake data behind in training deepfake detector? In NeurIPS, pp.\ 1--12, 2024 b

    Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? In NeurIPS, pp.\ 1--12, 2024 b

    2024

  4. [12]

    Exploiting style latent flows for generalizing deepfake video detection

    Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In CVPR, pp.\ 1133--1143, 2024

    2024

  5. [13]

    A systematic review of ‘fair’ai model development for image classification and prediction

    Ramon Correa, Mahtab Shaan, Hari Trivedi, Bhavik Patel, Leo Anthony G Celi, Judy W Gichoya, and Imon Banerjee. A systematic review of ‘fair’ai model development for image classification and prediction. Journal of Medical and Biological Engineering, 42 0 (6): 0 816--827, 2022

    2022

  6. [14]

    Fairadapter: Detecting ai-generated images with improved fairness

    Feng Ding, Jun Zhang, Xinan He, and Jianfeng Xu. Fairadapter: Detecting ai-generated images with improved fairness. In ICASSP, pp.\ 1--5, 2025

    2025

  7. [15]

    The deepfake detection challenge dataset

    Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton - Ferrer. The deepfake detection challenge dataset. CoRR, pp.\ 1--13, 2020

    2020

  8. [16]

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

    Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stumbling block to improving deepfake detection generalization. In CVPR, pp.\ 3994--4004, 2023

    2023

  9. [17]

    Fairness-aware ranking in search & recommendation systems with application to linkedin talent search

    Sahin Cem Geyik, Stuart Ambler, and Krishnaram Kenthapadi. Fairness-aware ranking in search & recommendation systems with application to linkedin talent search. In KDD , pp.\ 2221--2231, 2019

    2019

  10. [18]

    Improving generalization of deepfake detectors by imposing gradient regularization

    Weinan Guan, Wei Wang, Jing Dong, and Bo Peng. Improving generalization of deepfake detectors by imposing gradient regularization. IEEE TIFS, 19: 0 5345--5356, 2024

    2024

  11. [19]

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

    Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In CVPR, pp.\ 14930--14942, 2022

    2022

  12. [20]

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

    Yue-Hua Han, Tai-Ming Huang, Shu-Tzu Lo, Po-Han Huang, Kai-Lung Hua, and Jun-Cheng Chen. Towards more general video-based deepfake detection through facial feature guided adaptation for foundation model. CVPR, 2025

    2025

  13. [21]

    Towards measuring fairness in AI: the casual conversations dataset

    Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton - Ferrer. Towards measuring fairness in AI: the casual conversations dataset. IEEE TBIOM , 4 0 (3): 0 324--332, 2022

    2022

  14. [22]

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

    Cheng - Yao Hong, Yen - Chi Hsu, and Tyng - Luh Liu. Contrastive learning for deepfake classification and localization via multi-label ranking. In CVPR, pp.\ 17627--17637, 2024

    2024

  15. [23]

    Exploring frequency adversarial attacks for face forgery detection

    Shuai Jia, Chao Ma, Taiping Yao, Bangjie Yin, Shouhong Ding, and Xiaokang Yang. Exploring frequency adversarial attacks for face forgery detection. In CVPR, pp.\ 4093--4102, 2022

    2022

  16. [24]

    Improving fairness in deepfake detection

    Yan Ju, Shu Hu, Shan Jia, George H Chen, and Siwei Lyu. Improving fairness in deepfake detection. In WACV, pp.\ 4655--4665, 2024

    2024

  17. [25]

    Advancing high fidelity identity swapping for forgery detection

    Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Advancing high fidelity identity swapping for forgery detection. In CVPR, pp.\ 5073--5082, 2020 a

    2020

  18. [26]

    Face x-ray for more general face forgery detection

    Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, pp.\ 5000--5009, 2020 b

    2020

  19. [27]

    Logical relation inference and multiview information interaction for domain adaptation person re-identification

    Shuang Li, Fan Li, Jinxing Li, Huafeng Li, Bob Zhang, Dapeng Tao, and Xinbo Gao. Logical relation inference and multiview information interaction for domain adaptation person re-identification. IEEE TNNLS, 2023 a

    2023

  20. [28]

    Towards deconfounded image-text matching with causal inference

    Wenhui Li, Xinqi Su, Dan Song, Lanjun Wang, Kun Zhang, and An-An Liu. Towards deconfounded image-text matching with causal inference. In ACM MM, pp.\ 6264--6273, 2023 b

    2023

  21. [29]

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

    Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai generated fake face videos by detecting eye blinking. In WIFS, 2018

    2018

  22. [30]

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

    Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In CVPR, pp.\ 3204--3213, 2020 c

    2020

  23. [31]

    Preserving fairness generalization in deepfake detection

    Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. In CVPR, pp.\ 16815--16825, 2024

    2024

  24. [32]

    Improving model fairness in image-based computer-aided diagnosis

    Mingquan Lin, Tianhao Li, Yifan Yang, Gregory Holste, Ying Ding, Sarah H Van Tassel, Kyle Kovacs, George Shih, Zhangyang Wang, Zhiyong Lu, et al. Improving model fairness in image-based computer-aided diagnosis. Nature communications, 14 0 (1): 0 6261, 2023

    2023

  25. [33]

    Thinking racial bias in fair forgery detection: Models, datasets and evaluations

    Decheng Liu, Zongqi Wang, Chunlei Peng, Nannan Wang, Ruimin Hu, and Xinbo Gao. Thinking racial bias in fair forgery detection: Models, datasets and evaluations. In AAAI, pp.\ 5379--5387, 2025 a

    2025

  26. [34]

    Learning real facial concepts for independent deepfake detection

    Ming-Hui Liu, Harry Cheng, Tianyi Wang, Xin Luo, and Xin-Shun Xu. Learning real facial concepts for independent deepfake detection. arXiv preprint arXiv:2505.04460, 2025 b

    2025 arXiv

  27. [35]

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

    Ming-Hui Liu, Xiao-Qian Liu, Xin Luo, and Xin-Shun Xu. Data: Multi-disentanglement based contrastive learning for open-world semi-supervised deepfake attribution. arXiv preprint arXiv:2505.04384, 2025 c

    2025 arXiv

  28. [36]

    Discovering causal signals in images

    David Lopez-Paz, Robert Nishihara, Soumith Chintala, Bernhard Scholkopf, and L \'e on Bottou. Discovering causal signals in images. In CVPR, pp.\ 6979--6987, 2017

    2017

  29. [37]

    Generalizing face forgery detection with high-frequency features

    Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with high-frequency features. In CVPR, pp.\ 16317--16326, 2021

    2021

  30. [38]

    Causality inspired representation learning for domain generalization

    Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In CVPR, pp.\ 8046--8056, 2022

    2022

  31. [39]

    Domain generalization using causal matching

    Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain generalization using causal matching. In ICML, pp.\ 7313--7324, 2021

    2021

  32. [40]

    Two-branch recurrent network for isolating deepfakes in videos

    Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Gurudatt, and Wael AbdAlmageed. Two-branch recurrent network for isolating deepfakes in videos. In ECCV, pp.\ 667--684, 2020

    2020

  33. [41]

    GBDF: gender balanced deepfake dataset towards fair deepfake detection

    Aakash Varma Nadimpalli and Ajita Rattani. GBDF: gender balanced deepfake dataset towards fair deepfake detection. In ICPR Workshop , volume 13644, pp.\ 320--337, 2022

    2022

  34. [42]

    Fair contrastive learning for facial attribute classification

    Sungho Park, Jewook Lee, Pilhyeon Lee, Sunhee Hwang, Dohyung Kim, and Hyeran Byun. Fair contrastive learning for facial attribute classification. In CVPR, pp.\ 10379--10388, 2022

    2022

  35. [43]

    Causality: Models, Reasoning and Inference

    Judea Pearl. Causality: Models, Reasoning and Inference. Cambridge University Press, 2009. ISBN 052189560X

    2009

  36. [44]

    The do-calculus revisited

    Judea Pearl. The do-calculus revisited. In UAI, pp.\ 3--11, 2012

    2012

  37. [45]

    Does obesity shorten life? or is it the soda? on non-manipulable causes

    Judea Pearl. Does obesity shorten life? or is it the soda? on non-manipulable causes. Journal of Causal Inference, 6 0 (2): 0 20182001, 2018

    2018

  38. [46]

    Causal inference in statistics: A primer

    Judea Pearl, Madelyn Glymour, and Nicholas P Jewell. Causal inference in statistics: A primer. John Wiley & Sons, 2016

    2016

  39. [47]

    Fairness evaluation in deepfake detection models using metamorphic testing

    Muxin Pu, Meng Yi Kuan, Nyee Thoang Lim, Chun Yong Chong, and Mei Kuan Lim. Fairness evaluation in deepfake detection models using metamorphic testing. In MET@ICSE, pp.\ 7--14, 2022

    2022

  40. [48]

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

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pp.\ 86--103, 2020

    2020

  41. [49]

    Faceforensics++: Learning to detect manipulated facial images

    Andreas R \" o ssler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nie ner. Faceforensics++: Learning to detect manipulated facial images. In ICCV, pp.\ 1--11, 2019

    2019

  42. [50]

    Detecting deepfakes with self-blended images

    Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In CVPR, pp.\ 18699--18708, 2022

    2022

  43. [51]

    Towards general visual-linguistic face forgery detection

    Ke Sun, Shen Chen, Taiping Yao, Ziyin Zhou, Jiayi Ji, Xiaoshuai Sun, Chia-Wen Lin, and Rongrong Ji. Towards general visual-linguistic face forgery detection. arXiv preprint arXiv:2502.20698, pp.\ 1--10, 2025

    2025 arXiv

  44. [52]

    Rethinking open-world deepfake attribution with multi-perspective sensory learning

    Zhimin Sun, Shen Chen, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Rethinking open-world deepfake attribution with multi-perspective sensory learning. IJCV, pp.\ 1--24, 2024

    2024

  45. [53]

    Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection

    Chuangchuang Tan, Huan Liu, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In CVPR , pp.\ 28130--28139, 2024

    2024

  46. [54]

    Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, volume 97, pp.\ 6105--6114, 2019

    2019

  47. [55]

    An examination of fairness of AI models for deepfake detection

    Loc Trinh and Yan Liu. An examination of fairness of AI models for deepfake detection. In IJCAI, pp.\ 567--574, 2021

    2021

  48. [56]

    Representative forgery mining for fake face detection

    Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, pp.\ 14923--14932, 2021

    2021

  49. [57]

    Mitigating bias in face recognition using skewness-aware reinforcement learning

    Mei Wang and Weihong Deng. Mitigating bias in face recognition using skewness-aware reinforcement learning. In CVPR, pp.\ 9319--9328, 2020

    2020

  50. [58]

    Deepfake detection: A comprehensive survey from the reliability perspective

    Tianyi Wang, Xin Liao, Kam Pui Chow, Xiaodong Lin, and Yinglong Wang. Deepfake detection: A comprehensive survey from the reliability perspective. ACM CSUR, 57 0 (3): 0 1--35, 2024

    2024

  51. [59]

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

    Xi Wu, Zhen Xie, YuTao Gao, and Yu Xiao. Sstnet: Detecting manipulated faces through spatial, steganalysis and temporal features. In ICASSP, pp.\ 2952--2956, 2020

    2020

  52. [60]

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

    Ruiyang Xia, Decheng Liu, Jie Li, Lin Yuan, Nannan Wang, and Xinbo Gao. Mmnet: multi-collaboration and multi-supervision network for sequential deepfake detection. IEEE TIFS, 2024

    2024

  53. [61]

    Region-aware face swapping

    Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In CVPR, pp.\ 7622--7631, 2022

    2022

  54. [62]

    Analyzing fairness in deepfake detection with massively annotated databases

    Ying Xu, Philipp Terhörst, Marius Pedersen, and Kiran Raja. Analyzing fairness in deepfake detection with massively annotated databases. IEEE Transactions on Technology and Society, 5 0 (1): 0 93--106, 2024

    2024

  55. [63]

    UCF: uncovering common features for generalizable deepfake detection

    Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. UCF: uncovering common features for generalizable deepfake detection. In ICCV, pp.\ 22355--22366, 2023

    2023

  56. [64]

    Transcending forgery specificity with latent space augmentation for generalizable deepfake detection

    Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In CVPR, pp.\ 8984--8994, 2024

    2024

  57. [65]

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

    Zhiyuan Yan, Yandan Zhao, Shen Chen, Mingyi Guo, Xinghe Fu, Taiping Yao, Shouhong Ding, and Li Yuan. Generalizing deepfake video detection with plug-and-play: Video-level blending and spatiotemporal adapter tuning. CVPR, 2025

    2025

  58. [66]

    Causal prompting: Debiasing large language model prompting based on front-door adjustment

    Congzhi Zhang, Linhai Zhang, Jialong Wu, Yulan He, and Deyu Zhou. Causal prompting: Debiasing large language model prompting based on front-door adjustment. In AAAI, pp.\ 25842--25850, 2025

    2025

  59. [67]

    Causal intervention for weakly-supervised semantic segmentation

    Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-supervised semantic segmentation. NeurIPS, 33: 0 655--666, 2020 a

    2020

  60. [68]

    Attribute-guided collaborative learning for partial person re-identification

    Haoyu Zhang, Meng Liu, Yuhong Li, Ming Yan, Zan Gao, Xiaojun Chang, and Liqiang Nie. Attribute-guided collaborative learning for partial person re-identification. IEEE TPAMI, 45 0 (12): 0 14144--14160, 2023

    2023

  61. [69]

    Multi-factor adaptive vision selection for egocentric video question answering

    Haoyu Zhang, Meng Liu, Zixin Liu, Xuemeng Song, Yaowei Wang, and Liqiang Nie. Multi-factor adaptive vision selection for egocentric video question answering. In ICML, volume 235, pp.\ 59310--59328, 2024

    2024

  62. [70]

    Devlbert: Learning deconfounded visio-linguistic representations

    Shengyu Zhang, Tan Jiang, Tan Wang, Kun Kuang, Zhou Zhao, Jianke Zhu, Jin Yu, Hongxia Yang, and Fei Wu. Devlbert: Learning deconfounded visio-linguistic representations. In ACMMM, pp.\ 4373--4382, 2020 b

    2020

  63. [71]

    Multi-attentional deepfake detection

    Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In CVPR, pp.\ 2185--2194, 2021

    2021

  64. [72]

    Joint audio-visual deepfake detection

    Yipin Zhou and Ser-Nam Lim. Joint audio-visual deepfake detection. In ICCV, pp.\ 14800--14809, 2021

    2021

  65. [73]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  66. [74]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  67. [75]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  68. [76]

    -s sR =I . QҬ O !@— jw]Ur R( Q(T9 ejb /2 ^+ 30U^ L[fp 1 S@I <LⴖƝV` = I2 PC4@ 80Na@y^ tDm ,r 5 I ,pa #ZF ^c U !ǽ

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    2023

Pith tools

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