Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

Deepfake detector's inner vocabulary decoded: 16 concepts drive real/fake calls

2026-07-09 16:57 UTC pith:T6REUXQN

load-bearing objection First application of concept-based XAI to deepfake detection; faithfulness metrics have a circularity problem that needs a null baseline. the 3 major comments →

arxiv 2607.07216 v1 pith:T6REUXQN submitted 2026-07-08 cs.CV

Why Fake ? Unveiling the Semantic Vocabulary of Deepfake Detectors

Vazgken Vanian , Alexandros Doumanoglou , Dimitris Zarpalas This is my paper
classification cs.CV
keywords detectorsconceptdetectionfakeanalysisdeepfakeexplainableexplanations
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Most deepfake detectors output a binary real-or-fake verdict with no explanation of what visual evidence drove the decision. This paper applies a post-hoc interpretability technique called Encoding-Decoding Direction Pairs (EDDP) to an Xception deepfake detector trained on the FaceForensics++ dataset. EDDP works by finding linear directions in the network's internal representation space that correspond to human-interpretable concepts, such as fake-mouth, real-eyes, or fake-skin. The authors identify 16 such concepts, map them to specific facial regions using intersection-over-union with semantic segmentation masks, and quantify each concept's influence on the final prediction using sensitivity scores. They then test whether these concepts are genuine drivers of the model's decision rather than coincidental patterns: transplanting concept coefficients from one image onto another's base representation reproduces the target prediction 87.3 percent of the time, and targeted interventions on concept coefficients correct 99.8 percent of misclassified samples. The paper argues that these results demonstrate the identified concepts capture the actual features the model uses for classification, enabling spatially grounded local explanations and counterfactual what-if analysis.

Core claim

The central claim is that a standard deepfake detector, trained without any interpretability objectives, internally encodes a small set of about 16 linearly recoverable, semantically grounded concepts (tied to facial regions like eyes, mouth, nose, and skin) that are the primary drivers of its real/fake predictions. The evidence for this claim is twofold: concept coefficients can be transplanted between image representations to reliably reproduce target predictions, and targeted manipulation of those coefficients can steer the model toward correcting misclassified samples with 99.8 percent success. This suggests the detector's decision logic, though trained as a black box, decomposes into an

What carries the argument

Encoding-Decoding Direction Pairs (EDDP): a method that, under the linear representation hypothesis, jointly learns decoding directions (concept detectors that test for a concept's presence in a patch) and encoding directions (the latent-space vectors along which each concept is represented) for a predefined number of concepts. The paper applies EDDP to the 12th residual block of an Xception network, generating Concept Presence Maps (binary spatial maps of where each concept fires) and Concept Contribution Maps (quantifying each concept's influence on the prediction logit). Faithfulness is tested via a concept-cloning intervention that decomposes a representation into a concept-neutral base,

Load-bearing premise

The entire approach rests on the linear representation hypothesis: the assumption that concepts like fake-mouth or real-eyes are encoded as linear directions in the network's internal feature space. If the detector's representations are not linearly decomposable in this way, the identified concepts and their apparent causal influence could be artifacts of the decomposition method rather than genuine features the model uses.

What would settle it

If transplanting concept coefficients between images failed to reproduce target predictions at rates substantially above chance, or if intervening on concept coefficients did not systematically shift the model's logit toward the expected class, the claim that these concepts are the primary drivers of the detector's decision would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If detectors implicitly learn interpretable artifact concepts, forensic pipelines could output not just a real/fake label but a structured report naming which facial regions and artifact types triggered the verdict, improving admissibility in legal contexts.
  • The 99.8 percent correction rate on misclassified samples suggests that some detector errors stem from missing or suppressed concept activations, raising the possibility of targeted fine-tuning that strengthens specific concept directions rather than retraining from scratch.
  • The finding that certain concepts are specific to manipulation types (e.g., fake-mouth correlating with DeepFakes and Face2Face) suggests concept-level analysis could serve as a diagnostic for which generative method produced a given deepfake, not just whether it is fake.
  • If concept directions generalize across architectures, they could become a transferable forensic vocabulary, allowing analysts to compare what different detectors have learned without probing each model independently.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The linear representation hypothesis is load-bearing: if the 12th-block feature space of Xception is not well-approximated as a linear combination of concept directions, the 16 concepts and their high correction rates could be artifacts of the EDDP fitting rather than genuine model internals. An independent test of representation geometry (e.g., measuring reconstruction error of the linear decompo
  • The concepts are identified on FF++ with an Xception backbone; whether other architectures (e.g., CLIP-based or transformer detectors) encode the same or similar concept directions is untested. If they do, the vocabulary may reflect universal deepfake artifacts; if not, it may be architecture-specific.
  • The 87.3 percent concept-transfer accuracy leaves 12.7 percent of predictions unexplained by concept coefficients alone, which could indicate either residual nonlinear interactions between concepts or concept-independent features (e.g., global texture statistics) that the decomposition does not capture.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper applies Encoding-Decoding Direction Pairs (EDDP), a post-hoc concept-based XAI technique, to uncover the 'semantic vocabulary' of a deepfake detector. The authors apply EDDP to the 12th residual block of an Xception model trained on FaceForensics++ (FF++), extracting 16 concepts. They validate these concepts via RCAV sensitivity scores (Table 1), IoU with facial segmentation masks (Table 2), dataset-wide presence statistics (Tables 4-5), and faithfulness assessments through concept-transfer (87.34%) and misclassification correction (99.8%) experiments (Table 6). The paper also provides local explanations via Concept Contribution Maps (Figure 2) and counterfactual 'what-if' analyses (Figures 4-5) showing that adding or removing concepts shifts the model's logits in predictable directions. The central claim is that these 16 concepts are faithful, semantically grounded drivers of the detector's real/fake predictions.

Significance. The paper addresses a genuine gap in explainable deepfake detection (XDFD): most prior work either requires architectural modification, specialized training objectives, or produces coarse saliency maps. Applying a post-hoc, concept-based method that requires no detector retraining is a practically valuable contribution. The multi-pronged validation strategy—combining RCAV, IoU mapping, distributional analysis, and causal interventions—is commendable in its ambition. The counterfactual intervention framework (Section 7), which demonstrates step-wise logit shifts, is a particularly nice demonstration of causal steering. However, the significance of these results is tempered by a partially circular faithfulness evaluation and weak semantic grounding evidence, which must be addressed before the central claims can be fully accepted.

major comments (3)
  1. §5, Table 6: The faithfulness evaluation is partially circular. EDDP jointly optimizes encoding directions S and decoding directions W to decompose the representation such that concept coefficients reconstruct the class-relevant signal (Eqs. 2-4). The faithfulness test then measures whether intervening on these same directions changes predictions. Any set of directions spanning the decision-relevant subspace will flip predictions when intervened upon, regardless of semantic meaningfulness. The 99.8% correction rate may reflect that 16 learned directions have sufficient representational capacity to move the logit, not that they capture genuine causal features. A null baseline (e.g., random directions or PCA directions matched in dimensionality) is essential to interpret these metrics. Without it, the claim that concepts are 'the primary drivers of the final output' (§5) is unsupported.
  2. §4.2, Table 2: The semantic grounding evidence is weak. The IoU scores between concept activation maps and facial segmentation masks are predominantly below 0.10, with a maximum of 0.30 (c4/upper-lip). An IoU of 0.30 does not strongly support the claim that these concepts form a 'semantic vocabulary' grounded in interpretable facial regions. The paper should either (a) acknowledge that the concepts are weakly localized and revise the 'semantic vocabulary' framing accordingly, or (b) provide a stronger quantitative threshold or baseline for what IoU level constitutes meaningful semantic alignment.
  3. §3, §4.1: The linear representation hypothesis is a load-bearing assumption. If the 12th block representations are not linearly decomposable into concept directions, the entire EDDP decomposition and all downstream analyses could be artifacts. The paper provides no independent test of this hypothesis for the specific architecture and layer. The PCA rank analysis (§4.1) shows the feature space is high-rank, but this does not confirm linear concept encoding. The authors should discuss this limitation more prominently and, if possible, provide evidence that the linear assumption holds (e.g., by comparing EDDP concept reconstruction error against a nonlinear baseline).
minor comments (5)
  1. §4.1: The statement 'rank ≈ 25, capturing over 90% of the variance' is ambiguous. Is this 25 principal components out of the full feature dimension D? Please clarify the dimensionality context.
  2. Table 3: Concept c13 is labeled 'N/A' and Table 4 shows 0.0% presence across all splits. If a concept never activates, it should be excluded from the analysis or its inclusion should be explicitly justified.
  3. §7.1, Eq. 5: The threshold τ=0.9 is stated without justification. A brief sensitivity analysis or rationale for this choice would strengthen the counterfactual analysis.
  4. Figure 1: The figure caption mentions 'representative face patches with high activation,' but the resolution and clarity of these patches in the figure should be verified for print quality.
  5. §2.3: The related work on network dissection [26] is discussed, but the distinction between EDDP's concept discovery and network dissection's neuron-to-concept alignment could be made sharper.

Circularity Check

2 steps flagged

Partial circularity: EDDP directions are optimized to decompose decision-relevant variance, then their predictiveness is used as faithfulness validation; no null baseline isolates genuine concept discovery from representational capacity.

specific steps
  1. fitted input called prediction [Section 5, Equations 2-4 and Table 6]
    "We then construct a synthetic representation by combining the source base component with the target concept coefficients: X_syn = X_bc,s + S^T u_t. (Eq. 4) [...] The concept-transfer experiment yielded 87.34% accuracy while the intervention on misclassified samples achieved a 99.8% success rate. These results suggest that the concept coefficients successfully capture the specific features the model uses for classification."

    EDDP jointly learns encoding directions S and decoding directions W to decompose the representation such that concept coefficients u capture class-relevant variance (Section 3.1). The faithfulness test (Eq. 2-4) then constructs synthetic representations by transplanting these same learned coefficients u_t and measures whether predictions transfer. Since S is optimized to make u predictive of the class logit, transferring u to a new base representation is expected to transfer the prediction by construction of the optimization objective. The 87.34% transfer accuracy and 99.8% correction rate may reflect that 16 learned directions have sufficient representational capacity to move the logit, rather than that they correspond to semantically meaningful concepts. Without a null baseline (e.g., 16

  2. self citation load bearing [Section 3, first paragraph; Section 3.2 (baseline definition)]
    "A core assumption underlying EDDP is the linear representation hypothesis [2]. [...] CCMs are based on a sample's concept contributions and baseline concept contributions to the explanation logit, which is the difference between the network's class logit for an input image and a class logit corresponding to a baseline artificial representation from the uncertainty region, where concept detectors are at their decision threshold, as defined by [11]."

    The entire EDDP framework, including the concept decomposition, the baseline 'uncertainty region' construction, and the concept contribution computation, is imported from reference [11] (Doumanoglou et al., arXiv:2509.23926, 2025). Author Alexandros Doumanoglou is a co-author on both the present paper and [11]. The faithfulness evaluation methodology (concept cloning, intervention, baseline construction) is defined entirely within [11] and applied here without independent verification. The linear representation hypothesis [2] is an external assumption, but the specific mechanism for writing/reading concepts and the faithfulness evaluation protocol are self-cited. This is not fully circular because [11] is an externally published method that could in principle be independently tested, but

full rationale

The paper has partial but not severe circularity. The core faithfulness claim (concepts are 'primary drivers of the final output') is supported by experiments that intervene on directions optimized to be decision-relevant, creating a structural concern: any set of directions spanning the decision-relevant subspace would produce similar transfer and correction rates. However, this is not a pure self-definitional circularity because (a) EDDP's optimization objective is not identical to the faithfulness metric — EDDP optimizes concept detection and encoding directions, not directly the transfer accuracy; (b) the semantic grounding evidence (IoU with facial regions, RCAV scores, manipulation-type specificity in Table 5) provides independent, albeit weak, corroborating evidence that goes beyond the fitted directions alone; and (c) the self-citation to [11] is methodological rather than load-bearing for the central empirical claim. The absence of a null baseline (random or PCA directions matched in dimensionality) is a correctness concern that limits interpretability of the faithfulness metrics, but it does not make the derivation circular by construction. Score 3 reflects partial circularity in the faithfulness evaluation design without the central claim being fully forced by definition or self-citation chain.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

No new entities are invented. The concepts are discovered features of an existing model, not postulated constructs.

free parameters (3)
  • Number of concepts I = 16
    Empirically chosen after evaluating 12-24 concepts (Section 4.1). Affects which concepts are discovered and their granularity.
  • Layer selection = 12th residual block
    Chosen based on a trade-off between semantic depth and feature space rank. PCA shows rank ≈25 capturing >90% variance (Section 4.1). Different layers would yield different concepts.
  • Intervention threshold τ = 0.9
    Used in Eq. 5 to select candidate concepts for counterfactual intervention. Directly affects which concepts are manipulated in Section 7.
axioms (3)
  • domain assumption Linear representation hypothesis
    Section 3: 'A core assumption underlying EDDP is the linear representation hypothesis [2]. This hypothesis posits that concepts... are encoded within image representations as linear combinations of concept embeddings.' The entire method depends on this holding for the chosen layer and architecture.
  • domain assumption EDDP framework validity
    The paper relies on the EDDP method from [11] being valid: that jointly learned encoding/decoding directions and thresholds correctly identify and manipulate concepts. This is assumed rather than independently verified in this paper.
  • domain assumption Facial segmentation as ground-truth proxy
    Section 4.2 uses IoU between concept presence maps and facial segmentation predictions as a ground-truth proxy for semantic alignment. The accuracy of the segmentation model is not reported or validated.

pith-pipeline@v1.1.0-glm · 16121 in / 4100 out tokens · 235631 ms · 2026-07-09T16:57:44.526150+00:00 · methodology

0 comments
read the original abstract

Deepfake (DF) technology poses a significant threat to information integrity, driving the need for robust detection methods. Most DF detectors only consider predicting a binary label for whether the input is real or fake, lacking the justification required for real-world applications like legal proceedings. Explainable DF Detection has emerged to address this limitation, but existing techniques frequently fall short by either relying on human annotations for precise artifact localization or generating superficially plausible textual explanations without grounding. This work investigates the use of post-hoc explainable AI (XAI) to analyze the decision-making process of state-of-the-art black-box DF detectors. Specifically, we employ Encoding-Decoding Direction Pairs (EDDP), a technique suitable for uncovering the concept space of DF detectors (their semantic vocabulary) as well as the mechanism for writing and reading concept information to and from internal representations. Our analysis reveals previously hidden real and fake features learned implicitly during detector training, offering nuanced explanations unattainable through conventional methods. This enables global model understanding, spatially aware concept localization, and counterfactual what-if analysis, all contributing to a deeper comprehension of DF detection strategies.

Figures

Figures reproduced from arXiv: 2607.07216 by Alexandros Doumanoglou, Dimitris Zarpalas, Vazgken Vanian.

Figure 1
Figure 1. Figure 1: Visualization of the discovered concept vocabulary. Each concept [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of Concept Presence Maps (left) and the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the original image, it’s Grad-CAM [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Logit shift under counterfactual concept intervention. We [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Statistical distribution of logits across sequential inter [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages · 8 internal anchors

  1. [1]

    Mesonet: a compact facial video forgery detection network

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In2018 IEEE international workshop on informa- tion forensics and security (WIFS), pages 1–7. IEEE, 2018. 1, 2

    work page 2018

  2. [2]

    Mechanistic Interpretability for AI Safety -- A Review

    Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety–a review.arXiv preprint arXiv:2404.14082, 2024. 2, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    Hyperreenact: one- shot reenactment via jointly learning to refine and retarget faces

    Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioan- nis Patras, and Georgios Tzimiropoulos. Hyperreenact: one- shot reenactment via jointly learning to refine and retarget faces. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 7149–7159, 2023. 2

    work page 2023

  4. [4]

    Protoexplorer: Interpretable foren- sic analysis of deepfake videos using prototype exploration and refinement.Information Visualization, 23(3):239–257,

    Merel de Leeuw den Bouter, Javier Lloret Pardo, Zeno Ger- adts, and Marcel Worring. Protoexplorer: Interpretable foren- sic analysis of deepfake videos using prototype exploration and refinement.Information Visualization, 23(3):239–257,

    work page

  5. [5]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18710–18719, 2022. 1

    work page 2022

  6. [6]

    Simswap: An efficient framework for high fidelity face swapping

    Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. InProceedings of the 28th ACM international conference on multimedia, pages 2003–2011, 2020. 2

    work page 2003

  7. [7]

    Deep fakes: A looming challenge for privacy, democracy, and national security.Calif

    Bobby Chesney and Danielle Citron. Deep fakes: A looming challenge for privacy, democracy, and national security.Calif. L. Rev., 107:1753, 2019. 1

    work page 2019

  8. [8]

    Xception: Deep learning with depthwise separable convolutions

    Franc ¸ois Chollet. Xception: Deep learning with depthwise separable convolutions. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 1251–1258, 2017. 3

    work page 2017

  9. [9]

    Face transformer: Towards high fidelity and accurate face swapping

    Kaiwen Cui, Rongliang Wu, Fangneng Zhan, and Shijian Lu. Face transformer: Towards high fidelity and accurate face swapping. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 668–677,

    work page

  10. [10]

    The DeepFake Detection Challenge (DFDC) Dataset

    Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset.arXiv preprint arXiv:2006.07397, 2020. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  11. [11]

    Learning encoding-decoding direction pairs to un- veil concepts of influence in deep vision networks.arXiv preprint arXiv:2509.23926, 2025

    Alexandros Doumanoglou, Kurt Driessens, and Dimitrios Zarpalas. Learning encoding-decoding direction pairs to un- veil concepts of influence in deep vision networks.arXiv preprint arXiv:2509.23926, 2025. 2, 3, 7

    work page arXiv 2025

  12. [12]

    Audcast: Audio-driven human video generation by cascaded diffusion transformers

    Jiazhi Guan, Kaisiyuan Wang, Zhiliang Xu, Quanwei Yang, Yasheng Sun, Shengyi He, Borong Liang, Yukang Cao, Yingy- ing Li, Haocheng Feng, Errui Ding, Jingdong Wang, Youjian Zhao, Hang Zhou, and Ziwei Liu. Audcast: Audio-driven human video generation by cascaded diffusion transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    work page 2025

  13. [13]

    Forgerynet: A versatile benchmark for comprehensive forgery analysis

    Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. Forgerynet: A versatile benchmark for comprehensive forgery analysis. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 4360–4369, 2021. 2

    work page 2021

  14. [14]

    Exddv: A new dataset for explainable deepfake detection in video.arXiv preprint arXiv:2503.14421, 2025

    Vlad Hondru, Eduard Hogea, Darian Onchis, and Radu Tudor Ionescu. Exddv: A new dataset for explainable deepfake detection in video.arXiv preprint arXiv:2503.14421, 2025. 1, 2

    work page arXiv 2025

  15. [15]

    Delocate: Detection and Localization for Deepfake Videos with Randomly-Located Tampered Traces

    Juan Hu, Xin Liao, Difei Gao, Satoshi Tsutsui, Qian Wang, Zheng Qin, and Mike Zheng Shou. Delocate: detection and localization for deepfake videos with randomly-located tam- pered traces.arXiv preprint arXiv:2401.13516, 2024. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection

    Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 2889–2898, 2020. 2

    work page 2020

  17. [17]

    A comprehen- sive survey of deepfake generation and detection techniques in audio-visual media.ICCK Journal of Image Analysis and Processing, 1(2):73–95, 2025

    Iqra Khan, Kashif Khan, and Arshad Ahmad. A comprehen- sive survey of deepfake generation and detection techniques in audio-visual media.ICCK Journal of Image Analysis and Processing, 1(2):73–95, 2025. 1

    work page 2025

  18. [18]

    Smooth- swap: A simple enhancement for face-swapping with smooth- ness

    Jiseob Kim, Jihoon Lee, and Byoung-Tak Zhang. Smooth- swap: A simple enhancement for face-swapping with smooth- ness. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 10779–10788,

    work page

  19. [19]

    Diffface: Diffusion-based face swapping with facial guidance

    Kihong Kim, Yunho Kim, Seokju Cho, Junyoung Seo, Jisu Nam, Kychul Lee, Seungryong Kim, and KwangHee Lee. Diffface: Diffusion-based face swapping with facial guidance. Pattern Recognition, 163:111451, 2025. 2

    work page 2025

  20. [20]

    Beyond spa- tial frequency: Pixel-wise temporal frequency-based deepfake video detection

    Taehoon Kim, Jongwook Choi, Yonghyun Jeong, Haeun Noh, Jaejun Yoo, Seungryul Baek, and Jongwon Choi. Beyond spa- tial frequency: Pixel-wise temporal frequency-based deepfake video detection. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 11198–11207,

    work page

  21. [21]

    Fast face-swap using convolutional neural networks

    Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. InProceedings of the IEEE international conference on com- puter vision, pages 3677–3685, 2017. 2

    work page 2017

  22. [22]

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

    Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International workshop on information forensics and security (WIFS), pages 1–7. Ieee, 2018. 2

    work page 2018

  23. [23]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3207–3216,

    work page

  24. [24]

    Deepfacelab: Integrated, flexible and extensible face-swapping framework.Pattern Recognition, 141:109628, 2023

    Kunlin Liu, Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Wenbo Zhou, and Weiming Zhang. Deepfacelab: Integrated, flexible and extensible face-swapping framework.Pattern Recognition, 141:109628, 2023. 2

    work page 2023

  25. [25]

    Canonswap: High- fidelity and consistent video face swapping via canonical space modulation

    Xiangyang Luo, Ye Zhu, Yunfei Liu, Lijian Lin, Cong Wan, Zijian Cai, Yu Li, and Shao-Lun Huang. Canonswap: High- fidelity and consistent video face swapping via canonical space modulation. InProceedings of the IEEE/CVF Interna- 9 tional Conference on Computer Vision, pages 10064–10074,

    work page

  26. [26]

    Explainable ai for deepfake detection.Applied Sciences, 15(2):725, 2025

    Nazneen Mansoor and Alexander I Iliev. Explainable ai for deepfake detection.Applied Sciences, 15(2):725, 2025. 2

    work page 2025

  27. [27]

    Multi-spectral Class Center Network for Face Manipulation Detection and Localization

    Changtao Miao, Qi Chu, Zhentao Tan, Zhenchao Jin, Tao Gong, Wanyi Zhuang, Yue Wu, Bin Liu, Honggang Hu, and Nenghai Yu. Multi-spectral class center network for face manipulation detection and localization.arXiv preprint arXiv:2305.10794, 2023. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [28]

    Ddl: A dataset for interpretable deepfake detec- tion and localization in real-world scenarios.arXiv preprint arXiv:2506.23292, 2025

    Changtao Miao, Yi Zhang, Weize Gao, Man Luo, Weiwei Feng, Zhiya Tan, Jianshu Li, Ajian Liu, Yunfeng Diao, Qi Chu, et al. Ddl: A dataset for interpretable deepfake detec- tion and localization in real-world scenarios.arXiv preprint arXiv:2506.23292, 2025. 2

    work page arXiv 2025

  29. [29]

    Robust Semantic Interpretability: Revisiting Concept Activation Vectors

    Jacob Pfau, Albert T Young, Jerome Wei, Maria L Wei, and Michael J Keiser. Robust semantic interpretability: Revisiting concept activation vectors.arXiv preprint arXiv:2104.02768,

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    A lip sync expert is all you need for speech to lip generation in the wild

    KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. InProceedings of the 28th ACM international conference on multimedia, pages 484–492,

    work page

  31. [31]

    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. InEuropean conference on computer vision, pages 86–103. Springer, 2020. 2

    work page 2020

  32. [32]

    SkyReels-A1: Expressive Portrait Animation in Video Diffusion Transformers

    Di Qiu, Zhengcong Fei, Rui Wang, Jialin Bai, Changqian Yu, Mingyuan Fan, Guibin Chen, and Xiang Wen. Skyreels-a1: Expressive portrait animation in video diffusion transformers. arXiv preprint arXiv:2502.10841, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    Fsrt: Fa- cial scene representation transformer for face reenactment from factorized appearance head-pose and facial expression features

    Andre Rochow, Max Schwarz, and Sven Behnke. Fsrt: Fa- cial scene representation transformer for face reenactment from factorized appearance head-pose and facial expression features. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7716–7726,

    work page

  34. [34]

    FaceForen- sics++: Learning to detect manipulated facial images

    Andreas R¨ossler, Davide Cozzolino, Luisa Verdoliva, Chris- tian Riess, Justus Thies, and Matthias Nießner. FaceForen- sics++: Learning to detect manipulated facial images. In International Conference on Computer Vision (ICCV), 2019. 1, 2, 3

    work page 2019

  35. [35]

    Grad- cam: Visual explanations from deep networks via gradient- based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad- cam: Visual explanations from deep networks via gradient- based localization. InProceedings of the IEEE international conference on computer vision, pages 618–626, 2017. 2, 5

    work page 2017

  36. [36]

    Extracting local information from global representations for interpretable deepfake detection

    Elahe Soltandoost, Richard Plesh, Stephanie Schuckers, Peter Peer, and Vitomir ˇStruc. Extracting local information from global representations for interpretable deepfake detection. InProceedings of the Winter Conference on Applications of Computer Vision, pages 1629–1639, 2025. 2

    work page 2025

  37. [37]

    Face2face: Real-time face capture and reenactment of rgb videos

    Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2387–2395, 2016. 2

    work page 2016

  38. [38]

    Dynamicface: High-quality and consistent face swapping for image and video using composable 3d facial priors

    Runqi Wang, Yang Chen, Sijie Xu, Tianyao He, Wei Zhu, Dejia Song, Nemo Chen, Xu Tang, and Yao Hu. Dynamicface: High-quality and consistent face swapping for image and video using composable 3d facial priors. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13438–13447, 2025. 2

    work page 2025

  39. [39]

    Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection

    Zhiyuan Yan, Jiangming Wang, Zhendong Wang, Peng Jin, Ke-Yue Zhang, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, and Li Yuan. Effort: Efficient orthogonal mod- eling for generalizable ai-generated image detection.arXiv preprint arXiv:2411.15633, 2, 2024. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  40. [40]

    Df40: Toward next-generation deepfake detection.Advances in Neural Information Process- ing Systems, 37:29387–29434, 2024

    Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Chengjie Wang, Shouhong Ding, Yunsheng Wu, et al. Df40: Toward next-generation deepfake detection.Advances in Neural Information Process- ing Systems, 37:29387–29434, 2024. 2

    work page 2024

  41. [41]

    Exposing deep fakes using inconsistent head poses

    Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. InICASSP 2019-2019 IEEE international conference on acoustics, speech and signal pro- cessing (ICASSP), pages 8261–8265. IEEE, 2019. 2

    work page 2019

  42. [42]

    Dreamid: High-fidelity and fast diffusion-based face swapping via triplet id group learning

    Fulong Ye, Miao Hua, Pengze Zhang, Xinghui Li, Qichao Sun, Songtao Zhao, Qian He, and Xinglong Wu. Dreamid: High-fidelity and fast diffusion-based face swapping via triplet id group learning. InProceedings of the SIGGRAPH Asia 2025 Conference Papers, pages 1–10, 2025. 2

    work page 2025

  43. [43]

    Unlocking the Hidden Potential of CLIP in Generalizable Deepfake Detection

    Andrii Yermakov, Jan Cech, and Jiri Matas. Unlocking the hidden potential of clip in generalizable deepfake detection. arXiv preprint arXiv:2503.19683, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    Sadtalker: Learning realistic 3d motion coefficients for stylized audio- driven single image talking face animation

    Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio- driven single image talking face animation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8652–8661, 2023. 2

    work page 2023

  45. [45]

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

    Wenliang Zhao, Yongming Rao, Weikang Shi, Zuyan Liu, Jie Zhou, and Jiwen Lu. Diffswap: High-fidelity and controllable face swapping via 3d-aware masked diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8568–8577, 2023. 2 10

    work page 2023