Pith. sign in

REVIEW 4 major objections 6 minor 34 references

This paper argues that deepfake detection and explanation are separate capabilities, showing on a 200K-image benchmark that the most accurate classifier did not win because its explanations were less grounded in the correct visual evidence.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 08:42 UTC pith:CVM4EYYV

load-bearing objection A solid, reproducible challenge report whose central 'detection vs explanation' claim is plausible but leans on an unvalidated LLM grounding metric and a top-two gap the size of noise. the 4 major comments →

arxiv 2607.21007 v1 pith:CVM4EYYV submitted 2026-07-23 cs.CV

Explainable Deepfake Detection Challenge

Abhijeet Narang , Kartik Kuckreja , Shreya Ghosh , Muhammad Haris Khan , Usman Tariq , Jianfei Cai , Abhinav Dhall This is my paper
classification cs.CV
keywords explainable deepfake detectionbenchmarknatural-language explanationentity groundingevidence groundingvision-language modelsevaluation protocolclassification
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.

This paper reports on a challenge that asks deepfake detectors to do two things at once: label an image real or fake and explain, in natural language, the visual evidence behind that label—once in technical detail for experts and once simply for general readers. The hidden 200K-image test is scored by combining detection accuracy, semantic similarity, readability, and grounding measures that check whether explanations name the correct manipulated entities and supporting visual cues. The central result is that detection and explanation are distinct capabilities: the submission with the strongest detection macro-F1 (0.9479) finished second overall, while the winning entry (0.7612) had the best explanation-grounding scores. A sympathetic reader would take this as evidence that future deepfake systems should be optimized and evaluated jointly, not for classification alone.

Core claim

The challenge claims that a useful deepfake detector must justify its verdict, and that this justification can be measured. On 200,000 held-out images, the official score M_final = (D_F1 + E_exp)/2 equally weights detection macro-F1 and an explanation subscore that blends BERTScore against reference explanations, a simplicity estimate, and two LLM-derived grounding metrics: entity grounding (does the explanation identify the right manipulated object or region?) and evidence grounding (does it cite the right visual cues?). The results show that the best classifier by macro-F1 (0.9479) came second overall because its explanations scored lower on grounding, while the top-ranked system (0.7612)

What carries the argument

The load-bearing object is the two-component score M_final = (D_F1 + E_exp)/2. Detection is macro-F1 over the real/fake classes. The explanation subscore combines BERTScore-F1 against reference explanations, the normalized Simplicity Level Estimate for the simple explanation, and two LLM-based grounding scores (EntityScore and EvidenceScore) computed by extracting diagnostic entities and evidence claims from submitted and reference explanations, then checking bidirectional semantic support with a covariance function. These grounding scores do the argumentative work: they are what separate the best classifier from the best overall system.

Load-bearing premise

The entire ranking depends on trusting that the reference explanations and the LLM-based grounding evaluator actually measure explanation quality; the paper concedes the grounding evaluation's reliability still needs improvement, and a noisy judge could change the order of finishers.

What would settle it

Take a random sample of, say, 5,000 test images and have human annotators judge whether each submitted complex explanation names the correct manipulated entity and the correct visual evidence, without seeing the LLM scores. If human-grounded quality correlates with detection accuracy rather than with the LLM EntityScore/EvidenceScore, or if the humans re-rank the top teams, then the paper's conclusion that detection and explanation decouple is an artifact of the automated judge rather than a fact about the systems.

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

If this is right

  • Challenge leaderboards should evaluate explanation grounding jointly with classification, not detection alone.
  • Systems that optimize only detection risk ranking below systems with modestly lower accuracy but better-grounded explanations.
  • The grounding metrics reward explanations that name the correct manipulated entity and visual cue, not merely sentences that paraphrase the reference.
  • The winning entry's combination of a strong detection ensemble with an explanation model tuned on evidence-grounded rewards suggests a productive training direction.
  • The paper states that grounding evaluation reliability and efficiency remain open problems, pointing to future benchmark improvements.

Where Pith is reading between the lines

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

  • The decoupling of detection and explanation suggests that training detectors to localize evidence could improve classification on out-of-distribution manipulations, a possibility the paper does not explore.
  • The LLM-based grounding judge itself could be audited: if human raters re-rank explanations differently from the LLM, the 'distinct capabilities' conclusion would shift, making a human study the natural next test.
  • The same evaluation recipe—label plus dual-audience explanation, with grounding metrics—could transfer to other verification tasks such as medical imaging or document forensics, where explanations serve both experts and laypeople.
  • Since the best detector was not the best explainer, there may be an exploitable Pareto frontier between accuracy and grounded explanation; future work could map this trade-off explicitly.

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

4 major / 6 minor

Summary. The paper describes the Explainable Deepfake Detection Challenge at ACM Multimedia 2026, built on the authors' XPlainVerse benchmark. Participants submit a real/fake label plus two natural-language explanations (complex and simple) per image. The evaluation protocol combines macro-F1 detection with BERTScore semantic similarity, SLE simplicity, and LLM-based entity/evidence grounding scores (Eqs. 1–10). Results are reported for two baselines and five finalist teams on a 200K hidden test set. Pixel Sleuth ranks first (M_final=0.7612) while Team Antvengers has the highest detection macro-F1 (0.9479) but lower explanation-grounding scores. The paper concludes that detection and explanation are distinct capabilities.

Significance. If the evaluation protocol is reliable, the challenge provides a valuable community resource for joint deepfake detection and explanation generation: the hidden test is large, the participants are external, the evaluation script and baselines are released, and the equations in Section 5 are internally consistent. The finding that the strongest detector does not win the joint benchmark is an important, non-obvious outcome. However, the central claim depends heavily on an LLM-based grounding metric (G_ent/G_evid) that is not validated against human judgment, on hand-chosen weights, and on single-run scores without confidence intervals. These issues must be addressed before the 'distinct capabilities' conclusion can be considered established.

major comments (4)
  1. [Section 5, Eqs. 7–10; Section 8] The paper's central conclusion relies on the LLM grounding scores G_ent and G_evid, but these are not validated against human judgments. The top-two gap in M_final is only 0.0063 (Pixel Sleuth 0.7612 vs. Antvengers 0.7549). The E_LLM difference (0.02335) contributes about 0.0140 to the 0.0182 E_exp difference, i.e., roughly 77% of the explanation-score gap that decides the ranking. The covariance decisions c_p/c_r in Eq. (7) are binary LLM semantic-support calls with no reported human agreement, and the paper itself concedes 'future work should improve the reliability and efficiency of grounding evaluation.' If the LLM judge is noisy or biased, the observed ordering—and the claimed distinction between detection and explanation—could invert. Please provide a human validation study on a sample of the test explanations, or at minimum a sensitivity analysis using multiple LLM judges, differe
  2. [Table 3; Section 8] The claim that 'results demonstrate that detection and explanation are distinct capabilities' is too strong for five finalist submissions with no confidence intervals or significance testing. D_F1 differs by 0.0055 between Antvengers and Pixel Sleuth, while the grounding metrics differ by larger amounts, but there is no statistical analysis showing these differences are reliable. Also, the criteria for selecting the five 'finalist submissions' are not stated. Please report bootstrap or per-image paired confidence intervals for D_F1, E_exp, G_ent, G_evid, and M_final, and state how the finalists were chosen (e.g., top five on the public leaderboard). If the ranking is not robust, the conclusion should be tempered to 'in this challenge, the highest-detection team did not rank first.'
  3. [Section 3.1; Reference [23]] The benchmark dataset, the reference explanations, and the grounding evaluation metric all come from the authors' own XPlainVerse paper [23]. This creates a circularity risk: methods may be rewarded for matching the style of the references produced by the authors' pipeline rather than for producing explanations that humans judge as grounded or useful. Because the central claim is about explanation quality, independent validation is essential. Please report human evaluation of explanation quality (e.g., entity correctness, evidence correctness, usefulness) on a sample of submissions and correlate it with G_ent/G_evid and E_ref. This is a load-bearing issue, since the current evidence for 'distinct capabilities' is entirely mediated by the authors' own evaluator.
  4. [Section 5, Eqs. 4–10] The evaluation contains several hand-chosen weights and thresholds: the 0.7/0.3 split in Eq. (5), the 0.4/0.6 split in Eq. (10), the equal weighting of D_F1 and E_exp, and the SLE clipping range [-1,4] in Eq. (4). These are not derived from data or justified with sensitivity analysis. Since the final ranking and the conclusion depend on these choices, please show that the ranking and the 'distinct capabilities' claim are robust across a reasonable range of weight values, or provide a principled justification for the chosen values. If the ranking is highly sensitive to these parameters, that should be acknowledged as a limitation.
minor comments (6)
  1. [Section 5, Eq. (4)] The SLE clip range [-1,4] is not justified. Please cite the SLE scale or provide empirical support for this normalization.
  2. [Section 7] The paper does not explain how the five finalist submissions were selected from the 138 registered teams. Please state the selection rule (e.g., top on public leaderboard, deadline compliance, code availability).
  3. [Section 6] The LoRA fine-tuning description lacks hyperparameter details (rank, alpha, learning rate, number of steps). Please add them for reproducibility.
  4. [Section 3.1] The phrase 'adversarial perturbations' is used without specifying the attack method or budget. Please clarify what perturbations were applied and whether they are included in the hidden test.
  5. [Section 7] The platform is referred to as 'CodaBench'; please include a URL or citation, and clarify whether this is the same as CodaLab.
  6. [Section 6; Table 2] The baselines achieve much lower grounding scores than the finalists. A brief discussion of this gap (e.g., overfitting or task difficulty) would help readers calibrate the numbers.

Circularity Check

0 steps flagged

No significant circularity: leaderboard results are externally computed from hidden-test submissions, and the metric choices are explicit evaluation definitions rather than fitted predictions.

full rationale

This paper is a challenge report, not a derivation. The final ranking is computed by applying the explicitly stated protocol (Section 5, Equations 1-10) to external teams' submissions on a 200K-image hidden test set; Section 7 states that results were 'recomputed by the organizers on the full 200K-image hidden test set using the evaluation protocol described in Section 5.' No model parameter or metric weight is fitted to the test set, and no paper output is a renamed version of an input. The grounding metrics (Equations 7-9) and reference explanations are imported from the authors' own XPlainVerse paper [23], and the score weights (0.7/0.3, 0.4/0.6) are hand-chosen; these are transparent evaluation-design choices, not steps that make a derived quantity equal to its input by construction. The conclusion that detection and explanation are distinct capabilities is an interpretation of the externally produced leaderboard under that protocol. Concerns about LLM-evaluator reliability and the narrow margins between top teams—acknowledged by the paper's statement that 'future work should improve the reliability and efficiency of grounding evaluation'—are external-validity and noise concerns, not circularity. No specific equation reduces to itself or to a fitted parameter, so no circular step can be exhibited.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central evaluation rests on the validity of reference explanations, the LLM judge, and hand-chosen aggregation weights. These are assumed rather than demonstrated, making them the main epistemic cost of the paper.

free parameters (4)
  • Simple explanation score weighting = 0.7 BERTScore / 0.3 SLE
    Eq. 5 chooses these weights by hand; the readability-vs-fidelity tradeoff is not justified and affects E_s.
  • Explanation subscore weighting = 0.4 E_ref / 0.6 E_LLM
    Eq. 10 gives grounding metrics more weight than reference similarity; no sensitivity analysis is provided.
  • Final score weighting = 0.5 D_F1 / 0.5 E_exp
    Eq. 10 weights detection and explanation equally; alternative weightings could change the ranking.
  • SLE clip range = [-1, 4] mapped to [0,1]
    Eq. 4 uses an arbitrary clip range that changes the readability contribution to E_s.
axioms (4)
  • domain assumption XPlainVerse reference explanations are correct and grounded
    Section 3.1 says an edit-aware quality-control pipeline validates references, but no human agreement or external audit is reported.
  • domain assumption The LLM evaluator’s semantic coverage decisions are reliable
    Section 5, Eqs. 7-9, uses an LLM to decide whether entities/evidence in one explanation are supported by the other; no human validation is presented.
  • domain assumption BERTScore and SLE capture explanation quality
    Section 5 treats BERTScore-F1 as semantic fidelity and SLE as readability; SLE was trained on sentence simplification, not deepfake explanations.
  • domain assumption Hidden test split is free of leakage from public splits
    Section 3.2 describes split sizes but does not describe contamination controls or duplicate-image removal.

pith-pipeline@v1.3.0-alltime-deepseek · 8827 in / 14117 out tokens · 149451 ms · 2026-08-01T08:42:13.537443+00:00 · methodology

0 comments
read the original abstract

Deepfake detection is moving beyond binary classification decisions toward systems that can also explain the visual evidence supporting those decisions. This transition is important for real-world verification settings, where diverse users need to understand not only whether an image is manipulated, but also why it is considered suspicious. The Explainable Deepfake Detection Challenge at ACM Multimedia 2026 is designed to benchmark this joint capability. Built on XPlainVerse, a million-scale benchmark for explainable deepfake detection, the challenge evaluates methods on image classification and grounded natural-language explanation generation. Participants submit a real/fake label together with two explanations for each image: a detailed complex explanation for technical users and a concise simple explanation for general users. The evaluation combines classification metrics with semantic similarity, simplicity, and intent-aware grounding metrics that assess whether explanations identify the relevant manipulated entities and supporting visual evidence. The methodologies developed through the challenge will contribute to the development of next-generation explainable deepfake detectors. Evaluation script, baseline models, and accompanying code are available on https://github.com/Abhijeet8901/XPlainVerse-ACMChallenge.

Figures

Figures reproduced from arXiv: 2607.21007 by Abhijeet Narang, Abhinav Dhall, Jianfei Cai, Kartik Kuckreja, Muhammad Haris Khan, Shreya Ghosh, Usman Tariq.

Figure 1
Figure 1. Figure 1: Examples where semantic similarity misses ground [PITH_FULL_IMAGE:figures/full_fig_p003_1.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

34 extracted references · 1 canonical work pages

  1. [1]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    Pith/arXiv arXiv 2026

  2. [2]

    Zhixi Cai, Abhinav Dhall, Shreya Ghosh, Munawar Hayat, Dimitrios Kollias, Kalin Stefanov, and Usman Tariq. 2024. 1M-Deepfakes Detection Challenge. In Proceedings of the 32nd ACM International Conference on Multimedia. 11355–11359. doi:10.1145/3664647.3689145

    arXiv 2024

  3. [3]

    Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, and Kalin Stefanov. 2024. AV-Deepfake1M: A Large-Scale LLM- Driven Audio-Visual Deepfake Dataset. InProceedings of the 32nd ACM Interna- tional Conference on Multimedia. 7414–7423. doi:10.1145/3664647.3680795

    arXiv 2024

  4. [4]

    Zhixi Cai, Kartik Kuckreja, Shreya Ghosh, Akanksha Chuchra, Muhammad Haris Khan, Usman Tariq, Tom Gedeon, and Abhinav Dhall. 2025. AV-Deepfake1M++: A Large-Scale Audio-Visual Deepfake Benchmark with Real-World Perturbations. InProceedings of the 33rd ACM International Conference on Multimedia. 13686– 13691. doi:10.1145/3746027.3761979

    arXiv 2025

  5. [5]

    Liam Cripwell, Joël Legrand, and Claire Gardent. 2023. Simplicity Level Estimate (SLE): A Learned Reference-Less Metric for Sentence Simplification. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Singapore, 12053–12059. doi:10.18653/ v1/2023.emnlp-main.739

    2023

  6. [6]

    Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. 2020. The DeepFake Detection Challenge (DFDC) Dataset. arXiv:2006.07397 [cs.CV]

    Pith/arXiv arXiv 2020

  7. [7]

    Xiao Guo, Xiufeng Song, Yue Zhang, Xiaohong Liu, and Xiaoming Liu. 2025. Rethinking Vision-Language Model in Face Forensics: Multi-Modal Interpretable Forged Face Detector. InProceedings of the Computer Vision and Pattern Recogni- tion Conference (CVPR). 105–116

    2025

  8. [8]

    Parul Gupta, Shreya Ghosh, Tom Gedeon, Thanh-Toan Do, and Abhinav Dhall

  9. [9]

    Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. 2021. ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 4360–4369. doi:10.1109/ CVPR46437.2021.00434

    arXiv 2021

  10. [10]

    Vlad Hondru, Eduard Hogea, Darian Onchis, and Radu Tudor Ionescu. 2026. ExDDV: A New Dataset for Explainable Deepfake Detection in Video. InPro- ceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). 4273–4284

    2026

  11. [11]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations. https: //openreview.net/forum?id=nZeVKeeFYf9

    2022

  12. [12]

    Zhenglin Huang, Jinwei Hu, Xiangtai Li, Yiwei He, Xingyu Zhao, Bei Peng, Baoyuan Wu, Xiaowei Huang, and Guangliang Cheng. 2025. SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR). 28831–28841

    2025

  13. [13]

    Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. 2020. DeeperForensics-1.0: A Large-Scale Dataset for Real-World Face Forgery De- tection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2889–2898

    2020

  14. [14]

    Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Zou, Ling Lo, Sheng-Ping Yang, Yu- Wen Tseng, Kun-Hsiang Lin, Chia-Ling Chen, Yu-Ting Ta, Yan-Tsung Wang, Po-Ching Chen, Hongxia Xie, Hong-Han Shuai, and Wen-Huang Cheng. 2026. TriDF: Evaluating Perception, Detection, and Hallucination for Interpretable DeepFake Detection. InProceedings of the IEEE/CVF Conference o...

    2026

  15. [15]

    Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S. Woo. 2021. FakeAVCeleb: A Novel Audio-Video Multimodal Deepfake Dataset. InThirty-Fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    2021

  16. [16]

    Pavel Korshunov and Sébastien Marcel. 2019. Vulnerability Assessment and Detection of Deepfake Videos. In2019 International Conference on Biometrics (ICB). 1–6. doi:10.1109/ICB45273.2019.8987375

    arXiv 2019

  17. [17]

    Tianxiao Li, Zhenglin Huang, Haiquan Wen, Yiwei He, Xinze Li, Bingyu Zhu, Wuhui Duan, Congang Chen, Zeyu Fu, Yi Dong, Baoyuan Wu, Xiangtai Li, and Guangliang Cheng. 2026. Omni-Fake: Benchmarking Unified Multimodal Social Media Deepfake Detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 30299–30311

    2026

  18. [18]

    Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. 2020. Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 3207–3216

    2020

  19. [19]

    Vera Liao, Daniel Gruen, and Sarah Miller

    Q. Vera Liao, Daniel Gruen, and Sarah Miller. 2020. Questioning the AI: Inform- ing Design Practices for Explainable AI User Experiences. InProceedings of the 2020 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery, New York, NY, USA, 1–15. doi:10.1145/3313831.3376590

    arXiv 2020

  20. [20]

    Tim Miller. 2019. Explanation in Artificial Intelligence: Insights from the Social Sciences.Artificial Intelligence267 (2019), 1–38. doi:10.1016/j.artint.2018.07.007

    doi:10.1016/j.artint.2018.07.007 2019

  21. [21]

    Yisroel Mirsky and Wenke Lee. 2021. The Creation and Detection of Deepfakes: A Survey.Comput. Surveys54, 1, Article 7 (2021), 41 pages. doi:10.1145/3425780

    doi:10.1145/3425780 2021

  22. [22]

    Md Sarfaraz Momin, Abu Sufian, Debaditya Barman, Marco Leo, Cosimo Distante, and Naser Damer. 2025. Explainable Deepfake Detection across Different Modal- ities: An Overview of Methods and Challenges.Image and Vision Computing163 (2025), 105738. doi:10.1016/j.imavis.2025.105738

    arXiv 2025

  23. [23]

    Abhijeet Narang, Kartik Kuckreja, Shreya Ghosh, Muhammad Haris Khan, Jianfei Cai, and Abhinav Dhall. 2026. XPlainVerse: A Million-Scale Benchmark for Explainable Deepfake Detection. arXiv:2607.03562 [cs.CV] https://arxiv.org/abs/ 2607.03562

    Pith/arXiv arXiv 2026

  24. [24]

    Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. 2019. FaceForensics++: Learning to Detect Manipulated Facial Images. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 1–11. doi:10.1109/ICCV.2019.00009

    arXiv 2019

  25. [25]

    Kacper Sokol and Peter Flach. 2020. One Explanation Does Not Fit All: The Promise of Interactive Explanations for Machine Learning Transparency.KI – Künstliche Intelligenz34, 2 (2020), 235–250. doi:10.1007/s13218-020-00637-y

    work page doi:10.1007/s13218-020-00637-y 2020

  26. [26]

    Ke Sun, Shen Chen, Taiping Yao, Ziyin Zhou, Jiayi Ji, Xiaoshuai Sun, Chia-Wen Lin, and Rongrong Ji. 2025. Towards General Visual-Linguistic Face Forgery Detection. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR). 19576–19586

    2025

  27. [27]

    Woo, Priyanka Singh, Irena Irmalasari, Saakshi Gupta, and Dev Gupta

    Shahroz Tariq, Simon S. Woo, Priyanka Singh, Irena Irmalasari, Saakshi Gupta, and Dev Gupta. 2025. From Prediction to Explanation: Multimodal, Explainable, and Interactive Deepfake Detection Framework for Non-Expert Users. InPro- ceedings of the 33rd ACM International Conference on Multimedia. 11716–11725. doi:10.1145/3746027.3755786

    arXiv 2025

  28. [28]

    Konstantinos Tsigos, Evlampios Apostolidis, Spyridon Baxevanakis, Symeon Papadopoulos, and Vasileios Mezaris. 2024. Towards Quantitative Evaluation of Explainable AI Methods for Deepfake Detection. InProceedings of the 3rd ACM International Workshop on Multimedia AI against Disinformation. 37–45. doi:10.1145/3643491.3660292

    arXiv 2024

  29. [29]

    Luisa Verdoliva. 2020. Media Forensics and DeepFakes: An Overview.IEEE Journal of Selected Topics in Signal Processing14, 5 (Aug. 2020), 910–932. doi:10. 1109/JSTSP.2020.3002101

    arXiv 2020

  30. [30]

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, Zhaokai Wang, Zhe Chen, Hongjie Zhang, Ganlin Yang, Haomin Wang, Qi Wei, Jinhui Yin, Wenhao Li, Erfei Cui, Guanzhou Chen, Zichen Ding, Changyao Tian, Zhenyu Wu, Jingjing Xie, Zehao Li, Bowen Yang, Yuchen Duan, Xuehui Wang, Zhi Hou,...

    Pith/arXiv arXiv 2025

  31. [31]

    Xin Yang, Yuezun Li, and Siwei Lyu. 2019. Exposing Deep Fakes Using Inconsis- tent Head Poses. In2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 8261–8265. doi:10.1109/ICASSP.2019.8683164

    arXiv 2019

  32. [32]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi

  33. [2020]

    InInternational Confer- ence on Learning Representations

    BERTScore: Evaluating Text Generation with BERT. InInternational Confer- ence on Learning Representations. https://openreview.net/forum?id=SkeHuCVFDr

  34. [2025]

    arXiv:2506.00868 [cs.MM] https: //arxiv.org/abs/2506.00868

    Multiverse Through Deepfakes: The MultiFakeVerse Dataset of Person- Centric Visual and Conceptual Manipulations. arXiv:2506.00868 [cs.MM] https: //arxiv.org/abs/2506.00868

    Pith/arXiv arXiv