Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

LLMs Are Not Yet Ready for Deepfake Image Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that four leading vision-language models, tested zero-shot on 100 curated face images, are not yet dependable as standalone deepfake detectors even though they produce coherent explanations.

desk verdict A fair question and a plausible negative answer, but the accuracy numbers rest on an undisclosed scoring rule for abstentions and a very small sample. read the letter →

arxiv 2506.10474 v1 pith:B2DGWJLQ submitted 2025-06-12 cs.CV

Shahroz Tariq , David Nguyen , M.A.P. Chamikara , Tingmin Wu , Alsharif Abuadbba , Kristen Moore This is my paper
classification cs.CV
keywords DeepfakeLLMsVLMsForensicsGenerativeAISyntheticMediaLargeLanguageModelFoundational
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether general-purpose vision-language models can serve as standalone deepfake detectors without any training. The authors test ChatGPT, Claude, Gemini, and Grok in a zero-shot setting on 100 carefully selected images: real faces from regular, artistic, and studio settings, plus face-swapped, re-enacted, GAN-generated, and diffusion-generated fakes. They find that no model reaches the accuracy of specialised detectors, and each model has a characteristic failure mode, such as Grok trusting realistic backgrounds over distorted faces, Gemini over-predicting the 'real' class, and ChatGPT treating vintage-style synthetic images as authentic. The paper concludes that current VLMs are not yet dependable as autonomous detectors but are useful as explainable assistants that surface cues for human reviewers. The practical import is that these models belong in hybrid, human-in-the-loop forensic workflows rather than in fully automated detection pipelines.

What carries the argument

The carrying mechanism is a compact, hand-curated benchmark of 100 images, stratified by gender and perceptual difficulty, that spans three settings for real photos (regular, artistic, studio) and three deepfake families (faceswap, reenactment, and GAN/diffusion synthesis). Each of the four models is given the same zero-shot prompt, a persona instruction followed by a structured task prompt, and must answer in a fixed format covering decision, confidence, reasoning, and uncertainty. This uniform protocol makes the model outputs comparable and lets the authors separate classification errors from explanation quality. The trigger-pattern probing is a second component: after observing ChatGPT's vintage-image bias, the authors collected additional vintage synthetic samples and showed the misclassification reproduced, which gives the style-bias claim its evidential weight.

What would settle it

Run the same four models with the same prompt on a larger, randomly sampled benchmark—for example, 1,000 images drawn uniformly from FF++, DFDC, CelebDF, R3GAN, and current diffusion platforms, without hand-picking edge cases or adding sets after inspecting model behaviour. If any model maintains above 90 percent accuracy on both real and fake classes across repeated queries, then the paper's 'not yet dependable as standalone detectors' claim would no longer hold for that model.

Watch

Extended reading notes

Core claim

The paper's central claim, stated on its own terms, is that while current proprietary VLMs can produce coherent natural-language explanations and catch surface-level anomalies, their classifications are not reliable enough to serve as a strong signal for deepfake detection. Across the three deepfake categories—faceswap, reenactment, and synthetic generation—accuracy varies widely by model and image style, and repeated queries to the same model sometimes return different labels or confidence scores for the same image. The authors identify style-based trigger patterns as the clearest evidence of this unreliability: ChatGPT consistently labels vintage-styled diffusion images as real, and Gemini labels hyper-realistic studio-quality fakes as real because it is biased toward the 'real' class. They conclude that the strengths of VLMs lie in interpretability and contextual reasoning, making them candidates for collaboration with human analysts, not replacements for specialised detectors.

Load-bearing premise

The load-bearing premise is that 100 hand-picked images—about ten per class per source, chosen by the authors for diversity and difficulty, with a vintage set added after seeing ChatGPT's bias—are representative enough of real-world deepfakes and real photos that the measured accuracies support a general statement about VLM reliability.

Editorial extensions

If this is right

  • Organizations should not deploy current general-purpose VLMs as autonomous detectors in content moderation, legal review, or authentication pipelines.
  • A workable deployment is a hybrid pipeline in which a specialised detector scores the image and a VLM produces a natural-language rationale for a human reviewer to verify.
  • Fakes styled to look vintage or professionally polished are a systematic blind spot, so adversarial users could evade ChatGPT- and Gemini-style VLMs by applying retro or studio filters.
  • Non-deterministic answers across repeated queries mean VLM-assisted forensics needs logging and reproducibility controls before high-stakes use.
  • Validation on video, audio-visual, and open-source models is the immediate next step; the current verdict applies only to static images and four proprietary models.

Reading between the lines

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

  • A direct corollary the authors leave implicit is that aesthetic filters can be weaponised: apply a vintage grade (sepia, grain, soft focus) to any faceswap output and ChatGPT-style VLMs will likely call it real, an experiment a reader could run with paired images.
  • Because performance drops sharply between regular and artistic/studio real images, the paper's diagnosis implies VLMs lean on photographic-style priors rather than face-integrity cues; that predicts brittleness under compression, re-encoding, or social-media filters, none of which the paper tests.
  • If the diagnosis is right, a natural next benchmark is a hybrid detector that feeds VLM embeddings or explanations together with a specialised detector's score into a small classifier; recovering a large part of the accuracy gap would confirm the two signal types are complementary.
  • Since the benchmark contains 100 hand-selected images, the 'not ready' verdict is partially an artifact of sample size; a larger random sample is the fastest way to see whether the failure modes are stable or were amplified by the selection.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper evaluates four proprietary vision-language models (ChatGPT/GPT-4o, Claude/Sonnet 4, Gemini 2.5 Flash, Grok 3) on zero-shot deepfake image detection. The authors construct a curated benchmark of 100 images spanning real photos (regular, artistic, studio), faceswap and reenactment frames from FF++, DFDC, and CelebDF, and synthetic faces from GAN (R3GAN) and diffusion sources. Using a structured prompt that asks each model to return a decision, confidence, and reasoning, they report accuracy/precision/recall/F1 per model and category, plus qualitative failure modes. The central claim is that current VLMs are not dependable as standalone detectors but may serve as assistive tools in human-in-the-loop forensic workflows, with particular failure modes including over-reliance on surface aesthetics and biases such as ChatGPT's misclassification of vintage-style images.

Significance. If substantiated, the paper provides a useful, timely snapshot of general-purpose VLM capabilities for media forensics. Its strengths include covering three deepfake taxonomies with four major commercial VLMs, a defined zero-shot prompting protocol, explicit collection of reasoning traces, and identification of concrete style-biased trigger patterns. It also promises dataset release. However, the quantitative evidence as presented is not yet sufficient to support the strength of the 'not dependable' conclusion: the sample per cell is only about 10 images, the handling of 'Inconclusive' outputs is undisclosed, no uncertainty quantification or statistical tests are reported, and no specialized-detector baseline is run on the same benchmark. These issues are fixable and the central claim is plausible, but they currently leave the performance numbers ambiguous.

major comments (4)
  1. [§2.2, §3.1, Tables 1–2] The manuscript never specifies how 'Inconclusive' responses from the forced-choice Decision field were scored, yet this decision rule directly determines every reported accuracy, precision, recall, and F1 value. Section 2.2 instructs models to return 'Decision: [Real / Fake / Inconclusive]' and states that 'we recorded the predicted class,' but Section 3.1 does not say whether 'Inconclusive' was counted as an error, excluded from the denominator, or mapped to a class. Because the paper's central claim is a negative capability claim, the scoring rule is load-bearing: a model that abstains on difficult inputs and is otherwise accurate could be usable in a human-in-the-loop setting yet appear 'not dependable' under an undisclosed scoring rule. Please state the exact handling of 'Inconclusive', report the per-model distribution of Decision values, and if any abstentions were excluded, also report the accuracy on the full set with abstentions counted as errors, so readers can evaluate both interpretations.
  2. [§2.1, §2.2, §3.1, Tables 1–2] The quantitative conclusions rest on very small per-cell samples (10 images per source per class, with some cells as small as 10 total) and no error bars, confidence intervals, or statistical tests are provided. The paper itself notes in §2.2 and §4.2 that models are non-deterministic, so single-query point estimates are insufficient. For example, with 10 images per cell, a one-image flip changes accuracy by 10 percentage points, which is comparable to many of the model differences reported in Table 2 (e.g., Grok's 0.27 vs. Gemini's 0.27 average accuracy on synthesis). Please report per-image results, repeated queries for all or a representative subset, and either confidence intervals, bootstrap estimates, or an appropriate statistical test (e.g., McNemar's test for paired model comparisons).
  3. [Introduction, §3.1, Table 2] The Introduction claims that VLMs 'do not match the classification accuracy of specialised vision-based detectors' and §3.1 reiterates that none of the models achieve robust performance, but no specialized detector is evaluated on the same curated benchmark. The comparison relies on prior published results on different datasets and protocols (e.g., ref. [36]), which are not directly comparable to the 100-image set used here. Please add a same-benchmark baseline—for instance, a standard face-forensics CNN trained on FF++ or a pretrained deepfake detector from the literature—evaluated under the same real/fake split, or explicitly soften the comparative claim to avoid an unsupported contrast.
  4. [§3.2, Figure 3] The 'vintage trigger' finding for ChatGPT is presented as a key insight, but the text states that the additional vintage-style samples were collected after the authors observed ChatGPT's bias. This introduces selection bias: the trigger set is chosen from the distribution that the model is known to misclassify, so the 'all labelled as real' result is expected by construction. Please either describe the trigger set as an exploratory, post-hoc illustration (with the number of images, sources, and selection criteria transparently reported) or, if the claim is meant to establish a general style bias, add a pre-registered or independently sampled vintage test set that was not selected based on model behavior. Without this, the claim of a systematic 'style bias' is not supported.
minor comments (5)
  1. [§2.1] The citation 'Le at al. [36]' should be 'Le et al. [36]'.
  2. [Table 2] The column header 'TGID' appears to be a typo or abbreviation for the R3GAN source; the table should clearly state which source corresponds to the GAN column.
  3. [Tables 1–2] It is never stated whether precision and recall are computed with 'fake' or 'real' as the positive class. This makes the numbers ambiguous, especially in Table 1 where near-perfect precision on real images could trivially reflect a strong 'real' bias. Please define the positive class in the table captions or in the evaluation-procedure paragraph.
  4. [§2.1] The text says 'We handpicked samples from a diverse set of sources' for the real category, but the exact sampling procedure for Unsplash and Getty (e.g., search terms, filters, how gender balance was verified) is not described in enough detail for replication. The promised dataset release should include curation metadata.
  5. [Figure 2 and Figure 3] The captions for Figures 2 and 3 are minimal; consider adding a description of what each panel shows, including which model misclassified which image and what its stated decision was, so that the qualitative claims can be independently assessed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark measurements are external and are not entailed by the paper's assumptions or self-citations.

full rationale

The paper's central claim—that current VLMs are not yet dependable as standalone deepfake detectors—is an empirical negative result supported by direct zero-shot queries to four external proprietary VLMs on a hand-curated benchmark. There is no fitted parameter, no equation, and no quantity defined in terms of the conclusion; the reported accuracies are measurements of model outputs, not derivations from the paper's own assumptions. The numerous self-citations ([36], [13], [51]–[53]) supply a taxonomy, framing, prior detector comparisons, and a recommended human-AI collaboration direction, but none of these force or predict the measured accuracy values. The post-hoc collection of vintage-style samples after observing ChatGPT's bias (Section 3.2) is a methodological limitation of the exploratory trigger-pattern analysis rather than a circular step, because the vintage images were selected on style and the model's label was the measured outcome. The undisclosed treatment of the 'Inconclusive' decision option is a reporting and validity concern, not circularity. The evaluation is self-contained against external model behavior and public datasets, so the conclusion does not reduce to its own inputs.

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

There are no fitted free parameters and no invented entities in this empirical study. The central claims rely on benchmark representativeness, ground-truth correctness, prompt comparability, and model-version stability, which are listed as domain assumptions.

assumptions (4)
  • domain assumption Ground-truth labels for images from FF++, DFDC, CelebDF, R3GAN, and web sources are correct.
    Section 2.1 treats these dataset labels as the reference for computing accuracy, precision, recall, and F1. Mislabeled samples in the curated set would directly corrupt all reported metrics.
  • domain assumption 100 hand-picked images, about 10 per class per source, are representative enough to support claims about general VLM reliability.
    Section 2.1 says the sampling strategy emphasized diversity, gender balance, and perceptual difficulty to emulate realistic detection scenarios, but the selection was manual and included deliberately difficult edge cases. This representativeness is load-bearing for the 'not ready' conclusion.
  • domain assumption A unified persona prompt and structured output format produce comparable responses across the four proprietary VLMs.
    Section 2.2 uses one persona prompt plus a fixed task prompt. The authors acknowledge prompt sensitivity in Section 4.2, which means small prompt differences could change the measured accuracies.
  • domain assumption Web-interface outputs correspond to stable, identifiable model versions with deterministic-enough behavior for benchmarking.
    Section 2.2 names models such as GPT-4o and Gemini 2.5 Flash but does not pin access times, web UI versions, or sampling parameters. The authors report observed variability in repeated queries, so the measurements are model-instance dependent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLMs Are Not Yet Ready for Deepfake Image Detection." pith.science (2026). https://pith.science/paper/B2DGWJLQ

@misc{pith2026250610474,
  author       = {Pith},
  title        = {Pith review of: LLMs Are Not Yet Ready for Deepfake Image Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2DGWJLQ}},
  note         = {Machine review of arXiv:2506.10474}
}
read the original abstract

The growing sophistication of deepfakes presents substantial challenges to the integrity of media and the preservation of public trust. Concurrently, vision-language models (VLMs), large language models enhanced with visual reasoning capabilities, have emerged as promising tools across various domains, sparking interest in their applicability to deepfake detection. This study conducts a structured zero-shot evaluation of four prominent VLMs: ChatGPT, Claude, Gemini, and Grok, focusing on three primary deepfake types: faceswap, reenactment, and synthetic generation. Leveraging a meticulously assembled benchmark comprising authentic and manipulated images from diverse sources, we evaluate each model's classification accuracy and reasoning depth. Our analysis indicates that while VLMs can produce coherent explanations and detect surface-level anomalies, they are not yet dependable as standalone detection systems. We highlight critical failure modes, such as an overemphasis on stylistic elements and vulnerability to misleading visual patterns like vintage aesthetics. Nevertheless, VLMs exhibit strengths in interpretability and contextual analysis, suggesting their potential to augment human expertise in forensic workflows. These insights imply that although general-purpose models currently lack the reliability needed for autonomous deepfake detection, they hold promise as integral components in hybrid or human-in-the-loop detection frameworks.

Figures

Figures reproduced from arXiv: 2506.10474 by the authors.

Figure 1
Figure 1. Evaluation Datasets. We evaluated 4 leading LLM-based vision-language models from ChatGPT, Gemini, Grok, and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Failure cases of different models. Here, GT denotes [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Trigger images for ChatGPT. 3.2 Qualitative Analysis Beyond raw performance metrics, we conducted a qualitative anal￾ysis of model outputs to better understand their reasoning patterns, failure modes, and potential for human-AI collaboration. Failure Modes and Misclassification Patterns. Our anal￾ysis revealed several recurring failure patterns across models: • Surface-level reasoning: Grok frequently misclassified … view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. From Prediction to Explanation: Multimodal, Explainable, and Interactive Deepfake Detection Framework for Non-Expert Users

    cs.CV 2025-08 conditional novelty 5.0 of 10

    A pipeline combining a deepfake classifier, Grad-CAM heatmaps, image captioning, and an LLM generates layered explanations of deepfake verdicts for non-expert users.

  2. On the Reliability of Vision-Language Models Under Adversarial Frequency-Domain Perturbations

    cs.CV 2025-07 conditional novelty 5.0 of 10

    Vision-language model judgments about image realism and generated captions can be shifted by imperceptible perturbations confined to specific spatial frequency bands, even under black-box access.

  3. The Age of Sensorial Zero Trust: Why We Can No Longer Trust Our Senses

    cs.CR 2025-07 unverdicted novelty 3.0 of 10

    A position paper argues that organizations should apply Zero Trust verification principles to human sensory perception to defend against deepfake and voice-clone fraud.

Reference graph

Works this paper leans on

56 extracted references · 48 canonical work pages · cited by 3 Pith papers

  1. [36]

    Le, Jiwon Kim, Simon S

    Binh M. Le, Jiwon Kim, Simon S. Woo, Kristen Moore, Alsharif Abuadbba, and Shahroz Tariq. Sok: Systematization and benchmarking of deepfake detectors in a unified framework, 2025

    work page 2025

  2. [1]

    Deepfakes and stolen pii utilized to apply for remote work positions

    FBI. Deepfakes and stolen pii utilized to apply for remote work positions. https: //www.ic3.gov/Media/Y2022/PSA220628, June 2022. Accessed: 2022-07-01

    work page 2022

  3. [2]

    Deepfake in the metaverse: Security implications for virtual gaming, meetings, and offices

    Shahroz Tariq, Alsharif Abuadbba, and Kristen Moore. Deepfake in the metaverse: Security implications for virtual gaming, meetings, and offices. InProceedings of the 2nd Workshop on Security Implications of Deepfakes and Cheapfakes, WDC ’23, page 16–19, New York, NY, USA, 2023. Association for Computing Machinery

    work page 2023

  4. [3]

    Muhammad Shahid Muneer and Simon S. Woo. Towards safe synthetic image generation on the web: A multimodal robust nsfw defense and million scale dataset. InCompanion Proceedings of the ACM Web Conference 2025 (WWW Companion ’25). ACM, April 2025

    work page 2025

  5. [4]

    Classifying genuine face images from disguised face images

    Junyaup Kim, Siho Han, and Simon S Woo. Classifying genuine face images from disguised face images. In2019 IEEE International Conference on Big Data (Big Data), pages 6248–6250. IEEE, 2019

    work page 2019

  6. [5]

    Seeing is living? rethinking the security of facial liveness verification in the deepfake era

    Changjiang Li, Li Wang, Shouling Ji, Xuhong Zhang, Zhaohan Xi, Shanqing Guo, and Ting Wang. Seeing is living? rethinking the security of facial liveness verification in the deepfake era. In31st USENIX Security Symposium (USENIX Security 22), 2022

    work page 2022

  7. [6]

    Evaluating trustworthiness and racial bias in face recognition apis using deepfakes.Computer, 56(5):51–61, 2023

    Shahroz Tariq, Sowon Jeon, and Simon S Woo. Evaluating trustworthiness and racial bias in face recognition apis using deepfakes.Computer, 56(5):51–61, 2023

    work page 2023

  8. [7]

    When george clooney is not george clooney: Using genattack to deceive amazon’s and naver’s celebrity recognition apis

    Keeyoung Kim and Simon S Woo. When george clooney is not george clooney: Using genattack to deceive amazon’s and naver’s celebrity recognition apis. In IFIP International Conference on ICT Systems Security and Privacy Protection, pages 355–369. Springer, 2018

    work page 2018

Show all 56 references
  1. [8]

    Am i a real or fake celebrity? evaluating face recognition and verification apis under deepfake impersonation attack

    Shahroz Tariq, Sowon Jeon, and Simon S Woo. Am i a real or fake celebrity? evaluating face recognition and verification apis under deepfake impersonation attack. InProceedings of the ACM Web Conference 2022, pages 512–523, 2022

    2022

  2. [9]

    Le and Simon Woo

    Binh M. Le and Simon Woo. Quality-agnostic deepfake detection with intra- model collaborative learning. InProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), 2023

    2023

  3. [10]

    Self-supervised video forensics by audio-visual anomaly detection

    Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  4. [11]

    Aunet: Learning relations between action units for face forgery detection

    Weiming Bai, Yufan Liu, Zhipeng Zhang, Bing Li, and Weiming Hu. Aunet: Learning relations between action units for face forgery detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  5. [12]

    Capsule-forensics: Using capsule networks to detect forged images and videos

    Huy H Nguyen, Junichi Yamagishi, and Isao Echizen. Capsule-forensics: Using capsule networks to detect forged images and videos. InICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019

    2019

  6. [13]

    One detector to rule them all: Towards a general deepfake attack detection framework

    Shahroz Tariq, Sangyup Lee, and Simon Woo. One detector to rule them all: Towards a general deepfake attack detection framework. InProceedings of the web conference 2021, pages 3625–3637, 2021

    2021

  7. [14]

    Apparatus and method for detecting deepfake based on convolutional long short-term memory network, September 2023

    Simon Sungil Woo, Sang Yup Lee, and Shahroz Tariq. Apparatus and method for detecting deepfake based on convolutional long short-term memory network, September 2023. US Patent App. 18/114,416

    2023

  8. [15]

    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. Springer, 2020

    2020

  9. [16]

    Adaptive face forgery detection in cross domain

    Luchuan Song, Zheng Fang, Xiaodan Li, Xiaoyi Dong, Zhenchao Jin, Yuefeng Chen, and Siwei Lyu. Adaptive face forgery detection in cross domain. In European Conference on Computer Vision. Springer, 2022

    2022

  10. [17]

    Add: Frequency attention and multi-view based knowledge distillation to detect low-quality compressed deepfake images

    Binh M Le and Simon S Woo. Add: Frequency attention and multi-view based knowledge distillation to detect low-quality compressed deepfake images. In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence, 2022

    2022

  11. [18]

    Exploring the asynchronous of the frequency spectra of gan-generated facial images

    Binh M Le and Simon S Woo. Exploring the asynchronous of the frequency spectra of gan-generated facial images. InCEUR Workshop Proceedings, volume

  12. [19]

    Tar: Generalized forensic framework to detect deepfakes using weakly supervised learning

    Sangyup Lee, Shahroz Tariq, Junyaup Kim, and Simon S Woo. Tar: Generalized forensic framework to detect deepfakes using weakly supervised learning. InIFIP International Conference on ICT Systems Security and Privacy Protection, pages 351–366. Springer, 2021

    2021

  13. [20]

    Detecting hand- crafted facial image manipulations and gan-generated facial images using shallow-fakefacenet.Applied Soft Computing, 105:107256, 2021

    Sangyup Lee, Shahroz Tariq, Youjin Shin, and Simon S Woo. Detecting hand- crafted facial image manipulations and gan-generated facial images using shallow-fakefacenet.Applied Soft Computing, 105:107256, 2021

    2021

  14. [21]

    Ugad: Universal generative ai detector utilizing frequency fingerprints

    Inzamamul Alam, Muhammad Shahid Muneer, and Simon S Woo. Ugad: Universal generative ai detector utilizing frequency fingerprints. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 4332–4340, 2024

    2024

  15. [22]

    Altfreezing for more general video face forgery detection

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  16. [23]

    Learning on gradients: Generalized artifacts representation for gan-generated images detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  17. [24]

    Ptd: Privacy-preserving human face processing framework using tensor decomposition

    Jeongho Kim, Shahroz Tariq, and Simon S Woo. Ptd: Privacy-preserving human face processing framework using tensor decomposition. InProceedings of the 37th ACM/SIGAPP Symposium on Applied Computing, pages 1296–1303, 2022

    2022

  18. [25]

    Cored: Generalizing fake media detection with continual representation using distillation

    Minha Kim, Shahroz Tariq, and Simon S Woo. Cored: Generalizing fake media detection with continual representation using distillation. InProceedings of the 29th ACM International Conference on Multimedia, pages 337–346, 2021

    2021

  19. [26]

    Fretal: Generalizing deepfake de- tection using knowledge distillation and representation learning

    Minha Kim, Shahroz Tariq, and Simon S Woo. Fretal: Generalizing deepfake de- tection using knowledge distillation and representation learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001–1012, 2021

    2021

  20. [27]

    Detecting both machine and human created fake face images in the wild

    Shahroz Tariq, Sangyup Lee, Hoyoung Kim, Youjin Shin, and Simon S Woo. Detecting both machine and human created fake face images in the wild. In Proceedings of the 2nd international workshop on multimedia privacy and security, pages 81–87, 2018

    2018

  21. [28]

    Gan is a friend or foe?: a framework to detect various fake face images

    Shahroz Tariq, Sangyup Lee, Hoyoung Kim, Youjin Shin, and Simon S Woo. Gan is a friend or foe?: a framework to detect various fake face images. InProceedings of the 34th ACM/SIGAPP Symposium on Applied Computing, pages 1296–1303. ACM, 2019

    2019

  22. [29]

    A convolutional lstm based residual network for deepfake video detection.arXiv preprint arXiv:2009.07480, 2020

    Shahroz Tariq, Sangyup Lee, and Simon S Woo. A convolutional lstm based residual network for deepfake video detection.arXiv preprint arXiv:2009.07480, 2020

    2009 arXiv

  23. [30]

    Method for performing continual learning using representation learning and apparatus thereof, July

    Simon Sungil Woo, KIM Minha, and Shahroz Tariq. Method for performing continual learning using representation learning and apparatus thereof, July

  24. [31]

    Faceforensics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. InProceedings of the IEEE/CVF international conference on computer vision, pages 1–11, 2019

    2019

  25. [32]

    The deepfake detection challenge (dfdc) dataset.arXiv preprint arXiv:2006.07397, 2020

    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

    2006 arXiv

  26. [33]

    Evaluation of an audio-video multimodal deepfake dataset using unimodal and multimodal detectors

    Hasam Khalid, Minha Kim, Shahroz Tariq, and Simon S Woo. Evaluation of an audio-video multimodal deepfake dataset using unimodal and multimodal detectors. InProceedings of the 1st Workshop on Synthetic Multimedia-Audiovisual Deepfake Generation and Detection, pages 7–15, 2021

    2021

  27. [34]

    Fakeavceleb: A novel audio-video multimodal deepfake dataset

    Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S Woo. Fakeavceleb: A novel audio-video multimodal deepfake dataset. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021

    2021

  28. [35]

    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, 2020

    2020

  29. [37]

    Towards understanding of deepfake videos in the wild

    Beomsang Cho, Binh M Le, Jiwon Kim, Simon Woo, Shahroz Tariq, Alsharif Abuadbba, and Kristen Moore. Towards understanding of deepfake videos in the wild. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 4530–4537, 2023

    2023

  30. [38]

    Why do facial deepfake detectors fail? InProceedings of the 2nd Workshop on Security Implications of Deepfakes and Cheapfakes, WDC ’23, page 24–28, New York, NY, USA, 2023

    Binh Le, Shahroz Tariq, Alsharif Abuadbba, Kristen Moore, and Simon Woo. Why do facial deepfake detectors fail? InProceedings of the 2nd Workshop on Security Implications of Deepfakes and Cheapfakes, WDC ’23, page 24–28, New York, NY, USA, 2023. Association for Computing Machinery

    2023

  31. [39]

    Woo, and Shahroz Tariq

    Razaib Tariq, Simon S. Woo, and Shahroz Tariq. Exploring the impact of moiré pattern on deepfake detectors. InSpecial Session Recent Advances in Multimedia Forensics in the AI-Generative Era (ICIP), Abu Dhabi, UAE, 2024

    2024

  32. [40]

    Woo, and Shahroz Tariq

    Razaib Tariq, Minji Heo, Simon S. Woo, and Shahroz Tariq. Beyond the screen: Evaluating deepfake detectors under moiré pattern effects. InWorkshop on Media Forensics (CVPR), Seattle, 2024

    2024

  33. [41]

    Chatgpt: A conversational ai model, 2023

    OpenAI. Chatgpt: A conversational ai model, 2023. Accessed: 2025-06-05

    2023

  34. [42]

    Gemini: Multimodal ai by google deepmind, 2023

    Google DeepMind. Gemini: Multimodal ai by google deepmind, 2023. Accessed: 2025-06-05

    2023

  35. [43]

    Claude: Constitutional ai assistant, 2023

    Anthropic. Claude: Constitutional ai assistant, 2023. Accessed: 2025-06-05

    2023

  36. [44]

    Grok: Conversational ai by xai, 2023

    xAI. Grok: Conversational ai by xai, 2023. Accessed: 2025-06-05

    2023

  37. [45]

    Flamingo: a visual language model for few-shot learning.Advances in neural Shahroz Tariq, David Nguyen, M.A.P

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning.Advances in neural Shahroz Tariq, David Nguyen, M.A.P. Chami...

    2022

  38. [46]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  39. [47]

    Antifakeprompt: Prompt-tuned vision-language models are fake image detectors.arXiv preprint arXiv:2310.17419, 2023

    You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. Antifakeprompt: Prompt-tuned vision-language models are fake image detectors.arXiv preprint arXiv:2310.17419, 2023

    2023 arXiv

  40. [48]

    Can chatgpt detect deepfakes? a study of using multimodal large language models for media forensics

    Shan Jia, Reilin Lyu, Kangran Zhao, Yize Chen, Zhiyuan Yan, Yan Ju, Chuanbo Hu, Xin Li, Baoyuan Wu, and Siwei Lyu. Can chatgpt detect deepfakes? a study of using multimodal large language models for media forensics. InProceedings of the IEEE/CVF Conference on Computer Vision a...

    2024

  41. [49]

    Can multi-modal (reasoning) llms work as deepfake detectors?arXiv preprint arXiv:2503.20084, 2025

    Simiao Ren, Yao Yao, Kidus Zewde, Zisheng Liang, Ning-Yau Cheng, Xiaoou Zhan, Qinzhe Liu, Yifei Chen, Hengwei Xu, et al. Can multi-modal (reasoning) llms work as deepfake detectors?arXiv preprint arXiv:2503.20084, 2025

    2025 arXiv

  42. [50]

    The gan is dead; long live the gan! a modern gan baseline.Advances in Neural Information Processing Systems, 37:44177–44215, 2024

    Nick Huang, Aaron Gokaslan, Volodymyr Kuleshov, and James Tompkin. The gan is dead; long live the gan! a modern gan baseline.Advances in Neural Information Processing Systems, 37:44177–44215, 2024

    2024

  43. [51]

    A2c: A modular multi-stage collaborative decision framework for human–ai teams

    Shahroz Tariq, Mohan Baruwal Chhetri, Surya Nepal, and Cecile Paris. A2c: A modular multi-stage collaborative decision framework for human–ai teams. Expert Systems with Applications, 282:127318, 2025

    2025

  44. [52]

    Towards a criteria-based approach to selecting human-ai interaction mode.arXiv preprint arXiv:2411.07406, 2024

    Jessica Irons, Patrick Cooper, Melanie McGrath, Shahroz Tariq, and Andreas Duenser. Towards a criteria-based approach to selecting human-ai interaction mode.arXiv preprint arXiv:2411.07406, 2024

    2024 arXiv

  45. [53]

    Bridging expertise gaps: The role of llms in human-ai collaboration for cybersecurity.arXiv preprint arXiv:2505.03179, 2025

    Shahroz Tariq, Ronal Singh, Mohan Baruwal Chhetri, Surya Nepal, and Cecile Paris. Bridging expertise gaps: The role of llms in human-ai collaboration for cybersecurity.arXiv preprint arXiv:2505.03179, 2025

    2025 arXiv

  46. [54]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  47. [55]

    Minigpt-4: Enhancing vision-language understanding with advanced large lan- guage models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large lan- guage models. InThe Twelfth International Conference on Learning Representa- tions

  48. [2023]

    18/092,590

    US Patent App. 18/092,590

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.