Leveraging Large Language Models for Information Verification -- an Engineering Approach

Dao Minh Tu; Nguyen An Phuoc; Nguyen Dinh Mau; Nguyen Manh Duc Tuan; Nguyen Nang Hung; Nguyen Thanh Trong; Vuong Thanh Toan

REVIEW 4 major objections 6 minor 20 references

Leveraging Large Language Models for Information Verification -- an Engineering Approach

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a single general-purpose LLM, GPT-4o, can serve as the reasoning core of a fully automated multimedia news-verification pipeline.

desk verdict A clean, honest system description for the ACMMM25 verification challenge, but the core claim that the pipeline can verify news is backed only by a single case study and no quantitative evaluation. read the letter →

arxiv 2506.18274 v1 pith:SFDT5XLO submitted 2025-06-23 cs.LG

Nguyen Nang Hung , Nguyen Thanh Trong , Vuong Thanh Toan , Nguyen An Phuoc , Dao Minh Tu , Nguyen Manh Duc Tuan , Nguyen Dinh Mau This is my paper

classification cs.LG

keywords multimedianewsverificationGPT-4opipelinepromptengineeringkeyframeselectionK-meansclusteringaudiotranscriptionLLMfact-checkingforensicanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a single general-purpose LLM, GPT-4o, can serve as the reasoning core of an automated multimedia news-verification pipeline. The pipeline turns videos into a compact set of representative frames, transcribes the audio, gathers related articles through a web search, and then runs two prompt-driven passes — cross-validation and forensic analysis — that decide which parts of a claim are verifiable. The authors demonstrate the method on a challenge dataset of 50 samples, 72.56% of which are videos, and show a worked example that ends in coordinates, a date range with a consensus label, and an authenticity assessment. Their central claim is that with careful prompt engineering, the API version of GPT-4o can reproduce the multi-step reasoning of the web version while keeping a human involved only at the final validation step.

What carries the argument

The central mechanism is the pair of purpose-built prompts, Prompt 1 (cross-validation of information sources) and Prompt 2 (forensic analysis). Prompt 1 forces the LLM to reconcile web sources into a structured output with location, coordinates, a date-range consensus category, event consensus and conflicts, and tags; Prompt 2 then treats the selected video keyframes as the ground truth and asks the LLM to find evidence validating or contradicting that metadata, and to judge whether the content is synthetic, modified, or AI-generated. The supporting machinery is the frame-selection chain (shot detection, ViT feature extraction, and K-means clustering with silhouette-score cluster counting) that compresses video into a token-viable set of representative frames, along with audio transcription that adds a parallel textual channel.

What would settle it

Run the pipeline on a video that splices two unrelated events into one clip, with a description matching only one of them, and check whether the forensic prompt flags the splice and whether cross-validation still reports consensus; or feed it a known GAN-generated image and observe whether the LLM-only forensic analysis detects it, given the paper's own admission that general LLMs miss artifacts dedicated detectors are built to find.

Watch

Extended reading notes

Core claim

The paper's discovery is that multimedia verification can be reduced to a structured two-stage dialogue with one LLM. First, video is segmented by a shot detection model, frames are embedded with a Vision Transformer, and K-means clustering selects the most representative frames; a speech recognition model transcribes the audio; web search results are crawled and summarized. Then the first prompt aggregates the web sources into a JSON of location, coordinates, date with a Consensus/Partial/Non-verifiable label, event description, and tags, while the second prompt re-reads the selected keyframes against that JSON, validating location, event, and people, and rating the content's authenticity. The worked example of the Lyman sample shows the pipeline producing a coherent, explainable report whose coordinates match those from a map service with only small variations. The paper's contribution is the engineering arrangement itself: a fully automated pipeline whose output is a package of evidence — not a binary truth verdict — with human intervention limited to final validation.

Load-bearing premise

The pipeline assumes that unedited media content is the undeniable ground truth for verification, so if a video or image is manipulated, misattributed, or staged, the entire verification chain inherits that error with no independent synthetic-media check.

Editorial extensions

If this is right

A news claim can be processed end to end without human effort until the final sign-off, producing a structured verdict with location, a dated event window, an event summary, and an authenticity rating.
Because 72.56% of the challenge data is video, the frame-selection mechanism is what makes the LLM API practical; video verification works without feeding raw footage to the model.
The date-consensus categories (Consensus, Partial, Non-verifiable) provide a reusable, human-readable way to state how much confidence the evidence supports about when an event occurred.
The same prompt-driven arrangement can be pointed at new news items by swapping the search keywords and re-running the two prompts, so the pipeline is a candidate template for ongoing media monitoring.

Reading between the lines

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

Beyond the paper, the date-consensus threshold ('less than a month' = Consensus, one to three months = Partial) is an arbitrary design choice; a natural extension is to calibrate these boundaries against human annotators' judgments on the same events.
Beyond the paper, the pipeline's reliance on the LLM's internal geographic knowledge could be directly stress-tested by comparing its extracted coordinates with a dedicated map API across a broad sample, not just on the single reported example.
Beyond the paper, one can isolate how much of the accuracy comes from the frame-selection mechanism by rerunning the same two prompts on uniformly sampled frames and comparing the verdicts; this would turn the engineering pipeline into a measurable claim about the value of clustering.
Beyond the paper, the stated limitations suggest an upgrade path: plugging a dedicated provenance or watermarking checker into the authenticity stage would close the gap the authors acknowledge for broader news verification.

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.

Desk Editor's Note

read the letter

Colleague,

Read the ACMMM25 technical report from the FPT team. Short version: it's a clean, honest write-up of an LLM-based multimodal verification pipeline, but the central claim that the pipeline can verify news is supported only by one case study. No benchmark, no baseline, no error analysis. If you treat it as a system description for the challenge, it's fine; if you treat it as evidence that this approach works, it isn't.

What's actually new: not much in terms of algorithms. Every component — TransNetV2 shot detection, ViT embeddings, K-means keyframe selection, Whisper transcription, GPT-4o prompting, Google search — is off-the-shelf. The contribution is the integration and two carefully designed prompts. That's genuinely useful for practitioners entering the challenge. The report also deserves credit for laying out the pipeline step by step with a concrete sample output (ID115), and the limitation section is unusually candid: they admit they rely on GPT for geographic coordinates, they omit a dedicated synthetic-media detector, and they note the metadata retrieval is text-only.

The soft spots are real and load-bearing. First, no quantitative evaluation on the 50-sample challenge set. The single case study is a demonstration, not evidence of reliability. The claim that their report has 'better reasoning' than the web-service version is unsupported. Second, the forensic step in Section 3.3 assumes that unedited media is ground truth, but the judgment of 'unedited' is delegated to GPT-4o itself. They say a manual double-check survey showed no AI-generated content, but the size, protocol, and results of that survey are not reported. Since the pipeline validates metadata against the images rather than authenticating the images independently, a manipulated or misattributed video would corrupt the whole chain. They acknowledge this in Section 5.1, which mitigates but does not fix it.

Who's it for? Someone building a similar pipeline for the ACMMM verification challenge, or a researcher looking for a well-structured example of LLM-agent integration. It is not a research result. I would not send it to a serious peer-review venue in its current form; it needs at least a benchmark and an error analysis. Worth a skim for the prompts and the honest limitations.

Referee Report

4 major / 6 minor

Summary. The paper describes an automated pipeline for multimedia news verification developed for the ACMMM25 Grand Challenge. The pipeline retrieves web metadata via Google search based on keywords extracted from the news title/description, selects representative video keyframes using TransNetV2 shot detection and K-means clustering over ViT features, transcribes audio with Whisper, and then uses GPT-4o API through two prompts to cross-validate the retrieved metadata and perform forensic analysis of the media content. The authors claim the pipeline is fully automated with human intervention limited to final validation, and they demonstrate it on a single sample (ID115) with a detailed step-by-step output. The paper also includes an honest discussion of limitations, including the omission of a dedicated synthetic-media detector and reliance on GPT-4o's internal knowledge for geographic coordinates.

Significance. The engineering contribution is a concrete, modular architecture that combines established components (TransNetV2, ViT, K-means, Whisper) with prompt-based LLM reasoning, and the paper is transparent about its limitations. The provided prompts are potentially reusable and the pipeline is described clearly enough to be reproduced. However, the central claim that the pipeline can verify multimedia news with human intervention limited to final validation is not supported by quantitative evidence: the single case study is a demonstration of feasibility rather than reliability, no evaluation on the 50-sample challenge set is reported, and no comparison with baselines or error analysis is provided. If the authors add a rigorous evaluation, the pipeline could serve as a useful practical baseline for the challenge, but in its current form the effectiveness claim remains unsubstantiated.

major comments (4)

[Section 4 and abstract] The central claim that the pipeline can verify multimedia news with human intervention limited to final validation is supported only by a single qualitative case study (ID115). No quantitative results on the 50-sample challenge set, no agreement with ground-truth answers, no precision/recall or error analysis, and no comparison with baselines are reported. Since the abstract and the technical contributions in Section 2.3 make load-bearing assertions about the pipeline's effectiveness, the lack of evaluation leaves the core claim unsubstantiated.
[Section 3.3, Prompt 2 and the ground-truth assumption] The forensic analysis step assumes that media content, if found unedited, is the undeniable ground truth for verifying other information. However, the determination of whether content is unedited, synthetic, or modified is made by the same GPT-4o model that validates the metadata against the images, and the prompt itself instructs the model to perform this check without any independent tool. The manual double-check survey mentioned in Section 3.3 is not described (no sample size, protocol, or results), so the assumption is never stress-tested. Because this assumption is load-bearing for the entire verification chain, the paper must either integrate an independent synthetic-media detector or provide a documented validation of the assumption on the challenge data.
[Sections 5.1 and 3.1, coordinate output] The paper acknowledges that geographic coordinates are extracted from GPT-4o's internal knowledge base rather than from a specific geocoding tool, and states that accurate coordinates cannot be convincingly extracted without tools such as Google Maps. Since geographic coordinates are one of the three required output elements, and the authors state that 70% of the data is geographically verifiable, the accuracy of the coordinate output is directly load-bearing for the practical-effectiveness claim. The report should quantify coordinate errors on the challenge set or integrate a geocoding tool to make the claim credible.
[Section 2.3, comparison claim] The claim that the proposed pipeline shows 'high coherence, better reasoning, and in-depth cross-validation' compared to the GPT-4o web service is not supported by any controlled comparison. The case study in Section 2.3 uses sample ID221 with the web-service version, while the pipeline demonstration in Section 4 uses a different sample (ID115), so the two outputs cannot be directly compared. If this claim is retained, the authors must present paired outputs of the web service and the pipeline on the same inputs.

minor comments (6)

[Abstract] The final sentence of the abstract is a sentence fragment and lacks a period; it should be completed and punctuated.
[Section 2.2] The phrase 'geologically verifiable' should likely be 'geographically verifiable', since the context is location coordinates rather than geological features.
[Section 2.3] There is a typo in 'APT-4o' which should read 'GPT-4o'.
[Section 3.2.1] The description of determining the optimal cluster count via silhouette score is vague; the paper should specify the range of candidate cluster counts that was evaluated.
[References] Reference [6] is a web page summary of the ViT paper rather than a primary source; consider citing the original ViT paper or removing the secondary reference.
[Figure 1] Figure 1 is not explicitly referenced in the text and has no caption description; please add a reference and a caption explaining the pipeline flow.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline is externally grounded in Google search results, media content, and human final validation, with no derivation chain that reduces to its own inputs.

full rationale

The paper is an engineering report, not a derivation. Its load-bearing components are external: Google search results supply metadata, video frames and audio transcripts supply multimedia evidence, and GPT-4o is used as a summarizer and cross-checker. The one potential loop—GPT summarizes external web metadata in Prompt 1 and then validates that summary against images in Prompt 2—does not reduce to the paper's own output by construction, because the metadata content originates from independent web sources and the final report is subject to human validation. The authors explicitly acknowledge the weakness of relying on GPT for synthetic-media detection ('we are relying on GPT to find evidences of synthetic and modified content, which is not ideal') and the omission of a dedicated detector, which is a stated limitation rather than a disguised input. There are no fitted parameters renamed as predictions, no self-citations carrying the argument, and no uniqueness theorem imported from the authors' prior work. The effectiveness claim is supported only by a single qualitative case study and an unreported manual survey, but that is a lack of quantitative evidence, not circularity. Accordingly, no specific circular step can be quoted and exhibited, and the correct finding is no significant circularity.

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

The central claim rests on the reliability of external models and data sources rather than on fitted equations. The free parameters are engineering choices with no sensitivity evidence, and the axioms are standard domain assumptions for an LLM-based verification system.

free parameters (3)

K (number of keyframes per shot and number of search results) = 10 (default)
Chosen by hand; no sensitivity analysis. Used in Sections 3.1 and 3.2.1.
Audio chunk length for Whisper transcription = 30 seconds
Engineering choice in Section 3.2.2; no ablation or justification.
Keyframe downscaling resolution = 256x256 pixels
Chosen for token limits in Section 3.2.1; trade-off with visual fidelity is not measured.

assumptions (4)

domain assumption Media data, if unedited, is the undeniable ground-truth for verification.
Invoked in Section 3.3 Forensic Analysis; the forensic prompt uses video frames as evidence to validate metadata, so any manipulation in the media propagates to the conclusion. The authors note they rely on this assumption and omit dedicated synthetic detection because the provided data appeared to be real footage.
domain assumption Pre-trained models TransNetV2, ViT, and Whisper perform adequately on news-domain video and audio.
Section 3.2 uses TransNetV2 for shot detection, a pre-trained ViT for feature extraction, and Whisper for transcription; the pipeline's keyframe selection quality depends on these models generalizing to the challenge data.
domain assumption Google search retrieval covers the relevant event sources.
Section 3.1 builds metadata from the top-K search results; if the search engine misses the event or returns inaccessible pages, cross-validation lacks the necessary evidence. The demonstration shows several links failing to fetch.
domain assumption GPT-4o can reliably interpret downscaled keyframes and answer structured verification prompts.
Sections 3.3 and the case study rely on GPT-4o to extract coordinates, judge authenticity, and compare sources; errors in the LLM's perception or reasoning directly enter the final report.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Leveraging Large Language Models for Information Verification -- an Engineering Approach." pith.science (2026). https://pith.science/paper/SFDT5XLO

@misc{pith2026250618274,
  author       = {Pith},
  title        = {Pith review of: Leveraging Large Language Models for Information Verification -- an Engineering Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SFDT5XLO}},
  note         = {Machine review of arXiv:2506.18274}
}

read the original abstract

For the ACMMM25 challenge, we present a practical engineering approach to multimedia news source verification, utilizing Large Language Models (LLMs) like GPT-4o as the backbone of our pipeline. Our method processes images and videos through a streamlined sequence of steps: First, we generate metadata using general-purpose queries via Google tools, capturing relevant content and links. Multimedia data is then segmented, cleaned, and converted into frames, from which we select the top-K most informative frames. These frames are cross-referenced with metadata to identify consensus or discrepancies. Additionally, audio transcripts are extracted for further verification. Noticeably, the entire pipeline is automated using GPT-4o through prompt engineering, with human intervention limited to final validation.

Figures

Figures reproduced from arXiv: 2506.18274 by the authors.

**Figure 1.** Pipeline overview. examine the potential of Large Language Models, particularly GPT4o, in solving the problem at hands. This case study motivates us to develop companion tools to support GPT-4o as our backbone knowledge base. Section 3 details our pipeline and its main components: multimedia processing and cross verification. Section 4 is dedicated to practical implementations where code blocks and sample outputs … view at source ↗

**Figure 2.** Selected Image Frames. Step 3: Audio Processing Audio transcript is in Russian, which may or may not be fully cohesive depending on the quality of the audio itself. Here we omit showing the sample transcription due to technical difficulty in rendering Russian characters on LATEX. Step 4: Cross Validation of Information Sources Cross Validation Output { "location": { "location": "Lyman (formerly Krasnyi Lyman), Donet… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 13 canonical work pages

[1]

Videosummarization usingclustering

A.Goyal,M.Singh,andM.S.Patterh,“Videosummarization usingclustering”, InternationalJournalofComputerApplica- tions,vol.99,no.18,pp.1–4,2014. doi:10.5120/17446-8153

work page doi:10.5120/17446-8153 2014
[2]

Transnetv2:Aneffectivedeepnetwork architectureforfastshottransitiondetection

T.SoučekandJ.Lokoč,“Transnetv2:Aneffectivedeepnetwork architectureforfastshottransitiondetection”, arXivpreprint arXiv:2008.04838,2020.[Online].Available:https://arxiv.org/ abs/2008.04838

arXiv 2008
[3]

Asurveyoffakenews:Fundamental theories,detectionmethods,andopportunities

X.ZhouandR.Zafarani,“Asurveyoffakenews:Fundamental theories,detectionmethods,andopportunities”, ACMComput- ingSurveys,vol.53,no.5,pp.1–40,2020. doi:10.1145/3395046

doi:10.1145/3395046 2020
[4]

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

A. Dosovitskiy, L. Beyer, A. Kolesnikov,et al., “An image is worth 16x16 words: Transformers for image recognition at scale”,in InternationalConferenceonLearningRepresentations (ICLR),2021.[Online].Available:https://arxiv.org/abs/2010. 11929

work page 2021
[5]

Approachesfor fakenewsaboutcovid-19detection

Y.Kuznetsova,L.Tsvetkova,andN.Anikina,“Approachesfor fakenewsaboutcovid-19detection”, ComputaciónySistemas , vol.25,no.4,pp.783–793,2021. doi:10.13053/CyS-25-4-4091

work page doi:10.13053/cys-25-4-4091 2021
[6]

A. Perincherry,An image is worth 16x16 words: Transform- ersforimagerecognitionatscale-papersummary ,WebPage, 2021.[Online].Available:http://www.akhilperincherry.com/ paperSummaries/vitPaperSummary.html

work page 2021
[7]

Deepfakesgenerationanddetection:Ashortsur- vey

Z.Akhtar,“Deepfakesgenerationanddetection:Ashortsur- vey”,JournalofImaging ,vol.9,no.1,p.18,2023. doi:10.3390/ jimaging9010018

work page 2023
[8]

Areviewoffake newsdetectionapproaches:Acriticalanalysisofrelevantstud- iesandhighlightingkeychallengesassociatedwiththedataset, featurerepresentation,anddatafusion

S.K.Hamed,M.J.AbAziz,andM.R.Yaakub,“Areviewoffake newsdetectionapproaches:Acriticalanalysisofrelevantstud- iesandhighlightingkeychallengesassociatedwiththedataset, featurerepresentation,anddatafusion”, Heliyon,vol.9,no.10, e20382,2023. doi:10.1016/j.heliyon.2023.e20382

work page doi:10.1016/j.heliyon.2023.e20382 2023

Show all 20 references

[9]

Cluster-based video summarization with temporal context awareness

H.-D. Huynh-Lam, N.-P. Ho-Thi, M.-T. Tran, and T.-N. Le, “Cluster-based video summarization with temporal context awareness”, inPacific-Rim Symposium on Image and Video Technology,2023,pp.15–28.[Online].Available:https://arxiv. org/abs/2404.04511

2023 arXiv
[10]

Robustspeechrecognitionvialarge-scaleweaksu- pervision

A.Radford,J.W.Kim,T.Xu,G.Brockman,C.Mcleavey,andI. Sutskever,“Robustspeechrecognitionvialarge-scaleweaksu- pervision”,in Proceedingsofthe40thInternationalConference onMachineLearning ,ser.ProceedingsofMachineLearning Research,vol.202,PMLR,2023,pp.28492–28518.[Online]. Availabl...

2023
[11]

Developmentandclassi- ficationoffakenewsdetectiontechnology

X.Yang,Y.Li,Z.Zhang,andH.Liu,“Developmentandclassi- ficationoffakenewsdetectiontechnology”, AppliedSciences, vol.13,no.21,p.11877,2023. doi:10.3390/app132111877

2023 doi
[12]

Fakenewsdetection:Acompre- hensivesurveyofdatasets

A.AlsudaisandS.Al-Otaibi,“Fakenewsdetection:Acompre- hensivesurveyofdatasets”, arXivpreprintarXiv:2407.02122 , 2024

2024
[13]

Evaluatingtheuseofllmagentsforfact-checking

N. Kotonya, F. Boido, A. Tonon, A. Al-Ghadir, and Y. Al- Onaizan,“Evaluatingtheuseofllmagentsforfact-checking”, PLoSONE,vol.19,no.2,e0297837,2024. doi:10.1371/journal. pone.0297837

2024 doi
[14]

Assessingandim- proving the logical consistency of large language models on fact-checking

L.Pan,A.Al-Ghadir,andY.Al-Onaizan,“Assessingandim- proving the logical consistency of large language models on fact-checking”, inThe Twelfth International Conference on LearningRepresentations, 2024. [Online]. Available: https:// openreview.net/forum?id=SimlDuN0YT

2024
[15]

Deepfake generation and detection: A benchmark and survey

G. Pei, J. Zhang, M. Hu, et al., “Deepfake generation and detection: A benchmark and survey”,arXiv preprint arXiv:2403.17881,2024

2024
[16]

Asurvey ondeepfakedetectionthroughdeeplearning

P.K.Thai,S.Kalige,S.N.Ediga,andL.Chougoni,“Asurvey ondeepfakedetectionthroughdeeplearning”, WorldJournal ofAdvancedResearchandReviews ,vol.21,no.3,pp.2214–2217,
[17]

Gen- erativelargelanguagemodelsinautomatedfact-checking:A survey

I.Vykopal,M.Pikuliak,S.Ostermann,andM.Šimko,“Gen- erativelargelanguagemodelsinautomatedfact-checking:A survey”, arXivpreprintarXiv:2407.02351 ,2024

2024 arXiv
[18]

Contentcredentials

Coalition for Content Provenance and Authenticity (C2PA). “Contentcredentials”,CoalitionforContentProvenanceand Authenticity (C2PA). (2025), [Online]. Available: https:// contentcredentials.org/

2025
[19]

Synthiddetector—anewportaltohelpidentifyai- generatedcontent

P.Kohli,“Synthiddetector—anewportaltohelpidentifyai- generatedcontent”, GoogleDeepMindBlog ,May2025.[Online]. Available:https://blog.google/technology/ai/google-synthid- ai-content-detector/. 6 Hung Nguyen et al
[2024]

doi:10.30574/wjarr.2024.21.3.0946

2024 doi

Pith tools

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

[1] [1]

Videosummarization usingclustering

A.Goyal,M.Singh,andM.S.Patterh,“Videosummarization usingclustering”, InternationalJournalofComputerApplica- tions,vol.99,no.18,pp.1–4,2014. doi:10.5120/17446-8153

work page doi:10.5120/17446-8153 2014

[2] [2]

Transnetv2:Aneffectivedeepnetwork architectureforfastshottransitiondetection

T.SoučekandJ.Lokoč,“Transnetv2:Aneffectivedeepnetwork architectureforfastshottransitiondetection”, arXivpreprint arXiv:2008.04838,2020.[Online].Available:https://arxiv.org/ abs/2008.04838

arXiv 2008

[3] [3]

Asurveyoffakenews:Fundamental theories,detectionmethods,andopportunities

X.ZhouandR.Zafarani,“Asurveyoffakenews:Fundamental theories,detectionmethods,andopportunities”, ACMComput- ingSurveys,vol.53,no.5,pp.1–40,2020. doi:10.1145/3395046

doi:10.1145/3395046 2020

[4] [4]

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

A. Dosovitskiy, L. Beyer, A. Kolesnikov,et al., “An image is worth 16x16 words: Transformers for image recognition at scale”,in InternationalConferenceonLearningRepresentations (ICLR),2021.[Online].Available:https://arxiv.org/abs/2010. 11929

work page 2021

[5] [5]

Approachesfor fakenewsaboutcovid-19detection

Y.Kuznetsova,L.Tsvetkova,andN.Anikina,“Approachesfor fakenewsaboutcovid-19detection”, ComputaciónySistemas , vol.25,no.4,pp.783–793,2021. doi:10.13053/CyS-25-4-4091

work page doi:10.13053/cys-25-4-4091 2021

[6] [6]

A. Perincherry,An image is worth 16x16 words: Transform- ersforimagerecognitionatscale-papersummary ,WebPage, 2021.[Online].Available:http://www.akhilperincherry.com/ paperSummaries/vitPaperSummary.html

work page 2021

[7] [7]

Deepfakesgenerationanddetection:Ashortsur- vey

Z.Akhtar,“Deepfakesgenerationanddetection:Ashortsur- vey”,JournalofImaging ,vol.9,no.1,p.18,2023. doi:10.3390/ jimaging9010018

work page 2023

[8] [8]

Areviewoffake newsdetectionapproaches:Acriticalanalysisofrelevantstud- iesandhighlightingkeychallengesassociatedwiththedataset, featurerepresentation,anddatafusion

S.K.Hamed,M.J.AbAziz,andM.R.Yaakub,“Areviewoffake newsdetectionapproaches:Acriticalanalysisofrelevantstud- iesandhighlightingkeychallengesassociatedwiththedataset, featurerepresentation,anddatafusion”, Heliyon,vol.9,no.10, e20382,2023. doi:10.1016/j.heliyon.2023.e20382

work page doi:10.1016/j.heliyon.2023.e20382 2023