SEER: Semantic Enhancement and Emotional Reasoning Network for Multimodal Fake News Detection

Bin Chen; Keke Tang; Le Cheng; Lianwei Wu; Peican Zhu; Xiaodong Cui; Yubo Jing

REVIEW 4 major objections 5 minor 1 cited by

SEER: Semantic Enhancement and Emotional Reasoning Network for Multimodal Fake News Detection

T0 review · 4 major / 5 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 A multimodal network that augments text-image pairs with generated captions, CLIP features, and a Bayes-inspired emotion-reasoning loss achieves 0.929 accuracy on Weibo and 0.931 on Twitter, outperforming prior state-of-the-art baselines.

desk verdict Plausible architecture, unproven headline: the two-point gains over baselines rest on an uncontrolled comparison and dataset-tuned hyperparameters. read the letter →

arxiv 2507.13415 v1 pith:JF6KNVLN submitted 2025-07-17 cs.MM cs.AI

Peican Zhu , Yubo Jing , Le Cheng , Bin Chen , Xiaodong Cui , Lianwei Wu , Keke Tang This is my paper · ORCID

classification cs.MMcs.AI

keywords multimodalfakenewsdetectionsemanticenhancementBLIP-2imagecaptioningCLIPcross-modalalignmentemotionalreasoningexpertensembleco-attentionsocialmediamisinformation

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 argues that multimodal fake news detection is held back by two blind spots: existing systems read images at the pixel or entity level rather than understanding what the image actually depicts, and they ignore the emotional tilt of news. It proposes SEER, a network that injects BLIP-2-generated captions and CLIP embeddings into the text-image pipeline and adds an expert emotional-reasoning module that exploits the observation that fake news tends to carry more negative emotion. On the Weibo and Twitter benchmarks, SEER reports 0.929 and 0.931 accuracy respectively, exceeding the previous best baseline by roughly two points. The claim is that semantic enhancement plus emotion reasoning, not just cross-modal consistency, is what closes the remaining gap.

What carries the argument

The load-bearing object is a two-part network. The Multimodal Semantic Enhancement Module builds three multimodal views—original text-image, CLIP text-image, and CLIP-text with caption—interacts text, image, and caption through three co-attention blocks, and weights the fused representation by the CLIP text-image similarity $\theta=\cos(w^c_i, v^c_i)$. The Expert Emotional Reasoning Module runs $k=10$ experts over text and caption emotion features, blends them with a modality-weight $\lambda$, and converts the blended positive-emotion score $g_e$ into an authenticity estimate $\hat{y}^{\text{emo}}_i=\frac{\alpha}{\alpha+\beta}g_e+\frac{1-\alpha}{2-\alpha-\beta}(1-g_e)$, where $\alpha$ and $\beta$ are the probabilities of positive emotion in real and fake news. This predicted label feeds an auxiliary loss $\mathcal{L}_e$ that shapes the emotion features; the paper treats the Bayes relation as the bridge between sentiment statistics and veracity.

What would settle it

Train SEER on a third dataset while estimating $\alpha$ and $\beta$ from the training split's actual positive-emotion rates instead of fixing them; if accuracy drops compared with the paper's hand-set values, or if the emotion-score distributions of real and fake posts fail to separate, the Bayes-style emotional-reasoning mechanism is not the source of the reported improvement.

Watch

Extended reading notes

Core claim

SEER's central claim is that fusing scene-level captions and emotional priors into a multimodal detector closes a gap left by consistency-based methods. The paper reports 0.929 accuracy on Weibo and 0.931 on Twitter, beating all ten baselines, including the strongest, CCGN, by roughly two accuracy points. The ablation sequence shows that every added component—captions, CLIP, co-attention, self-attention, and the emotional-reasoning module—contributes to the final score, with CLIP removal being the most damaging on both datasets.

Load-bearing premise

The emotional-reasoning loss only works if the hand-set values of $\alpha$ and $\beta$ (the positive-emotion rates in real versus fake news) are calibrated to the data; if those priors are arbitrary, the auxiliary loss can distort emotion features and the reported gains may not carry to new datasets.

Editorial extensions

If this is right

If SEER's reported gains hold, adding cheap caption generation and emotion priors is a viable route beyond contrastive consistency for fake-news detection.
Removing CLIP is the single most damaging ablation, so aligned CLIP embeddings appear to be the largest contributor within the semantic-enhancement gains.
Caption-based emotion analysis outperforms image-based emotion analysis on both datasets, suggesting scene-level text descriptions carry sentiment that raw pixels do not.
The optimal number of emotion experts is around 10 on both datasets; too few or too many experts degrade performance.
Ablating either the semantic-enhancement module or the emotional-reasoning module drops accuracy back toward baseline levels, so the reported gain depends on both mechanisms together.

Reading between the lines

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

The paper leaves implicit that the hand-set $\alpha$ and $\beta$ values could be estimated from each dataset's training-split emotion distributions; testing whether learned priors change performance would reveal whether the loss is genuinely Bayesian or mostly a regularizer.
Because the emotional-reasoning loss assumes a positive/negative dichotomy, a natural extension is to test it on platforms or languages where ironic, sarcastic, or neutral fake news is common; the mechanism would only transfer if the sentiment skew is consistently present.
The two-point gain over CCGN may partly reflect that some baselines under-use CLIP; a fair head-to-head that gives every baseline the same CLIP features would isolate the contribution of SEER's specific architecture.
Using a different captioner than BLIP-2 would test whether the gains come from any scene-level description or from BLIP-2's particular alignment quality.

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

Two things to know up front. First, SEER is a reasonable architecture that combines three known ideas — BLIP-2 captions, CLIP alignment, and an emotion-reasoning auxiliary loss — and the ablations show each part contributes something. Second, the headline claim that SEER "significantly outperforms" all baselines is not actually established, because the comparison is not controlled: no split/protocol statement, no rerun of baselines, no error bars, and λ, k, α, β are set per dataset with no validation set described.

What is new here is the specific combination. The paper uses BLIP-2-generated captions as a third modality, CLIP embeddings to enhance both uni- and multi-modal features, and a multi-expert emotional reasoning module with a Bayes-flavored auxiliary loss. That combination is absent from the cited prior work. The ablations are thoughtfully designed and mostly tell a coherent story: CLIP removal hurts most, captions help more on Weibo, co-attention helps more on Twitter. The t-SNE visualization is illustrative rather than rigorous, but it does not contradict anything. The citation pattern looks fair; the authors engage with FEND-CLIP, CMMTN, CCGN and other close competitors, and the self-citations are on-topic.

Now the soft spots, in rough order of severity. The evaluation protocol is the load-bearing weakness. The paper says it follows "the same preprocessing steps and data split as the benchmark" but never names the benchmark, never says which baseline numbers were rerun in this framework, and reports no variance or significance tests. Table I has at least one typo (HMCAN's Twitter accuracy is "0,897"), which suggests the numbers were transcribed rather than reproduced. If the baseline numbers come from different splits and the authors' own hyperparameters were chosen on the test set via Figures 3–4, the reported 2-point gains may be optimistic bias, not real improvement. The Bayes derivation in Eq. 8 is also weaker than it looks. α and β are preset constants, not estimated from data; with the chosen values, the predicted probability is squeezed into [0.44, 0.60] on Weibo and [0.33, 0.68] on Twitter. That does not invalidate the auxiliary loss — it is still a supervised signal — but the "Bayes" framing is decorative, and the narrow range limits what the emotion module can express. Finally, the dataset description has a small but confusing inconsistency: Twitter is listed as 8720 real + 7448 fake news with only 514 images. This probably means 514 unique image URLs or post-level images, but it needs clarification.

Who is this for? Researchers in multimodal fake news detection who want to see whether BLIP-2 captions and emotion features help. The idea is worth trying, and the architecture is sensible enough to be a starting point. I would not trust the headline numbers until the protocol is fixed. I would send this to a serious referee with a request for code, explicit data split, rerun baselines, and validation-based hyperparameter selection. With those, the paper could become a solid contribution; without them, the superiority claim should not stand.

Referee Report

4 major / 5 minor

Summary. The paper proposes SEER, a multimodal fake-news detection network that augments text and image features with BLIP-2-generated image captions, CLIP embeddings, co-attention and self-attention modules, and an 'expert emotional reasoning' branch. The network is evaluated on the Weibo and Twitter datasets, reporting Accuracy and F1-score improvements over ten baselines, with ablations that remove individual components and a parameter analysis over the emotion-weight lambda and the number of experts k. The central claim is that SEER significantly outperforms all considered baselines, achieving 0.929 accuracy on Weibo and 0.931 on Twitter.

Significance. If the claimed performance holds under a controlled comparison, the paper makes a useful practical contribution: it demonstrates that large-multimodal-model captions and CLIP-based semantic enhancement can improve cross-modal alignment in fake-news detection, and it adds evidence that emotion-related features carry signal beyond standard text-image fusion. The ablation study is extensive and the architecture is clearly motivated. However, the current manuscript does not establish the central claim in a way that supports 'significantly outperforms': the comparison protocol is not described precisely, no error bars or significance tests are provided, several hyperparameters are dataset-specific and appear to be tuned without a described validation split, and the Bayes-based emotional-reasoning component rests on manually preset probabilities. These issues are correctable, but they are load-bearing for the paper's main conclusion.

major comments (4)

[Section IV-A/IV-B, Table I] The comparison underlying the central claim is not controlled. The text says the authors 'follow the same preprocessing steps and data split as the benchmark' but does not identify which benchmark split is used, which baseline numbers were re-run under the same framework, or which numbers are quoted from prior papers. The reported results have no standard deviations, no multiple-seed repeats, and no significance tests. Without these, the sentence in Section IV-B that SEER 'significantly outperforms all the considered baselines' is not established, especially because the reported gains over CCGN and HMCAN are on the order of one to two accuracy points.
[Section IV-A, Figures 3-4] The hyperparameters lambda, k, alpha, and beta are all dataset-specific, and lambda and k are selected via parameter-sweep plots without any described validation split. If these values were chosen based on test-set performance, the reported accuracies of 0.929 and 0.931 are optimistically biased. Please describe the validation procedure, report results for fixed hyperparameters, or provide nested cross-validation, and state explicitly whether the test set was used for any model selection step.
[Section III-C, Eq. (8)-(9)] The 'emotional reasoning' loss is not an independent source of emotional supervision. Equation (9) is a standard cross-entropy loss against the true labels y_i, and the Bayes-style mapping in Eq. (8) is computed from alpha and beta that are manually preset per dataset (Section IV-A) rather than estimated from data or learned. With the stated values, the predicted probability y_hat is confined to a narrow interval (about 0.44-0.60 on Weibo and 0.33-0.68 on Twitter), so the auxiliary loss can only weakly push predictions. The improvement from the EERM branch in Table IV could therefore come from the additional trainable feature branch rather than from emotional reasoning. Please estimate alpha and beta on a training split, treat them as learnable parameters, or otherwise demonstrate calibration; without this, the claimed interpretation of the module is not supported.
[Section IV-A, dataset descriptions] The reported Twitter image count is not credible as stated: 'Twitter contains 8720 real news and 7448 fake news with 514 images.' For a dataset with over 16,000 posts, a total of 514 images implies that most posts have no image, which would entirely change the meaning of the image-based and cross-modal results, including the ablations in Table II and the claim that 'images contain richer clues than text in Twitter.' Please clarify whether 514 is the number of images in a filtered subset, whether most posts lack images, and exactly how missing images are handled in the encoder, co-attention, CLIP, and caption-generation pipelines.

minor comments (5)

[Table I] Several cells in the Twitter block are missing separators or contain typographical corruption, for example 'HMCAN [16] 0,8970.9710.801' and '0.8860.9840.933' for CCGN. Please reformat the table and verify all numbers against the original runs.
[Section III-B, Eq. (3)] The description of the fusion features is hard to follow: sigma_i and MLP_M are used without explicit dimensional definitions, and the phrase 'passed to MLP' followed by 'concat' is ambiguous. Please define the input/output dimensions of each projection and state where the text-image similarity theta is applied.
[Section IV-A and Section II] The baselines list omits FEND-CLIP and MMFN even though they are discussed in Related Work; if they are not compared, please state why, since both are CLIP-based and directly relevant to the semantic-enhancement contribution.
[Section III-A] The phrase 'the products of large multimodal models' is vague. Please specify precisely which outputs are used (CLIP text/image embeddings, BLIP-2 captions, or other model outputs) and which model versions and decoding settings were used for caption generation.
[Section IV-E, Figure 5] The t-SNE visualization is described as showing that features 'clearly distinguish' real from fake news, but no quantitative cluster-quality or separation metric is provided. Consider reporting silhouette scores or a low-dimensional classification accuracy to support this claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SEER's components are externally pretrained models plus ordinary supervised losses, and the reported gains are not forced by construction.

full rationale

The derivation chain does not reduce to its own inputs. Semantic enhancement uses externally pretrained BLIP-2 and CLIP models plus trainable co-attention/self-attention blocks; the final detector (Eq. 10) is a standard supervised classifier. The emotional-reasoning branch (Eqs. 8-9) uses preset per-dataset constants α and β and the ground-truth labels y_i, making Le an ordinary auxiliary supervised loss rather than an independent prediction; the paper never claims this branch is unsupervised or that the final result is derived from emotion alone. The 'fake news is more negative' observation in Fig. 1(b) is an empirical dataset statistic used as motivation, not a theorem imported from the authors. Self-citations in the related-work list are not load-bearing for the headline comparison. I flag as non-circular correctness risks the under-specified statement in Section IV-A ('We follow the same preprocessing steps and data split as the benchmark') and the dataset-specific hyperparameter choices in Section IV-D, but these do not make any equation equivalent to its input by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the quality of pre-trained models (BLIP-2, CLIP), a presumed emotion-authenticity correlation, and hand-set parameters. No fully independent external benchmark anchors the emotion-reasoning component; the strongest anchors are the ablation experiments on the same two datasets.

free parameters (4)

lambda = 0.75 (Weibo), 0.25 (Twitter)
Weight balancing text and caption emotion contributions; tuned per dataset in Section IV-A.
alpha = 0.45 (Weibo), 0.65 (Twitter)
Preset probability that real news has a positive emotion, used in Eq. 8; chosen by hand per dataset.
beta = 0.3 (both datasets)
Preset probability that fake news has a positive emotion, used in Eq. 8; chosen by hand.
k (number of experts) = 10
Selected from the parameter analysis in Figure 4 as the best value; an architectural hyperparameter.

assumptions (5)

domain assumption Fake news is more inclined to contain negative emotions than real news.
Motivates the emotional reasoning module; supported only by a qualitative figure (Fig. 1b) without quantitative statistics.
standard math The prior probability of a news item being real equals that of being fake (0.5).
Used in the Bayes derivation for Eq. 8; plausible but not justified by data.
domain assumption BLIP-2 generated captions accurately summarize image semantics and align with text.
The semantic enhancement relies on the quality of pre-trained BLIP-2 captions; no manual inspection or error analysis is provided.
domain assumption CLIP embeddings provide a useful cross-modal alignment signal for fake news detection.
Relies on the pre-trained CLIP model; effectiveness is shown only via ablations.
domain assumption The data split and preprocessing match prior benchmarks exactly.
They state they follow prior work but do not specify the exact split, which is known to vary across papers and affect results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SEER: Semantic Enhancement and Emotional Reasoning Network for Multimodal Fake News Detection." pith.science (2026). https://pith.science/paper/JF6KNVLN

@misc{pith2026250713415,
  author       = {Pith},
  title        = {Pith review of: SEER: Semantic Enhancement and Emotional Reasoning Network for Multimodal Fake News Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JF6KNVLN}},
  note         = {Machine review of arXiv:2507.13415}
}

read the original abstract

Previous studies on multimodal fake news detection mainly focus on the alignment and integration of cross-modal features, as well as the application of text-image consistency. However, they overlook the semantic enhancement effects of large multimodal models and pay little attention to the emotional features of news. In addition, people find that fake news is more inclined to contain negative emotions than real ones. Therefore, we propose a novel Semantic Enhancement and Emotional Reasoning (SEER) Network for multimodal fake news detection. We generate summarized captions for image semantic understanding and utilize the products of large multimodal models for semantic enhancement. Inspired by the perceived relationship between news authenticity and emotional tendencies, we propose an expert emotional reasoning module that simulates real-life scenarios to optimize emotional features and infer the authenticity of news. Extensive experiments on two real-world datasets demonstrate the superiority of our SEER over state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2507.13415 by the authors.

**Figure 2.** Illustration of the proposed SEER. It contains of four components: Feature Representations, Multimodal Semantic Enhancement, Expert Emotional [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Impact of emotions from different modalities on two datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Effect of varying the number of experts ( [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Visualization of the learned representations of test data on Weibo. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Maximizing Mutual Information Between Prompt and Response Improves LLM Performance With No Additional Data
cs.LG 2026-03 unverdicted novelty 6.0 of 10

MIPO constructs contrastive preference pairs from correct versus random prompts and uses DPO to maximize mutual information between prompts and responses, producing 3-40% gains on personalization and 1-18% on math tas...

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages · cited by 1 Pith paper

[1]

Verifying multimedia use at mediaeval 2015

Christina Boididou, Katerina Andreadou, Symeon Papadopoulos, Duc Tien Dang Nguyen, Giulia Boato, Michael Riegler, Yiannis Kompatsiaris, et al. Verifying multimedia use at mediaeval 2015. In MediaEval 2015, volume 1436. CEUR-WS, 2015

work page 2015
[2]

Cross-modal ambiguity learning for multimodal fake news detection

Yixuan Chen, Dongsheng Li, Peng Zhang, Jie Sui, Qin Lv, Lu Tun, and Li Shang. Cross-modal ambiguity learning for multimodal fake news detection. InProceedings of the ACM web conference 2022, pages 2897–2905, 2022

work page 2022
[3]

Effi- cient source detection in incomplete networks via sensor deployment and source approaching.IEEE Transactions on Information Forensics and Security, 2025

Le Cheng, Peican Zhu, Keke Tang, Chao Gao, and Zhen Wang. Effi- cient source detection in incomplete networks via sensor deployment and source approaching.IEEE Transactions on Information Forensics and Security, 2025

work page 2025
[4]

Ccgn: consistency contrastive-learning graph network for multi-modal fake news detection.Multimedia Systems, 31(2):119, 2025

Shaodong Cui, Kaibo Duan, Wen Ma, and Hiroyuki Shinnou. Ccgn: consistency contrastive-learning graph network for multi-modal fake news detection.Multimedia Systems, 31(2):119, 2025

work page 2025
[5]

Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018

arXiv 2018
[6]

Kan: Knowledge-aware attention network for fake news detection

Yaqian Dun, Kefei Tu, Chen Chen, Chunyan Hou, and Xiaojie Yuan. Kan: Knowledge-aware attention network for fake news detection. InProceedings of the AAAI conference on artificial intelligence, volume 35, pages 81–89, 2021

work page 2021
[7]

Multimodal fake news detection through data augmentation- based contrastive learning.Applied Soft Computing, 136:110125, 2023

Jiaheng Hua, Xiaodong Cui, Xianghua Li, Keke Tang, and Peican Zhu. Multimodal fake news detection through data augmentation- based contrastive learning.Applied Soft Computing, 136:110125, 2023

work page 2023
[8]

Multimodal fusion with recurrent neural networks for rumor detection on microblogs

Zhiwei Jin, Juan Cao, Han Guo, Yongdong Zhang, and Jiebo Luo. Multimodal fusion with recurrent neural networks for rumor detection on microblogs. InProceedings of the 25th ACM international conference on Multimedia, pages 795–816, 2017

work page 2017

Show all 31 references

[9]

Mvae: Multimodal variational autoencoder for fake news detection

Dhruv Khattar, Jaipal Singh Goud, Manish Gupta, and Vasudeva Varma. Mvae: Multimodal variational autoencoder for fake news detection. InThe world wide web conference, pages 2915–2921, 2019

2019
[10]

Rumor detection with field of linear and non-linear propagation

An Lao, Chongyang Shi, and Yayi Yang. Rumor detection with field of linear and non-linear propagation. InProceedings of the web conference 2021, pages 3178–3187, 2021

2021
[11]

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

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

2023
[12]

Swin transformer: Hierarchical vision transformer using shifted windows

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021

2021
[13]

Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and- language tasks.Advances in neural information processing systems, 32, 2019

Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and- language tasks.Advances in neural information processing systems, 32, 2019

2019
[14]

Improving fake news detection by using an entity-enhanced framework to fuse di- verse multimodal clues

Peng Qi, Juan Cao, Xirong Li, Huan Liu, Qiang Sheng, Xiaoyue Mi, Qin He, Yongbiao Lv, Chenyang Guo, and Yingchao Yu. Improving fake news detection by using an entity-enhanced framework to fuse di- verse multimodal clues. InProceedings of the 29th ACM International Conference o...

2021
[15]

Ex- ploiting multi-domain visual information for fake news detection

Peng Qi, Juan Cao, Tianyun Yang, Junbo Guo, and Jintao Li. Ex- ploiting multi-domain visual information for fake news detection. In 2019 IEEE international conference on data mining (ICDM), pages 518–527. IEEE, 2019

2019
[16]

Hierarchical multi-modal contextual attention network for fake news detection

Shengsheng Qian, Jinguang Wang, Jun Hu, Quan Fang, and Chang- sheng Xu. Hierarchical multi-modal contextual attention network for fake news detection. InProceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 153–162, 2021

2021
[17]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

2021
[18]

Very deep convolu- tional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

Karen Simonyan and Andrew Zisserman. Very deep convolu- tional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

2014 arXiv
[19]

Spotfake: A multi-modal framework for fake news detection

Shivangi Singhal, Rajiv Ratn Shah, Tanmoy Chakraborty, Ponnu- rangam Kumaraguru, and Shin’ichi Satoh. Spotfake: A multi-modal framework for fake news detection. In2019 IEEE fifth international conference on multimedia big data (BigMM), pages 39–47. IEEE, 2019

2019
[20]

Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

2008
[21]

Positive unlabeled fake news detection via multi-modal masked trans- former network.IEEE Transactions on Multimedia, 26:234–244, 2024

Jinguang Wang, Shengsheng Qian, Jun Hu, and Richang Hong. Positive unlabeled fake news detection via multi-modal masked trans- former network.IEEE Transactions on Multimedia, 26:234–244, 2024

2024
[22]

Eann: Event adversarial neural networks for multi-modal fake news detection

Yaqing Wang, Fenglong Ma, Zhiwei Jin, Ye Yuan, Guangxu Xun, Kishlay Jha, Lu Su, and Jing Gao. Eann: Event adversarial neural networks for multi-modal fake news detection. InProceedings of the 24th acm sigkdd international conference on knowledge discovery & data mining, pages ...

2018
[23]

Fake news detection via knowledge-driven multimodal graph convolutional networks

Youze Wang, Shengsheng Qian, Jun Hu, Quan Fang, and Changsheng Xu. Fake news detection via knowledge-driven multimodal graph convolutional networks. InProceedings of the 2020 international conference on multimedia retrieval, pages 540–547, 2020

2020
[24]

Multimodal fusion with co-attention networks for fake news detection

Yang Wu, Pengwei Zhan, Yunjian Zhang, Liming Wang, and Zhen Xu. Multimodal fusion with co-attention networks for fake news detection. InFindings of the association for computational linguistics: ACL- IJCNLP 2021, pages 2560–2569, 2021

2021
[25]

Gamc: an unsupervised method for fake news detection using graph autoencoder with masking

Shu Yin, Peican Zhu, Lianwei Wu, Chao Gao, and Zhen Wang. Gamc: an unsupervised method for fake news detection using graph autoencoder with masking. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 347–355, 2024

2024
[26]

: Similarity-aware multi- modal fake news detection

Xinyi Zhou, Jindi Wu, and Reza Zafarani. : Similarity-aware multi- modal fake news detection. InPacific-Asia Conference on knowledge discovery and data mining, pages 354–367. Springer, 2020

2020
[27]

Multi-modal fake news detection on social media via multi-grained information fusion

Yangming Zhou, Yuzhou Yang, Qichao Ying, Zhenxing Qian, and Xinpeng Zhang. Multi-modal fake news detection on social media via multi-grained information fusion. InProceedings of the 2023 ACM International Conference on Multimedia Retrieval, pages 343– 352, 2023

2023
[28]

Multimodal fake news detection via clip-guided learning

Yangming Zhou, Yuzhou Yang, Qichao Ying, Zhenxing Qian, and Xinpeng Zhang. Multimodal fake news detection via clip-guided learning. In2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2825–2830. IEEE, 2023

2023
[29]

Multimodal fake news detection through intra-modality feature aggregation and inter-modality semantic fusion.Complex & Intelligent Systems, 10(4):5851–5863, 2024

Peican Zhu, Jiaheng Hua, Keke Tang, Jiwei Tian, Jiwei Xu, and Xiaodong Cui. Multimodal fake news detection through intra-modality feature aggregation and inter-modality semantic fusion.Complex & Intelligent Systems, 10(4):5851–5863, 2024

2024
[30]

Ken: Knowledge augmentation and emotion guidance network for multimodal fake news detection.arXiv preprint arXiv:2507.09647, 2025

Peican Zhu, Yubo Jing, Le Cheng, Keke Tang, and Yangming Guo. Ken: Knowledge augmentation and emotion guidance network for multimodal fake news detection.arXiv preprint arXiv:2507.09647, 2025

2025 arXiv
[31]

A general black-box adversarial attack on graph-based fake news detectors

Peican Zhu, Zechen Pan, Yang Liu, Jiwei Tian, Keke Tang, and Zhen Wang. A general black-box adversarial attack on graph-based fake news detectors. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 568–576, 2024

2024

Pith tools

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

[1] [1]

Verifying multimedia use at mediaeval 2015

Christina Boididou, Katerina Andreadou, Symeon Papadopoulos, Duc Tien Dang Nguyen, Giulia Boato, Michael Riegler, Yiannis Kompatsiaris, et al. Verifying multimedia use at mediaeval 2015. In MediaEval 2015, volume 1436. CEUR-WS, 2015

work page 2015

[2] [2]

Cross-modal ambiguity learning for multimodal fake news detection

Yixuan Chen, Dongsheng Li, Peng Zhang, Jie Sui, Qin Lv, Lu Tun, and Li Shang. Cross-modal ambiguity learning for multimodal fake news detection. InProceedings of the ACM web conference 2022, pages 2897–2905, 2022

work page 2022

[3] [3]

Effi- cient source detection in incomplete networks via sensor deployment and source approaching.IEEE Transactions on Information Forensics and Security, 2025

Le Cheng, Peican Zhu, Keke Tang, Chao Gao, and Zhen Wang. Effi- cient source detection in incomplete networks via sensor deployment and source approaching.IEEE Transactions on Information Forensics and Security, 2025

work page 2025

[4] [4]

Ccgn: consistency contrastive-learning graph network for multi-modal fake news detection.Multimedia Systems, 31(2):119, 2025

Shaodong Cui, Kaibo Duan, Wen Ma, and Hiroyuki Shinnou. Ccgn: consistency contrastive-learning graph network for multi-modal fake news detection.Multimedia Systems, 31(2):119, 2025

work page 2025

[5] [5]

Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018

arXiv 2018

[6] [6]

Kan: Knowledge-aware attention network for fake news detection

Yaqian Dun, Kefei Tu, Chen Chen, Chunyan Hou, and Xiaojie Yuan. Kan: Knowledge-aware attention network for fake news detection. InProceedings of the AAAI conference on artificial intelligence, volume 35, pages 81–89, 2021

work page 2021

[7] [7]

Multimodal fake news detection through data augmentation- based contrastive learning.Applied Soft Computing, 136:110125, 2023

Jiaheng Hua, Xiaodong Cui, Xianghua Li, Keke Tang, and Peican Zhu. Multimodal fake news detection through data augmentation- based contrastive learning.Applied Soft Computing, 136:110125, 2023

work page 2023

[8] [8]

Multimodal fusion with recurrent neural networks for rumor detection on microblogs

Zhiwei Jin, Juan Cao, Han Guo, Yongdong Zhang, and Jiebo Luo. Multimodal fusion with recurrent neural networks for rumor detection on microblogs. InProceedings of the 25th ACM international conference on Multimedia, pages 795–816, 2017

work page 2017