Source-linked AI summary
A Survey on Multimodal Disinformation Detection
Firoj Alam, Stefano Cresci, Tanmoy Chakraborty, Fabrizio Silvestri, Dimiter Dimitrov, Giovanni Da San Martino, Shaden Shaar, Hamed Firooz, Preslav Nakov
TL;DR
Multimodal disinformation research must account for falsehood and harmful intent, yet these aspects and modalities are often handled separately. This survey reviews research across modalities and argues for considering both dimensions together. It finds that multimodal datasets generally cover only two to three modalities, and none cover both factuality and harmfulness.
Problem
The survey identifies the challenge of combining different, often unaligned modalities and the lack of datasets that support this research.
Source: §7 · p. 9Method
The paper surveys multimodal disinformation research and synthesizes findings, challenges, and directions across modalities.
Source: §7 · p. 9Results
Most multimodal datasets cover just 2–3 modalities, and none cover both factuality and harmfulness; the study addressing both covered only text.
Source: C Lessons Learned · p. 18Takeaways & Limitations
The survey argues that disinformation detection should combine multiple modalities with factuality and harmfulness in the same framework.
Source: §7 · p. 9The authors may not have covered all relevant work on factuality and harmfulness across modalities.
Source: Limitations · p. 10
Abstract
from arXiv · showhide
Recent years have witnessed the proliferation of offensive content online such as fake news, propaganda, misinformation, and disinformation. While initially this was mostly about textual content, over time images and videos gained popularity, as they are much easier to consume, attract more attention, and spread further than text. As a result, researchers started leveraging different modalities and combinations thereof to tackle online multimodal offensive content. In this study, we offer a survey on the state-of-the-art on multimodal disinformation detection covering various combinations of modalities: text, images, speech, video, social media network structure, and temporal information. Moreover, while some studies focused on factuality, others investigated how harmful the content is. While these two components in the definition of disinformation (i) factuality, and (ii) harmfulness, are equally important, they are typically studied in isolation. Thus, we argue for the need to tackle disinformation detection by taking into account multiple modalities as well as both factuality and harmfulness, in the same framework. Finally, we discuss current challenges and future research directions
1 Introduction
The introduction frames disinformation as both falsehood and intentional harm, increasingly conveyed through multimodal online content. The survey reviews research across modalities and argues that both dimensions should be studied together.
1 Introduction
- Disinformation combines false information with an intent to deceive and harm, while misinformation lacks that intent.The survey treats factuality and harmfulness as distinct, equally important aspects; harmful content can also be factually true, and non-factual content can be harmful.Source: §1 · pp. 1–2
- Multimodal messages can attract more attention and spread further than text alone.Images and videos can be easier to consume and perceived as more credible.Source: §1 · p. 2
- Multimodal disinformation detection remains under-explored, with prior models relying mainly on text, networks, and temporal data.A meta-review found that 91.3% used text, 47.8% network structure, and 26% temporal data; only a handful used images or videos.Source: §1 · p. 2
- The survey covers social media, news, video, and TV-show data, and calls for frameworks that combine modalities with both factuality and harmfulness.It also notes that factuality and harmfulness were correlated in COVID-19 infodemic content, with the relationship varying across languages.Source: §1 · p. 2
- Prior surveys typically focused on factuality or harmfulness, while multimodality was not their main focus.This survey addresses that gap across text, images, speech, video, network structure, and temporal information.Source: §1 · p. 2
2 Multimodal Factuality Prediction
The survey reviews factuality prediction across text, images, speech, video, and network and temporal information. It highlights that modalities may contribute differently across tasks and over time.
2.1 Text
- Textual fact-checking is a substantial research area, but this survey does not detail individual text-only studies.The section situates text claims within multimodal settings that may also involve images, speech, video, and social networks.Source: §2 · p. 2, §2.1 · p. 3
2.2 Image
- Image-based factuality work examines misleading text-image content, fake images, and fauxtography, including image and comment signals.Fauxtography includes false images paired with false claims and true images paired with false claims; doctored images were associated with increased engagement on Twitter and Reddit.Source: §2.2 · pp. 3–4
2.3 Speech/Audio
- Acoustic signals improved performance over text and metadata alone for political-debate factuality prediction, and over text-only inputs for political-bias detection.The survey also describes deception-detection studies combining acoustic, prosodic, lexical, and phonotactic representations.Source: §2.3 · p. 4
2.4 Video
- Video-based deception studies combine transcripts or dialogue with visual cues such as facial expressions, gestures, and frame-level features.The cited courtroom dataset contains 61 deceptive and 60 truthful videos.Source: §2.4 · p. 4
2.5 Network and Temporal Information
- Propagation and interaction networks are used alongside temporal, linguistic, and user features for factuality and rumor-veracity tasks.Rumor Gauge recomputes other features at regular time steps to form multiple time series, and modality contributions can change over time.Source: §2.5 · pp. 4–5
- Temporal features contributed most to fake-news classification in the cited PolitiFact and GossipCop experiments, followed by network and linguistic features.Another framework found social-context features more informative than news-content features.Source: §2.5 · p. 5
3 Multimodal Harmful Content Detection
The survey covers harmful-content detection across text, images, speech, video, and network information. The studies address varied harms, datasets, and modality combinations.
3 Multimodal Harmful Content Detection
- The survey covers harmful content including child abuse material, violence, hate speech, graphic and sexual content, and spam.It organizes studies by content type, modality, data source, annotation, language, and method.Source: §3 · p. 5
3.1 Text
- Text studies range from hate-speech datasets and fine-grained categories to analyses of how hateful posts diffuse.One study found posts from hateful users diffused faster, wider, and with greater outreach than posts from non-hateful users.Source: §3.1 · pp. 5–6
3.2 Image
- Image-related work includes cyberbullying, hateful memes, multimodal hate speech, and propaganda detection.For fine-grained propaganda techniques, VisualBERT improved 19% over image-only ResNet-152 and 11% over text-only BERT.Source: §3.2 · pp. 6–7
3.3 Speech/Audio
- Speech and audio studies address violent sounds, cyberbullying, and cues such as screams and gunshots, but are comparatively few because data are scarce.One study found audio and visual features associated with cyberbullying and complementary to textual features.Source: §3.3 · p. 7
3.4 Video
- Video studies examine cyberbullying, bullying topics, propaganda-related emotional responses, and likely attacks through hateful comments.A multimodal dataset of more than 96 hours of Hollywood and YouTube videos supported the reported benefit of audio-and-image approaches.Source: §3.4 · p. 7
3.5 Network and Temporal Information
- Network analysis addresses coordinated harmful campaigns and user roles, combining profile, textual, and network features.Coordinated Inauthentic Behavior involves activities intended to mislead and manipulate; detection typically analyzes interaction networks, users, and shared content.Source: §3.5 · pp. 7–8
4 Modeling Techniques
The survey describes fusion strategies for combining modalities and learning setups used in disinformation detection. It also reviews supervised and less common unsupervised and semi-supervised approaches.
4 Modeling Techniques
- Multimodal modeling uses early, late, or hybrid fusion to combine features and decisions from different modalities.Early fusion combines low-level features before a single prediction model; late fusion combines unimodal decisions; hybrid fusion combines both strategies.Source: §4 · p. 8
- Learning setups include unsupervised, semi-supervised, supervised, and self-supervised methods.The survey notes that unsupervised approaches are limited and that semi-supervised work leverages unlabeled data because labeled data are scarce.Source: §4 · p. 8
- For propaganda detection, cited self-supervised joint models performed better than other fusion methods, with VisualBERT reported as strongest among the listed models.The comparison covered MMBT, ViLBERT, and VisualBERT.Source: §4 · p. 8
- Supervised approaches include adversarial models for emerging, time-critical fake news and sentiment-aware multimodal embeddings using readers’ comments.The survey also notes autoencoder-based methods as another explored supervised-learning type.Source: §4 · p. 8
5 Major Challenges
The survey identifies challenges in combining diverse modalities and obtaining datasets that cover both factuality and harmfulness. Existing fusion approaches have limits, and cross-modal inconsistency adds difficulty.
Models Combining Multiple Modalities
- Early and late fusion are limited and do not always yield strong results, while current multimodal transformers consider only text and visual modalities.The survey notes that disinformation-related content can involve additional modalities, including speech, video, and networks.Source: §5 · p. 9
- Cross-modal inconsistency in social media content poses another challenge for multimodal detection.The survey also identifies the need for mechanisms that combine modalities so one can complement the others.Source: §5 · p. 9
Datasets
- Appropriate datasets remain difficult to obtain, especially datasets covering diverse modalities and both factuality and harmfulness.The survey also calls for integrating data across platforms such as news, Twitter, Reddit, and Instagram.Source: §5 · p. 9
6 Future Directions
The survey proposes research directions in explainability, broader signals, and knowledge-based methods. These directions address gaps in model interpretation, the limits of content and network signals, and the constrained use of prior fact-checks.
Explainability
- Future work could explain model behavior and provide evidence for why claims are false, while examining what models learn.The survey notes that current explainable fact-checking work focuses more on explaining predictions than model explanations; knowledge-base approaches may fail on new topics such as COVID-19.Source: §6 · p. 9
Beyond Content and Network Signals
- Future multimodal methods could add common-sense and user information because content and network signals omit personal preferences and cultural aspects.The survey also envisions larger models with more heterogeneous inputs and pretraining across a wider variety of tasks.Source: §6 · p. 9
Knowledge-based Method
- Knowledge-based methods could use a shared repository of previously fact-checked claims and harmful content, since some claims are often repeated.The survey identifies existing work on previously fact-checked claims, sentence- and document-level context, and cross-language claim matching.Source: §6 · p. 9
7 Conclusion
The survey reviews multimodal disinformation research with a focus on information that is false and intended to harm. It identifies integration and dataset challenges, and argues for addressing multiple modalities, factuality, and harmfulness together.
7 Conclusion
- The survey covers prior work on multimodal disinformation detection, defining its focus as information that is false and intended to harm.Its stated scope includes the major research topics of factuality and disinformation.Source: §7 · p. 9
- A central challenge is combining modalities that may be unaligned and represented differently, alongside a lack of datasets to foster research.The conclusion gives text, speech, and network structure as examples of differing representations.Source: §7 · p. 9
- The survey argues that disinformation detection should address multiple modalities together with both factuality and harmfulness in one framework.It also points to research directions in addition to highlighting challenges.Source: §7 · p. 9
Limitations
The authors acknowledge that their survey may not cover all relevant work on factuality and harmfulness across modalities.
Limitations
- The survey may omit relevant work on factuality and harmfulness across different modalities.Source: Limitations · p. 10
A Examples of Factuality and Harmful Content
The section illustrates how factuality and harmfulness can appear across textual and visual content, and introduces multimodal modeling approaches.
A Examples of Factuality and Harmful Content
- Examples include content that is both harmful and false, harmful text promoting a bad cure, a true image paired with a false claim, and fear-appealing text-image content.These examples distinguish cases where text and image contribute separately or together.Source: A Examples of Factuality and Harmful… · p. 17
B Modeling Techniques
- The surveyed multimodal approaches include early fusion, late fusion, and joint modal learning; the hybrid approach is not shown.Source: B Modeling Techniques · p. 17
C Lessons Learned
The survey finds that multimodal datasets generally cover only a few modalities and do not combine factuality with harmfulness. It highlights the need to integrate both dimensions and to account for how useful signals vary over time and across modeling approaches.
C Lessons Learned
- The survey argues that detection models should integrate falseness and intentional harmfulness, which are mostly studied in isolation.Examples include fauxotographic content that is harmful and false, and harmful content conveyed jointly through text and image.Source: C Lessons Learned · p. 17, Figure 2 · p. 18
- Most multimodal datasets cover 2–3 modalities, and none addresses both factuality and harmfulness; Alam et al. (2021) covers both only in text.Alam et al. (2021) reports that 56% of Arabic false content is harmful.Source: C Lessons Learned · p. 18
- Early factuality prediction relies more on user and content features, while later analysis benefits more from network and temporal data.As information spreads, propagation features become richer and more informative, while content-feature contributions remain constant.Source: C Lessons Learned · p. 18
- The network illustration distinguishes nodes, ego neighborhoods, triads, and communities as levels or structures for describing users and spreaders.Source: Figure 3 · p. 18
- The illustrated modeling approaches include early and late fusion and joint modal learning, but omit the hybrid combination of early and late fusion.Source: Figure 4 · p. 18
- Combining text, speech, and metadata improved factuality performance over individual modalities; multimodal gains were also observed for hateful memes and propaganda detection.Source: C Lessons Learned · p. 18
D More Challenges
The section identifies challenges involving context, metadata, generalization, evolving topics, transparency, and label granularity. It describes limitations in current systems and directions for addressing them.
D More Challenges
- Most detection methods are noncontextualized, leaving reader responses, user networks, propagation dynamics, and cross-platform mentions outside the broader news context.The section calls for integrating these signals to capture a fuller perspective on news.Source: D More Challenges · p. 19
- News authenticity, author credibility, factuality, and whether an attack is coordinated or individual are relevant metadata for understanding disinformation severity.Source: D More Challenges · p. 19
- Performance is limited by dataset demographics and cultural context; a model trained on Indian political data may not generalize to US health data.Source: D More Challenges · p. 19
- Models may fail on evolving-event topics such as COVID-19 and vaccines, making zero-shot or few-shot learning a future avenue to explore.Source: D More Challenges · p. 19
- Detection models should provide interpretable, unbiased outcomes, including the real news behind a judgment and the counterfeit portion of the information.The section also notes a lack of datasets pairing disinformation explanations with corresponding real information.Source: D More Challenges · p. 19
- Most systems use binary labels, while finer categories such as true, satire, misleading, manipulated, or imposter content can support better decisions.The section proposes multiclass classification, ordinal regression, or regression as alternatives that may help prioritize disinformation for response.Source: D More Challenges · p. 19