Source-linked AI summary
Advanced Machine Learning Techniques for Fake News (Online Disinformation) Detection: A Systematic Mapping Study
Michal Choras, Konstantinos Demestichas, Agata Gielczyk, Alvaro Herrero, Pawel Ksieniewicz, Konstantina Remoundou, Daniel Urda, Michal Wozniak
TL;DR
Fake news threatens societies and individuals, yet detecting it remains challenging. This paper maps advanced ML approaches and related resources, highlighting research challenges including adaptation to changing data and explainability.
Problem
Fake-news detection is challenging, and the paper identifies a need to scrutinize current knowledge, possible solutions, and methodological gaps.
Source: §1 · pp. 1–2Method
The paper critically maps advanced ML-based detection approaches and surveys related datasets, initiatives, projects, and challenges.
Source: §1 · p. 2, §5 · p. 17Results
The mapping covers text, reputation, network, and image-manipulation analysis, alongside datasets and open challenges.
Source: §1 · p. 2, §5 · p. 17Takeaways & Limitations
The paper identifies streaming adaptation and explainability as challenges for ML-based fake-news detection systems.
Source: §5.1 · p. 17, §5.3 · p. 18Only a few papers address streaming data, although detection systems need mechanisms to adapt to concept drift.
Source: §5.1 · p. 17
Abstract
from arXiv · showhide
Fake news has now grown into a big problem for societies and also a major challenge for people fighting disinformation. This phenomenon plagues democratic elections, reputations of individual persons or organizations, and has negatively impacted citizens, (e.g., during the COVID-19 pandemic in the US or Brazil). Hence, developing effective tools to fight this phenomenon by employing advanced Machine Learning (ML) methods poses a significant challenge. The following paper displays the present body of knowledge on the application of such intelligent tools in the fight against disinformation. It starts by showing the historical perspective and the current role of fake news in the information war. Proposed solutions based solely on the work of experts are analysed and the most important directions of the application of intelligent systems in the detection of misinformation sources are pointed out. Additionally, the paper presents some useful resources (mainly datasets useful when assessing ML solutions for fake news detection) and provides a short overview of the most important R&D projects related to this subject. The main purpose of this work is to analyse the current state of knowledge in detecting fake news; on the one hand to show possible solutions, and on the other hand to identify the main challenges and methodological gaps to motivate future research.
1. Introduction
Fake news threatens societies and individuals, while its definition and detection remain challenging. This paper critically maps advanced ML approaches and related resources to describe possible solutions and motivate future research.
1. Introduction
- The paper complements prior surveys with a focus on advanced ML, its own method categorization, datasets, initiatives, projects, and future challenges.It reviews methods analyzing text, reputation, networks, and image manipulation.Source: §1 · p. 2
1.1. A historic perspective
- Disinformation predates modern media, but internet and social-media adoption have enabled its spread at unprecedented scale across elections, healthcare, climate, and other topics.The paper traces examples from ancient times through print-era hoaxes and wartime propaganda.Source: §1.1 · p. 3
1.2. Overview of definitions: what is meant by fake news?
- Definitions of fake news vary, and automatic classification must distinguish deliberate deception from irony or satire despite their blurred boundary.The paper also notes that conspiracy theories are not always considered fake news; text or images and the source’s motivation may matter.Source: §1.2 · p. 4, Figure 3 · p. 4
1.3. Why is fake news dangerous?
- Examples range from false COVID-19 claims about causes and cures to a mistaken election-era claim about buses, which spread widely before being labelled false.The paper presents these as examples of disinformation’s potential impact on citizens and political discourse.Source: §1.3 · pp. 4–5
2. Current initiatives worldwide to fight against disinformation
Responses to disinformation span governmental, regulatory, educational, technological, and community initiatives. The paper also highlights a deepfake-detection challenge: the contest-winning model performed poorly on new data.
2.1. Large-scale political initiatives addressing the issue of fake news
- The IGC concluded that global technology firms cannot alone combat harmful content, hate speech, and electoral interference, making self-regulation insufficient.The committee called for collaboration with national audiovisual regulators, independent fact-checkers, and researchers.Source: §2.1 · pp. 5, 7
- The EU action plan includes stronger cooperation and a Rapid Alert System for real-time alerts on disinformation campaigns.It also calls for improved detection capabilities, platform collaboration, and public awareness and resilience.Source: §2.1 · pp. 6–7
- Initiatives can be understood through three angles: technological, legislative, and educative.The paper applies this framing to actions proposed by the IGC, Australia, the US, and the EU.Source: §2.1 · p. 7
- The IGC identified a continuing challenge in reconciling internet regulation with freedom of speech under national laws.The paper also describes national and international regulatory proposals, alongside education initiatives such as media literacy programs.Source: §2.1 · pp. 7–8
2.2. Other noteworthy initiatives and solutions
- Community volunteers called Lithuanian “elves” scan social media and report fake information, while Facebook described supervised classifiers and image processing for policy-violating content.Facebook said its tools complement 35,000 human supervisors; the AI Foundation’s Reality Defender was intended to help users identify fake content, but technical details were unavailable.Source: §2.2 · p. 8
- The contest-winning deepfake detector achieved 65.18% precision on new data, leaving robust detection an open challenge.The result was reported for the best model on published data when validated with new data.Source: §2.2 · p. 8
3. A systematic overview of ML approaches for fake news detection
The paper critically reviews ML-based fake-news detection across four kinds of input: text, reputation, network data, and image manipulation.
3. A systematic overview of ML approaches for fake news detection
- The review organizes ML-based detection approaches around text and NLP, reputation, network analysis, and image-manipulation recognition.These approaches analyze different types of digital content, as represented in Fig. 1.Source: §3 · p. 9
- The section presents a comprehensive, critical analysis of previous ML-based approaches to false-news detection.Source: §3 · p. 9
- The overview includes reputation analysis alongside content-focused methods such as text, network data, and image manipulation.Source: §3 · p. 9
3.1. Text analysis
Text-based detection spans simple word representations, psycholinguistic and syntactic features, and methods that incorporate non-linguistic information. Prior work reports results for these approaches, including a 90% F-score for one psycholinguistic system.
3.1. Text analysis
- Text analysis ranges from bag-of-words and N-grams to psycholinguistic, syntactic, and semantic features.The paper notes that social context matters, while features extracted directly from article content are a basic source for pattern-recognition systems.Source: §3.1 · p. 9
3.1.1. NLP-based data representation
- PHEME experiments found that combining unigrams and bigrams was effective for Twitter disinformation classification using TF or TF-IDF features and five base classifiers.The passage describes this as a common starting point for further classifier and ensemble studies.Source: §3.1.1 · p. 9
3.1.2. Psycholinguistic features
- Psycholinguistic features studied include long sentences, limited vocabulary, repetition, reduced formality, and measurable language cues.The text describes these factors as potentially useful in designing pattern-recognition systems.Source: §3.1.2 · p. 9
- A system using 19 attributes and Multinomial Naive Bayes achieved a 90% F-score.The attributes included grammatical information and sentiment or slang cues.Source: §3.1.2 · p. 10
3.1.3. Syntax-based
- Syntactic methods represent sentence structure with Probability Context Free Grammars and can also model whole discussion structures.The paper describes discussion-structure extraction as promising for combating fake-news propagation [51] [52].Source: §3.1.3 · p. 10
3.1.4. Non-linguistic methods
- Non-linguistic approaches examine message creators and readers, social-media context, images, author behavior, and stylometric metrics.Other cited approaches include automatic news verification using image data and streaming analysis of social posts.Source: §3.1.4 · p. 10
- Some studies treat false-news detection as classification, while others consider anomaly detection, clustering, or credibility regression.Source: §3.1.4 · pp. 10–11
3.2. Reputation analysis
Reputation analysis uses scores derived from sources, content, and online behavior to assess trustworthiness. The reviewed systems report detection results, with some passages noting training or evaluation constraints.
3.2. Reputation analysis
- Reputation estimates can be represented numerically as scores, computed centrally or through local or global trust metrics.For news, reputation can come from content and reviews or from IP and domain information.Source: §3.2 · p. 11
- Reputation indicators include publisher anonymity, time spent on a site, and domain popularity; the survey reported that trustworthy domains were more popular than disinformation domains.Source: §3.2 · p. 11
- The MED classifier scores webpages using a reputation vector of domain, hosting, IP, popularity, and runtime features.Notos uses DNS characteristics and clustering across network-, zone-, and evidence-based features to score domains.Source: §3.2 · p. 11
- Notos achieved a 96.8% true positive score and 0.38% false positive score on real-world DNS traffic from vast ISP networks.Its domain reputation scores used clustering of network-based, zone-based, and evidence-based features.Source: §3.2 · pp. 11–12
- A deep neural model trained on 567 million aggregated passive DNS queries identified 95% of suspicious hosts with a 1:1000 false positive rate.Training took exceptionally long because of the data volume, and delay information was not assessed.Source: §3.2 · p. 12
- Segugio reached an 85% true positive rate and a false positive rate below 0.1%, but these rates were based on only 53 new domains.Source: §3.2 · p. 12
- GSVM-BA repeatedly removes positive support vectors to find an optimum decision boundary, using breadth and spectral feature vectors [69].Source: §3.2 · p. 12
3.3. Network data analysis
Network-based detection evaluates news using social-context data from propagation and related sources, representing relations as homogeneous or heterogeneous networks. The section describes feature extraction and learning, alongside network-pattern and tri-relationship models.
3.3. Network data analysis
- Homogeneous networks model one kind of node and edge, while heterogeneous networks can encode multiple kinds and relations.Credibility networks use social-media viewpoints to assess news; knowledge, stance, and interaction networks are examples of heterogeneous networks.Source: §3.3 · p. 12
- Network-based detection formalizes evaluating a news item’s truth value as classification using extracted features and learned models.Network representation learning can draw on friendship networks, temporal user engagement, and interaction networks; knowledge networks can support path-finding and flow-optimization operations.Source: §3.3 · p. 13
- False-news detection uses network patterns in dissemination, spreaders, and their relations, including patterns found at node, ego, triad, community, and network levels.The cited study reports that, compared with true news, false news can spread further, engage more spreaders, and involve spreaders who are more absorbed and more closely connected.Source: §3.3 · p. 13
- TriFN models relations among news articles, spreaders, and publishers through entity representation, relation modeling, and semi-supervised learning [72].The passage reports that TriFN performs well at detecting news in its initial stages of dissemination [72].Source: §3.3 · p. 13
3.4. Image based analysis and detection of image manipulations
Image-based approaches address manipulation ranging from copy-move and splicing to colorization and other edits, using image features and classifiers to detect or localize tampering. The methods vary in their image representations, descriptors, and classification strategies.
3.4. Image based analysis and detection of image manipulations
- Image-editing tools can alter semantic content without visible clues, motivating detection methods for copy-move, splice, inpainting, and image-wise modifications.The passage describes image manipulation as making fake news easier to spread.Source: §3.4 · p. 13
- CNNs extract camera-model fingerprints to identify images assembled from multiple devices and locate altered regions.The approach assumes untampered pixels appear to have been captured by a single device.Source: §3.4 · p. 13
- A cloud-based modification detector uses SSIM instead of MSE or PSNR; its authors claim cloud deployment improves speed and accessibility.Source: §3.4 · pp. 13–14
- The OELTP descriptor extracts features from overlapping YCbCr blocks, reduces dimensionality using feature energy, and supplies sorted features to an SVM.Source: §3.4 · p. 14
- For JPEG tampering, double-quantization artifacts in DCT coefficients can arise when differently compressed image regions are merged and resaved.Source: §3.4 · p. 14
- A grayscale image method combines HOG, DWT, and LBP features in a Logistic Regression model to distinguish manipulated from authenticated pictures.The authors attribute improved detection correctness to combining spatial- and frequency-based features.Source: §3.4 · p. 14
- Haar wavelets with HOG and LBP variance use Euclidean distance, while SURF and BRISK descriptors target single and multiple copy-move forgeries.The passage describes SURF as robust to rotation, blurring, and additive noise, and BRISK as robust for scale-invariant forged regions and poorly localized keypoints.Source: §3.4 · p. 14
- Block statistics with DWT coefficients and SVMs detect contrast enhancement, while HSV histogram and statistical features assess colorization authenticity.Source: §3.4 · p. 14
4. Research interest and popular datasets used for detecting fake news
The paper describes growing research activity and EU initiatives, then surveys resources for evaluating image tampering and fake-news detection. The datasets span manipulated images and videos, labeled claims, news articles, and social context.
4.1. Is fake news an important issue for the research society?
- Scopus publications on “fake news detection” rose from 5 in 2016 to 371 in 2019; Web of Science rose from 4 to 86 and IEEExplore from 3 to 133 over those years.Source: §4.1 · p. 14
- The paper lists EU-funded research projects compiled from CORDIS in July 2020 after searches for fake news, disinformation, and deepfake, and concludes that increasing amounts of economic resources are being invested.Source: §4.1 · pp. 15–16, Table 2 · p. 16
4.2. Image tampering datasets
- Image-tampering resources include CASIA versions with increasingly realistic post-processing, varied splice and copy effects, and CG-1050’s 1,050 tampered images with masks.The Palermo sets distinguish translated, rotated, and scaled copy-paste alterations.Source: §4.2 · pp. 16–17
- The Deepfake Detection Challenge Dataset contains 124,000 videos modified using eight facial-modification algorithms [18].Source: §4.2 · p. 17
4.3. Fake news datasets
- Fake-news datasets include LIAR’s nearly 13,000 manually labeled statements, an unbalanced political-news collection, and CREDBANK’s roughly 60 million crowd-checked tweets [89].LIAR uses six truth labels; CREDBANK covers more than 1,000 events over 96 days, with each event checked by 30 editors.Source: §4.3 · p. 17
- FakeNewsNet combines news and social-background details, ISOT provides over 12,600 true and false items each, and a multimodal approach uses textual and visual features.Source: §4.3 · p. 17
5. Conclusions, further challenges and way forward
The conclusion draws on a systematic mapping of fake-news detection research and identifies open challenges for future work. It highlights streaming, lifelong learning, explainability, and deepfakes as important considerations.
5.1. Streaming nature of fake news
- Most studies ignore streaming; Wang and Terano's method is limited to short streams, while Ksieniewicz et al. report usefulness on real-life datasets.Fake-news profiles can shift as spreaders alter message characteristics to evade detection, a change known as concept drift.Source: §5.1 · p. 17
5.2. Lifelong learning solutions
- Lifelong learning fits rapidly changing fake-news content, but systems must balance learning new information against retaining prior knowledge and avoiding catastrophic forgetting.The paper identifies feature modeling, storing and transferring prior learning, updating knowledge, and user feedback as features of such systems.Source: §5.2 · p. 18
5.3. Explainability of ML-based fake news detection systems
- Black-box outputs lack explanations, yet trust across journalism and security communities depends on decision-makers understanding why a system gives an answer.The paper describes explainability as the extent to which a person can understand and explain a system's internal mechanics.Source: §5.3 · p. 18
5.4. The emergence of deepfakes
- Deepfakes span face-swapped images and videos and simulated voices; one portrait-video detector reported 97.29% accuracy using biological-signal patterns [104].The paper also calls for legal, educational, and political contributions; identification rates must rise as disinformation grows more sophisticated.Source: §5.4 · pp. 18–19
5.5. Final remarks
- The study systematically maps research, projects, initiatives, and resources including methods and datasets, then identifies unresolved challenges to motivate further work.Source: §5.5 · p. 19