FNR: A Similarity and Transformer-Based Approach to Detect Multi-Modal Fake News in Social Media

Ghorbanpour, Faeze; Ramezani, Maryam; fazli, MohammadAmin; Rabiee, Hamid R.

doi:10.36227/techrxiv.17099096.v1

Cited by 1 publication

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, they encountered difficulties capturing multi-modal inconsistencies because of the semantic gap between the two types of features [37]. Ghorbanpour et al [38] proposed the Fake-News-Revealer (FNR) method, which uses a Vision-transformer [39] and BERT [5] to extract image and text features respectively. The model extracted textual and visual features separately and determined their similarities by loss.…”

Section: ) Problems In Multi-modalitymentioning

confidence: 99%

Utilizing Ensemble Learning for Detecting Multi-Modal Fake News

Luqman,

Faheem,

Ramay

et al. 2024

IEEE Access

View full text Add to dashboard Cite

The spread of fake news has become a critical problem in recent years due extensive use of social media platforms. False stories can go viral quickly, reaching millions of people before they can be mocked, i.e., a false story claiming that a celebrity has died when he/she is still alive. Therefore, detecting fake news is essential for maintaining the integrity of information and controlling misinformation, social and political polarization, media ethics, and security threats. From this perspective, we propose an ensemble learning-based detection of multi-modal fake news. First, it exploits a publicly available dataset Fakeddit consisting of over 1 million samples of fake news. Next, it leverages Natural Language Processing (NLP) techniques for preprocessing textual information of news. Then, it gauges the sentiment from the text of each news. After that, it generates embeddings for text and images of the corresponding news by leveraging Visual Bidirectional Encoder Representations from Transformers (V-BERT), respectively. Finally, it passes the embeddings to the deep learning ensemble model for training and testing. The 10-fold evaluation technique is used to check the performance of the proposed approach. The evaluation results are significant and outperform the state-of-the-art approaches with the performance improvement of 12.57%, 9.70%, 18.15%, 12.58%, 0.10, and 3.07 in accuracy, precision, recall, F1-score, Matthews Correlation Coefficient (MCC), and Odds Ratio (OR), respectively.

show abstract