Pang, Ervyn Zhen jun (2025) Utilizing image file properties and image content for advance fake news detection. Masters thesis, Sunway University.
Full text not available from this repository.Abstract
Fake news on social media continues to significantly impact society, often appearing in various formats such as misleading text, manipulated images, or deepfake videos. To combat this, researchers have explored multimodal approaches that combine both textual and visual data. This study proposes a novel method that extracts both Image File Properties (e.g., metadata) and Image Content (converted to text using the Google Cloud Vision API) to improve fake news detection. These features are then integrated into traditional text-based models for evaluation. At the time of writing, the use of Image File Properties and the transformation of Image Content into text format remains underexplored in the domain of fake news detection. Unlike previous studies that either focused on metadata in unrelated domains or employed CNNs to extract visual features directly, this research explores the combined effect of Image File Properties and textually extracted Image Content within a fake news detection framework. To achieve this, the Fakeddit dataset was used, as it includes both textual content from Reddit posts and their corresponding images. Statistical tests, including Point-Biserial Correlation and Chi-square analysis, were conducted to evaluate feature significance. Both Linear Regression and Random Forest models were applied to address Research Questions 1 and 2 (RQ1–RQ2), examining the individual impact of each image-derived component. RQ3 assessed the collective effect of combining these features, while RQ4 tested different machine learning methods and data partition strategies to identify the most effective ensemble model. Findings reveal that Image File Properties alone are not statistically significant predictors (RQ1), while Image Content provides valuable features for fake news classification (RQ2). When both components are combined, model performance improves (RQ3). Notably, the ensemble model developed in RQ4 outperforms individual models and effectively reduces false positives, an important consideration to prevent the incorrect flagging of legitimate news content. This research demonstrates the value of incorporating previously underutilized imagederived features into fake news detection models, offering a novel direction for future multimodal systems. However, as this is a newly explored area, further validation using datasets from different social media platforms is necessary to generalize the findings. Future work may also investigate integrating Large Language Models (LLMs) for deeper semantic reasoning, providing additional opportunities for feature enrichment and improved model accuracy. Additionally, future work may explore how Large Language Models (LLMs) can handle predictions involving Image File Properties and Image Content, and whether they can be used to explain the complex relationships between these variables.
| Item Type: | Thesis (Masters) |
|---|---|
| Uncontrolled Keywords: | ensemble model; fake news detection; image metadata; AI in journalism; disinformation |
| Subjects: | P Language and Literature > PN Literature (General) T Technology > TR Photography Z Bibliography. Library Science. Information Resources > ZA Information resources |
| Divisions: | Sunway University > Faculty of Engineering and Technology [formerly School of Engineering and Technology until 14 November 2025; School of Science and Technology until 2020] > School of Engineering [formerly Dept. of Engineering until 14 March 2025] |
| Depositing User: | Ms Yong Yee Chan |
| Date Deposited: | 26 Jul 2026 07:55 |
| Last Modified: | 26 Jul 2026 07:56 |
| URI: | http://repository.sunway.edu.my/id/eprint/3554 |
Actions (login required)
 |
View Item |