Fake News Detection and Recommendation System Using BERT and Federated Learning

Abstract

The rapid spread of misinformation on digital plat-forms necessitates effective news verification systems. This project presents a Hybrid Fake News Detection System that combines BERT-based text classification with semantic similarity analysis using Sentence-BERT (SBERT). The BERT model analyzes linguistic patterns to classify news as real or fake, while SBERT compares the news with content obtained through web scraping from trusted sources to validate its authenticity. A combined decision mechanism improves overall detection accuracy. The system also incorporates federated learning to enhance news recommendation while preserving user privacy by avoiding centralized data sharing. Additionally, it provides social features such as posting, liking, commenting, and following, along with an admin module for monitoring and controlling misinformation. The application is developed using HTML, CSS, and JavaScript for the frontend and Python Django for the backend. This approach improves both reliability and privacy compared to traditional methods.

Introduction

The text describes a Hybrid Fake News Detection System designed to combat the growing spread of misinformation on digital and social media platforms. Fake news is a major issue because it can mislead people and influence public opinion, while traditional manual or rule-based detection methods are too slow and ineffective for large-scale data.

To solve this, the proposed system uses advanced AI and NLP techniques, combining:

• BERT for contextual text classification (real vs fake news)
• Sentence-BERT (SBERT) for semantic similarity comparison with real news from trusted sources
• Web scraping to validate news against real-time information

The system works by first preprocessing user-submitted news, then analyzing it using BERT to detect fake or real content. At the same time, it compares the news with external trusted headlines using SBERT embeddings and cosine similarity. A hybrid decision mechanism combines both results to improve accuracy and reliability.

The system is built using a Django backend with a web frontend (HTML, CSS, JavaScript). It includes:

• A User module for posting, interacting, and viewing news
• An Admin module for monitoring users, removing fake content, and enforcing rules
• A federated learning-based recommendation system that personalizes news while preserving user privacy

The architecture ensures:

• Real-time fake news detection
• User interaction features (likes, comments, sharing)
• Secure and privacy-preserving learning

Existing research shows that while BERT and deep learning models perform well in text classification, they alone are insufficient because fake news can mimic real news. Similarly, semantic similarity alone is not enough. The key research gap is the lack of systems combining both approaches effectively.

Conclusion

In this paper, we proposed a Hybrid Fake News Detection System that combines BERT-based classification with SBERT-based semantic similarity analysis to improve the accuracy and reliability of fake news detection. By integrating real-time web scraping, the system validates user-submitted news against trusted sources, reducing the chances of misclassification caused by misleading writing styles. The inclusion of a hybrid decision mechanism enables the system to leverage both contextual understanding and real-world verification, resulting in better performance compared to standalone approaches. Furthermore, the integration of federated learning allows personalized news recommendation while preserving user privacy, addressing critical concerns in modern data-driven applications. The addition of user and admin modules ensures effective interaction, monitoring, and control of misinformation within the platform. Experimental results demonstrate that the hybrid approach outperforms individual detection methods in terms of accuracy and reliability. The system is scalable, efficient, and suitable for real-time applications. In future work, we aim to enhance the model by incorporating larger and more diverse datasets, improving real-time performance, and exploring advanced deep learning techniques for better accuracy. Additionally, optimizing the recommendation system and strengthening misinformation detection using multimodal data (such as images and videos) can further improve the overall effectiveness of the platform.

References

[1] T. H. C. Chiang, C. S. Liao, and W. C. Wang, “Investigating the Difference of Fake News Source Credibility Recognition between ANN and BERT,” Journal of Information Technology and Media Studies, vol. 9, no. 2, pp. 45–52, 2022. [2] V. D. Husza´r, V. K. Adhikarla, I. Ne´gyesi, and C. Krasznay, “Toward Fast and Accurate Violence Detection for Automated Video Surveillance Applications,” IEEE Access, vol. 11, pp. 18772–18793, 2023. [3] M. Khan, A. E. Saddik, W. Gueaieb, G. De Masi, and F. Karray, “VD-Net: An Edge Vision-Based Surveillance System for Violence Detection,” IEEE Access, vol. 12, pp. 43796–43808, 2024. [4] B. Wang and X. Zhang, “An Efficient Violence Detection Method Based on Temporal Attention Mechanism,” Instrumentation, vol. 10, no. 2, 2023. [5] V. Lomte, S. Singh, S. Patil, and D. Pahurkar, “Video Anomaly Detection through Live Surveillance,” IJRESM, vol. 2, no. 4, 2019. [6] M. Ahmed et al., “Real-Time Violent Action Recognition Using Key Frames Extraction and Deep Learning,” 2021. [7] Z. Trabelsi et al., “Real-Time Attention Monitoring System for Class-room: A Deep Learning Approach for Student’s Behavior Recognition,” Big Data and Cognitive Computing, vol. 7, no. 1, 2023. [8] J. H. Piater, S. Richetto, and J. L. Crowley, “Event-Based Activity Analysis in Live Video Using a Generic Object Tracker,” in Proc. IEEE Int. Workshop on Performance Evaluation of Tracking and Surveillance, 2002, pp. 1–8. [9] J. Xiao, S. Li, and Q. Xu, “Video-Based Evidence Analysis and Extraction in Digital Forensic Investigation,” IEEE Access, vol. 7, pp. 55432–55442, 2019. [10] A. Patwardhan and G. Knapp, “Aggressive Actions and Anger Detection from Multiple Modalities Using Kinect,” arXiv preprint arXiv:1607.01076, 2016. [11] L. Zhang, X. Ruan, and J. Wang, “WiVi: A Ubiquitous Violence Detection System With Commercial WiFi Devices,” IEEE Access, vol. 8, pp. 66626–66672, 2020. [12] B. Wang and X. Zhang, “An Efficient Violence Detection Method Based on Temporal Attention Mechanism,” Instrumentation, vol. 10, no. 2, 2023. [13] D. Tsiktsiris, A. Lalas, M. Dasygenis, and K. Votis, “Multimodal Abnormal Event Detection in Public Transportation,” IEEE Access, vol. 12, pp. 133469–133480, 2024. [14] A. Dembele, R. W. Mwangi, and A. O. Kube, “A Lightweight Convo-lutional Neural Network with Hierarchical Multi-Scale Feature Fusion for Image Classification.” [15] A. G. Howard et al., “Searching for MobileNetV3,” in Proc. IEEE/CVF ICCV, 2019. [16] B. Jacob et al., “Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference,” in Proc. IEEE CVPR, 2018, pp. 2704–2713. [17] H. Fan, K. Zhao, and X. Wang, “VideoMix: Rethinking Data Augmentation for Video Classification,” arXiv preprint, 2020. [18] A. Arnab et al., “Video Vision Transformers for Violence Detection,” arXiv preprint, 2022. [19] C. Wang and Y. Liu, “Exploring Temporally Dynamic Data Augmentation for Video Recognition,” 2022. [20] M. Patel, “Real-Time Violence Detection Using CNN-LSTM,” 2021. [21] J. Wang et al., “DETRs Beat YOLOs on Real-Time Object Detection,” arXiv preprint, 2023.

Copyright

Copyright © 2026 Adithya A A, Goutham Krishnan, Hana Habeeb Rahman , Sylasree V A, Reshma P D. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.