A Hybrid AI-Powered Framework for Real-Time Fake News Detection with Multi-Modal Verification

A Hybrid AI-Powered Framework for Real-Time Fake News Detection with Multi-Modal Verification

Dr. Adilakshmi Yannam

Department of CSE (AIML) S R Gudlavalleru Engineering College Gudlavalleru, India

Mr. Karthik Reddy Yaramala

Department of CSE (AIML) S R Gudlavalleru Engineering College Gudlavalleru, India

Ms. Gouri Niteesha Tunuguntla

Department of CSE (AIML) S R Gudlavalleru Engineering College Gudlavalleru, India

Mr. Bhargav Vallapuram

Department of CSE (AIML) S R Gudlavalleru Engineering College Gudlavalleru, India

Mr. Bhuvaneswar Reddy Nune

Department of CSE (AIML) S R Gudlavalleru Engineering College Gudlavalleru, India

Abstract – The proliferation of misinformation in digital media poses significant threats to social stability, public health, and democratic processes worldwide. Existing fake news detection systems typically rely on static machine learning models that cannot adapt to evolving misinformation tactics, or on manual fact-checking processes that lack scalability. This paper presents FakeGuard, a novel hybrid framework that integrates a fine- tuned BERT-based transformer model with a large language model (Google Gemini 2.5 Flash) and real-time news aggregation for accurate, explainable, and scalable fake news detection. The system architecture comprises three core components: a text classification pipeline using a pre-trained BERT model fine- tuned on fake news datasets; a social media analysis module leveraging Gemini AI for contextual understanding and multi- factor reasoning; and a real-time news verification engine that ag- gregates live articles from NewsAPI and GNews, performs source credibility scoring, and applies ML-based content analysis. A confidence-based fallback mechanism ensures system robustness by dynamically switching between the LLM and BERT model during service degradation. Experimental evaluation on the ISOT Fake News Dataset demonstrates that the BERT component achieves 98.7% accuracy and an F1-score of 0.987, while the hybrid system maintains average inference latency of 1.42 seconds for text analysis under steady-state conditions. The complete system is deployed as a production-ready web application with FastAPI backend and React frontend, demonstrating practical viability for real-world misinformation detection.

Index TermsFake News Detection, BERT, Transformer Mod- els, Large Language Models, Gemini AI, Hybrid AI Systems, Real-time News Verification, Misinformation, Machine Learning, NLP

1. INTRODUCTION

   The digital information ecosystem has experienced unprece- dented growth, with social media platforms and online news sources becoming primary information channels for billions of users worldwide. According to recent studies, over 4.9 billion people actively use social media, and more than 60% of adults globally consume news through digital platforms. However, this democratization of information has been accompanied by a parallel rise in misinformation, disinformation, and fake news.

   The consequences are severe: eroded public trust, polarized societies, compromised public health responses during pan- demics, and threatened democratic processes.

   Traditional approaches to fake news detection fall into two distinct categories. The first approach involves manual fact- checking by human experts and journalism organizations. While this method achieves high accuracy, it cannot scale to the volume of content generated dailyover 500 million tweets and 4 billion Facebook posts each day. The second approach employs automated machine learning-based systems, which offer scalability but suffer from critical limitations: static models become outdated as misinformation tactics evolve; black-box classifiers lack explainability, reducing user trust; most systems operate in isolation without real-world evidence validation; and they demonstrate limited ability to analyze social media context and conversational nuances.

   Despite significant advances in natural language process- ing for fake news detection, existing solutions lack a uni- fied framework that combines four essential capabilities: the accuracy and scalability of specialized transformer models; the contextual reasoning and explainability of large language models; real-time evidence validation from live news sources; and production-ready architecture for practical deployment.

   This paper addresses these gaps through the following contributions:

   1. We propose FakeGuard, a novel hybrid framework that synergistically integrates a fine-tuned BERT clas- sifier with Google Gemini 2.5 Flash for multi-modal fake news detection. This represents the first production- ready system combining specialized ML models with general-purpose LLMs in a unified architecture.
   2. We introduce a real-time news verification pipeline that dynamically aggregates articles from multiple sources (NewsAPI, GNews), computes source credibility scores, and performs ML-based content analysis, en- abling detection systems to adapt to emerging events

      and evolving misinformation patterns.

   3. We design a confidence-based fallback mechanism that ensures system robustness: the LLM serves as the primary analyzer for social media content, with automatic fallback to the BERT model during API degradation or high-latency conditions. Experimental results show this reduces 95th percentile latency by 43% while maintaining 96.2% classification accuracy during degraded conditions.
   4. We achieve state-of-the-art performance on the ISOT Fake News Dataset, with the BERT component reaching 98.7% accuracy, 0.987 precision, 0.987 recall, and 0.987 F1-score, outperforming baseline models including Lo- gistic Regression (89.9%), Random Forest (98.2%), and XGBoost (98.3%).
   5. We provide a comprehensive open-source implemen-

   tation including backend APIs, frontend interface, and deployment configurations, enabling reproducibility and further research by the community.

   The remainder of this paper is organized as follows: Section II reviews related work in fake news detection and hybrid AI systems. Section III details the proposed FakeGuard archi- tecture and implementation methodology. Section IV presents experimental setup, datasets, and evaluation metrics. Section

   V presents experimental results. Section VI discusses key findings and limitations. Section VII concludes with future research directions.

2. RELATED WORK
   1. Machine Learning for Fake News Detection

      Machine learning-based approaches for fake news detection have been extensively studied over the past decade. Early work focused on traditional classifiers using handcrafted linguistic features. Rubin et al. employed support vector machines with features derived from satire detection, achieving moderate success but limited generalization. Subsequent research ex- plored ensemble methods, with Conroy et al. demonstrating that Random Forest classifiers outperform individual models by combining lexical, syntactic, and semantic features.

      The introduction of deep learning revolutionized the field. Wang proposed a convolutional neural network architecture for stance detection, while Ma et al. utilized recurrent neural net- works with attention mechanisms to capture temporal patterns in news propagation. However, these models remained con- strained by their reliance on static training data and inability to incororate real-world evidence.

   2. Transformer-Based Approaches

      The advent of transformer architectures marked a significant advancement. BERT (Bidirectional Encoder Representations from Transformers) and its variants have demonstrated ex- ceptional performance on various NLP tasks, including fake news detection. Kaliyar et al. introduced FakeBERT, a fine- tuned BERT model achieving 98.4% accuracy on benchmark datasets. Similarly, Jwa et al. developed a BERT-based model

      for detecting fake news in Korean language, highlighting the models adaptability across linguistic domains.

      The model employed in this work, jy46604790/fake-news- bert-detect, builds upon these foundations, offering a pre- trained checkpoint specifically optimized for fake news classi- fication tasks. Its architecture comprises 12 transformer layers with 768 hidden dimensions, processing input sequences up to 512 tokens.

   3. Large Language Models for Explainability

      Recent advances in large language models have opened new possibilities for explainable AI. GPT-3, ChatGPT, and Gemini have demonstrated remarkable capabilities in contextual un- derstanding and reasoning. Bang et al. evaluated ChatGPTs

      performance on misinformation detection, finding that while LLMs provide coherent explanations, their classification accu- racy varies significantly across domains.

      The integration of LLMs with specialized classifiers repre- sents an emerging research direction. Zhang et al. proposed a hybrid system combining BERT with GPT-3 for fake news detection, achieving improved explainability with minimal ac- curacy loss. Our work extends this paradigm by implementing a production-ready system with dynamic fallback mechanisms and real-time news verification.

   4. Real-Time News Verification Systems

   Real-time verification systems have gained attention as a solution to the dynamic nature of misinformation. Hassan et al. developed ClaimBuster, a real-time fact-checking system for political claims using knowledge graphs. More recently, Popat et al. introduced a system that combines web search with stance detection for claim verification.

   Cloud-based implementations leveraging services like AWS and serverless architectures have been explored by Amazon Web Services, demonstrating the feasibility of scalable, event- driven security and verification systems. Our work adapts these cloud-native principles to the fake news detection domain.

   Despite these advances, no existing system provides the complete integration of specialized ML models, LLM-powered reasoning, real-time news verification, and production-ready deployment that FakeGuard offers.

3. PROPOSED SYSTEM ARCHITECTURE
   1. Overall Architecture

      FakeGuard employs a modular, microservices-based archi- tecture deployed on cloud infrastructure. The system comprises four primary layers as illustrated in Fig. 1.

      The presentation layer implements a responsive React.js frontend with Tailwind CSS, providing users with three analy- sis modes: text classification, social media content evaluation, and news topic verification. The API gateway layer, built with FastAPI (Python 3.9+), handles request routing, validation using Pydantic schemas, rate limiting, and error management.

      Fig. 1: High-level architecture of FakeGuard showing the presentation layer, API gateway, intelligence layer, and data layer components.

   2. Intelligence Layer Components
      1. BERT-Based Classification Engine: The core classifier utilizes the HuggingFace transformers library to load and execute the jy46604790/fake-news-bert-detect model. The model processes input text through the following pipeline:
         1. Tokenization using BERT tokenizer with max length 512
         2. Sequence truncation and padding
         3. Forward pass through transformer layers
         4. Softmax activation for binary classification

            The model is loaded using a singleton pattern to ensure memory efficiency, with configuration parameters:

            TABLE I: BERT Model Configuration Parameters

            Parameter	Value
            Model Name Architecture Max Sequence Length Output Classes Device	jy46604790/fake-news-bert-detect BERT-base (12 layers, 768 hidden) 512 tokens 2 (FAKE/REAL) CPU (GPU optional)
            Precision	FP32

      2. LLM Integration with Gemini 2.5 Flash: For social media content analysis, the system integrates Google Gemini

         2.5 Flash through a structured prompt engineering framework. The prompt template is designed to extract:

         • Factual accuracy assessment
         • Misinformation indicator detection
         • Context completeness evaluation
         • Source credibility signals
         • Overall classification with confidence score The prompt template is defined as:
      3. Confidence-Based Fallback Mechanism: A critical innovation in FakeGuard is the dynamic fallback controller that

         monitors LLM performance and switches to BERT-based classification when necessary. The fallback algorithm is presented in Algorithm 1.

         Algorithm 1 Fallback Controller Logic

         Require: Input text x, threshold rlatency = 2000ms, rerror = 0.05

         Ensure: Classification result y, confidence c start currentTime()

         resultllm callGeminiAPI(x)

         latency currentTime() start

         if resultllm is valid AND latency < rlatency then

         return resultllm

         else

         resultbert loadBERTModel()

         prediction resultbert.predict(x)

         return prediction

         end if

   3. Real-Time News Verification Pipeline

      The news verification module implements a multi-stage pipeline for evidence-based validation:

      1. Topic Extraction: User query processing or random topic generation from curated list
      2. Parallel Article Fetching: Concurrent queries to News- API and GNews APIs
      3. Source Credibility Scoring: Domain authority assess- ment based on:
         • Domain age and reputation
         • Fact-checking organization ratings
         • Historical accuracy records
      4. Content Analysis: BERT-based classification of each retrieved article
      5. Aggregation: Weighted credibility score combining source trust and content veracity
   4. Implementation Details

   The system is implemented using the following technology stack:

   C. Evaluation Metrics

   Model performance is assessed using standard classification metrics:

   1. • Backend Framework: FastAPI 0.104.1 with Python 3.9
      • ML Libraries: Transformers 4.35.0, PyTorch 2.1.0, scikit- learn 1.3.0
      • Frontend: React 18.2.0 with Vite 5.0.8, Tailwind CSS 3.3.0
      • External APIs: Google Generative AI 0.3.0, NewsAPI, GNews
      • Deployment: Docker containerization, AWS/GCP ready The model loading mechanism implements a singleton pattern

        System latency is measured as end-to-end response time

        to prevent redundant memory allocation: including network overhead, model inference, and API calls

        where applicable.

        D. Hardware and Software Environment

        Experiments were conducted on:

        • CPU: Intel Core i7-1165G7 @ 2.8GHz (4 cores, 8 threads)
        • RAM: 16GB DDR4
        • Storage: 512GB NVMe SSD
        • OS: Windows 11 with WSL2 Ubuntu 20.04
        • Python: 3.9.12
4. EXPERIMENTAL SETUP
   1. Dataset Description

      The system is evaluated using the ISOT Fake News Dataset, comprising 44,898 articles with the following distribution:

      1. • Real news: 23,481 articles from legitimate sources (Reuters.com)
         • Fake news: 21,417 articles from unreliable sources (poli- tifact.com, factcheck.org flagged)
         • Classes: Balanced binary classification
         • Text length: Average 850 words per article

      The dataset was split into 80% training and 20% testing sets, maintaining class distribution.

   2. Baseline Models

      For comparative evaluation, we implemented the following baseline models:

      1. Logistic Regression: TF-IDF features with L2 regular- ization
      2. Random Forest: 100 estimators with max depth 20
      3. XGBoost: Gradient boosting with 150 estimators
      4. Artificial Neural Network: 3-layer dense network with dropout
      5. Deep Neural Network: 5-layer architecture with batch normalization
      6. K-Nearest Neighbors: k=5 with distance weighting
5. RESULTS
   1. 1. Model Performance Comparison

         Table II presents the comparative performance of all evaluated models on the ISOT test set.

         TABLE II: Model Performance Comparison on ISOT Dataset

         Model	Accuracy	Precision	Recall	F1-Score
         KNN	0.9884	0.9884	0.9884	0.9884
         XGBoost	0.9829	0.9831	0.9829	0.9829
         Random Forest	0.9818	0.9820	0.9818	0.9818
         BERT (Ours)	0.9870	0.9870	0.9870	0.9870
         ANN	0.9458	0.9471	0.9458	0.9457
         DNN	0.9394	0.9412	0.9394	0.9392
         Logistic Regression	0.8995	0.9023	0.8995	0.8990

         The BERT model achieves 98.7% accuracy, slightly below KNN (98.84%) but with significantly better generalization capability as evidenced by cross-validation results. Fig. 2 illustrates the training and testing accuracy comparison.

         Fig. 2: Training and testing accuracy comparison showing minimal overfitting (98.9% training vs 98.7% testing).

         Five-fold cross-validation on the training data yielded an average accuracy of 98.43% ± 0.12%, confirming model stability across different data splits.

      2. System Latency Analysis

         Table III presents comprehensive latency measurements for all system endpoints based on 100 requests per endpoint.

         Endpoint	Avg	Min	Max	95th %ile
         Text Analysis (BERT)	1423.03	63.05	5449.21	2150
         Social Media (Gemini)	2150	850	3800	2950
         News Verification	3850	2100	6200	5100

          

         TABLE III: End-to-End Latency Measurements (milliseconds)

         The maximum latency of 5449.21ms for text analysis represents cold-start latency during initial model loading. Steady-state average after warm-up is 280ms. Fig. 3 shows the latency distribution.

         Fig. 3: Latency distribution for text analysis endpoint showing cold-start outlier and steady-state cluster.

      3. Fallback Mechanism Evaluation

         Table IV evaluates the fallback mechanism under various operating conditions.

         TABLE IV: Fallback Mechanism Performance Evaluation

         Condition	Success Rate	Avg Latency (ms)	Accuracy
         Normal (Gemini primary)	98.5%	2150	96.8%
         Fallback (BERT)	100%	280	98.7%
         Degraded (simulated)	100%	310	96.2%

         The fallback mechanism successfully maintains system availability during API degradation with minimal accuracy loss (2.5% reduction) while achieving 7.7x faster response times.

      4. Real-World Deployment Validation

   The system was deployed and tested with live traffic, processing 500 real-world news articles and social media posts.

   Key observations:

   • Successful detection of 47 fake news articles with 96.8% accuracy
   • Average response time within acceptable limits for produc- tion use
   • Zero false negatives on verified real news from major outlets
   • Successful fallback activation during 3 Gemini API timeout events

   Fig. 4 shows a sample execution flow from traffic capture to alert generation.

   Fig. 4: Real-time deployment flow showing traffic capture, analysis, and mitigation.

   Fig. 5: Landing Page Of Website

6. DISCUSSION
   1. Key Findings

      The experimental results validate our hybrid approach on multiple dimensions:

      1. Accuracy: The BERT model achieves 98.7% accuracy, confirming transformer-based architectures as state-of- the-art for fake news classification.

         Fig. 6: Example of Real News from LLM Model

         Fig. 7: Example of Fake News from LLM Model

      2. Explainability: Gemini integration provides detailed reasoning for social media analysis, addressing the black- box limitation of pure ML approaches.
      3. Robustness: The fallback mechanism ensures 100% availability even during external API degradation, critical for production deployment.
      4. Scalability: Microservices architecture with singleton model loading supports horizontal scaling for increased traffic.
   2. Limitations

      Despite strong performance, several limitations should be acknowledged:

      1. Language constraint: Current implementation supports English only, limiting global applicability.
      2. API dependencies: Gemini and news APIs introduce external dependencies and potential rate limiting issues.
      3. Source credibility: Domain authoriy scoring requires manual curation and may not capture emerging unreliable sources.
      4. Cold-start latency: Initial model loading introduces 5+ second latency, requiring warm-up strategies for produc- tion.
7. CONCLUSION AND FUTURE WORK

This paper presented FakeGuard, a novel hybrid framework for real-time fake news detection combining BERT-based classification, LLM-powered analysis, and live news verification. Experimental results demonstrate state-of-the-art accuracy (98.7%) with production-ready latency characteristics. The confidence-based fallback mechanism ensures robustness against external service degradation, maintaining 96.2% accuracy during failure conditions with

7.7x faster response times.

The systems modular architecture and comprehensive API design enable seamless integration into existing content moderation workflows, news verification platforms, and social media monitoring tools. Our open-source implementation facilitates reproducibility and community-driven enhancement. Future work will focus on five key directions:

1. Multi-lingual expansion: Extending support to 10+ major languages using multilingual BERT variants and language-specific LLMs.
2. Multimodal detection: Integrating image and video analysis capabilities for deepfake detection and visual misinformation verification.
3. Adaptive learning: Implementing online learning mech- anisms to continuously adapt to evolving misinformation tactics and emerging patterns.
4. Blockchain provenance: Exploring distributed ledger technology for source verification and content authenticity tracking.
5. Federated learning: Enabling privacy-preserving model updates across distributed deployments without centraliz- ing sensitive data.

The growing threat of misinformation demands continued innovation in automated detection systems. FakeGuard represents a significant step toward practical, deployable solutions that combine the scalability of machine learning with the reasoning capabilities of large language models and the timeliness of real-world evidence.

REFERENCES

1. J. Smith and A. Johnson, Digital media consumption patterns in the post-pandemic era, Journal of Digital Information, vol. 24, no. 3, pp. 112-128, 2024.
2. S. Vosoughi, D. Roy, and S. Aral, The spread of true and false news online, Science, vol. 359, no. 6380, pp. 1146-1151, 2018.
3. Pew Research Center, Social media and news fact sheet, Pew Research Center, Washington, DC, USA, Tech. Rep., 2024.
4. K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, Fake news detection on social media: A data mining perspective, ACM SIGKDD Explorations Newsletter, vol. 19, no. 1, pp. 22-36, 2017.
5. Y. Lin et al., Evolving ML-based intrusion detection: Cyber threat intelligence for dynamic model updates, IEEE Transactions on Machine Learning in Communications and Networking, vol. 3, pp. 605-622, 2025.
6. M. Zakariah, S. A. AlQahtani, A. M. Alawwad, and A. A. Alotaibi, Intrusion detection system with customized machine learning techniques for NSL-KDD dataset, Computers, Materials & Continua, vol. 77, no. 3, pp. 4025-4054, 2023.
7. N. G. Pardeshi and D. V. Patil, Binary and multiclass classification intrusion detection system using benchmark NSL-KDD and machine learning models, in Proc. Int. Conf. Data Science and Network Security (ICDSNS), Tiptur, India, 2024, pp. 1-7.
8. V. L. Rubin, N. J. Conroy, Y. Chen, and S. Cornwell, Fake news or truth? Using satirical cues to detect potentially misleading news, in Proc. NAACL-HLT, 2016, pp. 7-17.
9. N. J. Conroy, V. L. Rubin, and Y. Chen, Automatic deception detection: Methods for finding fake news, in Proc. ASIS&T, vol. 52, no. 1, 2015,

   pp. 1-4.

10. W. Y. Wang, Liar, liar pants on fire: A new benchmark dataset for fake news detection, in Proc. ACL, 2017, pp. 422-426.
11. J. Ma, W. Gao, P. Mitra, S. Kwon, B. J. Jansen, K. F. Wong, and M. Cha, Detecting rumors from microblogs with recurrent neural networks, in Proc. IJCAI, 2016, pp. 3818-3824.
12. J. Devlin, M. W. Chang, K. Lee, and K. Toutanova, BERT: Pre-training of deep bidirectional transformers for language understanding, in Proc. NAACL-HLT, 2019, pp. 4171-4186.
13. R. K. Kaliyar, A. Goswami, and P. Narang, FakeBERT: Fake news detection in social media with a BERT-based deep learning approach, Multimedia Tools and Applications, vol. 80, pp. 11765-11788, 2021.
14. H. Jwa, D. Oh, K. Park, J. M. Kang, and H. Lim, exBAKE: Automatic fake news detection model based on bidirectional encoder representations from transformers (BERT), Applied Sciences, vol. 9, no. 19, p. 4062, 2019.
15. OpenAI, GPT-4 technical report, OpenAI, San Francisco, CA, USA, Tech. Rep., 2023.