SYNOPSIS
TrustGuard AI
Real-Time Content Authenticity Detection System
Guide Name: Dr. A. Mallikarjuna Reddy
Designation: Head, Dept. of AI
Team No: 02
Team Members:
Pantangi Divya Anila-22EG107A43
Pinnaboyina Harshitha-23EG507A03
P. Sir Arthur Cotton-22EG107A50
Kandadi Praneeth Reddy-22EG107A24
Contents:
1. Abstract
2. Introduction
3. Literature Survey
4. Objectives
5. Problem Statement
6. Methodology
7. Reference
Abstract
In today’s digital world, fake text, images, and videos are increasingly difficult
to identify, posing a major risk to public trust. To address this challenge, we
developed TrustGuard AI, a real-time content authenticity detection system. The
framework employs DistilBERT for text analysis and MobileNetV2 for image
classification, categorizing content as Legitimate, Suspicious, or Fake. To
enhance transparency, we integrated Explainable AI techniques: SHAP
highlights key words in text influencing predictions, and Grad-CAM visualizes
critical regions in images. By combining accuracy with interpretability,
TrustGuard AI not only detects manipulated content but also explains its
reasoning. In future, the system will extend to audio and video detection and be
deployed as a browser extension for seamless real-time use.
Introduction:
With the rise of AI tools capable of generating fake text, images, and videos,
online misinformation has become increasingly difficult to detect, creating a
serious threat to public trust. TrustGuard AI addresses this challenge by serving
as a real-time authenticity detection system that classifies digital content into
three categories:Legitimate, Suspicious, or Fake,while providing an instant trust
score. The system combines efficiency and accuracy using DistilBERT for text
analysis and MobileNetV2 for image classification, both lightweight yet
powerful deep learning models. For transparency, SHAP highlights influential
words in text predictions, while Grad-CAM visualizes critical regions in
images, making decisions explainable. In the future, TrustGuard AI will extend
to audio and video detection and be deployed as a browser extension for
seamless real-time verification.
Literature Survey
1. Explainable Fake News Detection based on BERT and SHAP applied to
COVID-19 (Men & Mariano, 2024)
This study introduced an explainable framework for fake news detection using
BERT for classification and SHAP for interpretability. Applied to COVID-19
misinformation, the model not only achieved reliable accuracy but also
enhanced transparency by highlighting influential words driving predictions,
thereby increasing user trust.
2. Fake News Detection via NLP: A Survey of Deep Learning Techniques
(Zhou, Han & Morstatter, 2024)
This survey paper reviewed deep learning methods for fake news detection
using natural language processing. It analyzed various models such as CNNs,
RNNs, and transformers, identifying their advantages, limitations, and
application scenarios. The work provided a strong foundation for understanding
the strengths of NLP-based approaches while pointing out challenges like data
imbalance and generalization.
3. Exposing Deepfake Videos by Detecting Face Warping Artifacts (Li,
Chang & Lyu, 2025)
This research focused on detecting manipulated videos by identifying face
warping artifacts, a common flaw in deepfake generation. Using forensic
techniques and deep learning, the study achieved high accuracy in
distinguishing authentic from fake content, particularly in video frames. This
work is significant for advancing multimedia authenticity verification.
Problem Statement
With the rapid growth of AI-generated and digitally manipulated content, it has
become increasingly difficult for users to distinguish between genuine and fake
information online. Fake text, images, and multimedia spread misinformation,
erode public trust, and pose potential security threats. Current detection methods
often lack accuracy, transparency, and accessibility, leaving everyday users
without reliable tools to verify content credibility in real time.
Objectives
● To design and develop a deep learning model(DistilBERT) for detecting
fake and legitimate content.
● To design and develop an image classification model(MobileNetV2) for
identifying fake and manipulated images.
● To integrate explainable AI methods(SHAP for text and Grad-CAM for
images) into models for transparent decision-making.
Methodology: The methodology consists of two pipelines:
[Link] Pipeline Processing (Text + Image):
Text data is preprocessed and analyzed using DistilBERT, while images are
preprocessed and passed through MobileNetV2 for feature extraction.
2. Classification and Trust Scoring:
Both pipelines classify the input into Legitimate, Suspicious, or Fake, and a
Trust Score is generated based on model confidence.
3. Explainability Integration:
To ensure transparency, SHAP highlights influential words in text, and
Grad-CAM generates heatmaps for images, which are displayed through a
Streamlit interface for user interpretation.
Both pipelines converge into a Trust Score Module, categorizing inputs
into Legitimate, Suspicious, or Fake. Finally, results are displayed to the
user via a Streamlit App.
Future Work Plan
Week 6: Explainable AI methods are integrated into the system. SHAP is
applied to the text model to highlight the most influential words, while
Grad-CAM is added to the image model to visualize key regions. A Streamlit
web interface is developed to let users upload text or images and get predictions
instantly. The interface also displays the trust score along with explanations,
making the results clear and transparent.
Week 7: Focuses on final testing, documentation, and presentation. The full
system is validated with multiple text and image samples to check prediction
accuracy and explanation consistency. Any bugs or integration issues are fixed,
and the complete project report is prepared with results and visuals. Presentation
slides are finalized, and a live demo of the Streamlit app is conducted to
showcase real-time detection before submission.
References
[1] X. Men and V. Y. Mariano, “Explainable fake news detection based on
BERT and SHAP applied to COVID-19,” International Journal of Modern
Education and Computer Science (IJMECS), vol. 16, no. 1, pp. 11–19, Jan.
2024, doi:10.5815/ijmecs.2024.01.02.
[2] P. Zhou, J. Han, and D. Morstatter, “Fake news detection via NLP: A survey
of deep learning techniques,” IEEE Transactions on Neural Networks and
Learning Systems, vol. 35, no. 3, pp. 456–472, Mar. 2024,
doi:10.1109/TNNLS.2023.3258742.
[3] Y. Li, M. Chang, and S. Lyu, “Exposing deepfake videos by detecting face
warping artifacts,” IEEE Transactions on Information Forensics and Security,
vol. 19, no. 4, pp. 2110–2122, Apr. 2025, doi:10.1109/TIFS.2025.3338124.
