Department of Electronics &
Communication
Visvesvaraya Technological
University
Freq-Fusion Ideathon Idea Presentation
Theme:
Team Cryptex
TrustAI
Title:”SafeLe
ns”
[Link] Kallumanti(4PA23IC036)
[Link] K M(4PA23IC047)
[Link] Mahaloof(4PA24IC402)
[Link] S M(4PA23IC048)
INTRODUCTION
• The digital landscape is being flooded with hyper-realistic synthetic media, known as deepfakes ,AI-generated
videos, audio, and images.
• This technology poses an unprecedented threat to individual privacy, national security, financial systems, and
democratic processes.
• Current detection methods are often single mode (analyzing only video or only audio) and struggle against
rapidly evolving generative AI.
• There is a critical need for a robust, proactive, and explainable detection system.
• Our project, Safelens, addresses this by developing a multi-modal framework that analyzes visual, auditory,
and temporal inconsistencies.
• It is designed to be integrated as an API into social media platforms, video conferencing tools, and news
verification services.
• The ultimate goal is to restore and ensure digital trust by providing a shield against malicious AI-generated
content.
PROBLEM
STATEMENT
• What problem we are trying to solve?
• We are solving the problem of identifying and flagging AI-generated deepfake media that is used to spread
misinformation, commit fraud, and create non-consensual content.
• AI generated images and videos are now very easy to make which causes people to
misuse it.
• Why is it important? (relevance, real-world impact)
⚬ Relevance: The accessibility of AI generative models has made creating convincing
deepfakes easier than ever.
⚬ Real-World Impact:
￭ Political: Can be used to manipulate public opinion and elections.
￭ Financial: Enables sophisticated identity theft and fraud through fake video KYC.
￭ Social: Harms individuals through reputational damage and cyberbullying via
synthetic imagery.
￭ Trust: Erodes the public's trust in digital media and institutions.
METHODOLOG
Y
Our Solution
• Safelens is a detection system that uses separate AI models to analyze a video's visual features,
its audio track, and how well they sync together, combining all three results to accurately identify
deepfakes.
• How it addresses the problem
• It addresses the limitations of single-source detectors by using a multi-layered, fusion-based
approach. This makes it significantly harder to fool, as a forgery would need to be perfect across
visual, audio, and temporal domains simultaneously.
Enhanced Proactive Explainable Scalable
Accuracy Defense Results Integration
Multi-modal analysis Designed to evolve with Provides confidence scores
Easy API integration for
generative AI, offering future- and specific reasons for
drastically reduces false various digital platforms.
proof protection. detection.
positives and negatives.
Working
How TrustAI Works
• Break It Down: The system takes a video and separates it into moving pictures
and sound.
• Check the Video: AI analyzes the pictures for glitches real people don't have,
like weird blinking or strange skin.
• Check the Audio: AI analyzes the sound for robotic or unnatural voice
patterns.
• Check the Sync: AI checks if the person's lip movements perfectly match the
words being spoken.
• Combine the Evidence: The results from all three checks are combined to
make a single, super-accurate decision.
• Get the Result: The system gives a final answer: "Authentic" or "Deepfake
Detected."
REFERENCES
• D. Güera and E. J. Delp, "Deepfake Video Detection Using Recurrent Neural Networks..."
• Link: [Link]
• S. Agarwal, H. Farid, et al., "Protecting World Leaders Against Deep Fakes..."
• Link:
[Link]
Deep_Fakes_CVPRW_2019_paper.html
• A. Rössler, D. Cozzolino, et al., "FaceForensics++: Learning to Detect Manipulated Facial Images..."
• Link: [Link]
• Also check the project page: [Link]
• H. Khalid, S. Tariq, et al., "FakeAVCeleb: A Novel Audio-Video Multimodal Deepfake Dataset..."
• Link: [Link]
• This is crucial for your multi-modal approach as it provides the training data for audio-visual deepfakes.
• Y. M. Said, "Deepfake Detection using Deep Learning Methods: A Systematic Review..."
• Link: [Link]
Thank
You
