Accomplishments

ExposeAI: A Lightweight, Multimodal, and Explainable Framework for Efficient Deepfake Detection


  • Details
  • Share
Category
Conference
Authors
Abhijit Patil , Vedant Hundare & Zainuddin Fatakadawala
Conference Name
International Conference on Artificial Intelligence and Emerging Technology (AI Summit)
Conference From
19-Nov-2025
Conference To
21-Nov-2025
Conference Venue
Noida, India
  • Abstract

Deepfakes are a big concern because they make it impossible for people to believe what they see online. How can we know what's true anymore when anybody can make phony films that look authentic with just a smartphone app? That's why we created ExposeAI, a program that helps regular people find deepfakes without having to know a lot about computers. ExposeAI employs a cutting-edge AI technique called Vision Transformer (ViT-Base). You may think of it as a digital detective that is really adept at finding the little, almost imperceptible signs that something is phony. We trained our system with thousands of pictures from the Celeb-DF v1 dataset, and then we put it to the test with videos from the harder Celeb-DF v2 dataset. What were the results? It worked quite well, showing that it can handle both movies and pictures. But here's what makes ExposeAI different: it doesn't just tell you "this is fake" and leave you wondering why. Instead, it uses simple visual heatmaps to show you precisely what gave it away. These bright overlays show the parts of a picture or video that appear suspect. It might be anything strange about the eyes, the skin texture, or the lighting that people usually overlook. This openness isn't simply a good feature - it's vital. When individuals understand how the technology works and can see the proof for themselves, they're more inclined to trust and really utilize the tool. The creators built ExposeAI to put the power back in consumers' hands, giving them the confidence to question what they see online and make educated judgments regarding digital material validity.

© Somaiya 2026 / All rights reserved.
Get in Touch