Inspiration

In today's world, with deepfakes and manipulated media becoming more convincing, it’s harder to know what’s real online. While AI can be incredibly precise at spotting fakes, accuracy alone isn’t enough if users don’t trust the results. The key is explainability—if people can’t see for themselves why something is flagged as fake, they won’t believe it. That’s why we designed our project not only to detect fake media but also to highlight exactly what’s suspicious, helping users verify the AI's findings on their own. This way, we’re building trust through both precision and transparency.

What it does

Our project isn’t just about telling you if something’s real or fake—it also shows you why. We built a Telegram bot with a straightforward interface where you can upload images, videos, or audio files. The bot runs an analysis and gives you a probability score of how likely it is to be fake. It also highlights specific areas that look suspicious, whether it’s something off in an image, weird movements in a video, or strange audio glitches.

How we built it

We took state-of-the-art (SOTA) models for detecting fakes in images and audio, then added explainability using Grad-CAM to highlight the most suspicious areas. We treated videos as a series of images, applying the same approach to each frame.

Challenges we ran into

Embed an explainable model to the open-source SOTA models.

Accomplishments that we're proud of

Implementing AI explainability in tg-bot format.

What we learned

How to deal with fakes in 3 modalities!

What's next for (ノ ゜Д゜)ノ ︵ ┻━┻

Implement more advanced method for video fake detection

Built With

Share this project:

Updates