AI Deepfake Detector Extension

Inspiration

My mother sent me a post that was AI generated, she fully believed it and it was difficult to convince her of the truth, AI has gotten so intelligent that it is fooling the older and non technical folks of our society at an unprecedented level. Furthermore, there are nefarious webpages dedicated to deepfakes of real women in explicit poses, without their consent.

What it does

This is a google extension that allows a person to manually detect if an image, video, audio, or text is AI generated. The magnum opus is that it is able to run in real time and remove altered and nefarious AI posts from your YouTube, Twitter (X), Instagram, Tiktok etc. feeds, removing the worry that something is altered and protecting people like my mother who could also get potentially misinformed, scammed, or worse.

How we built it

We (I) built it using multiple LLM's for guidance and ran huggingface models locally that have 100+M parameters alongside LLM's (Llama via Groq) dedicated to detecting deepfake faces, deepfake audio, deepfake texts, and deepfake videos. We used fastapi as our backend along with Javascript and google web extensions to build the extension.

Challenges we ran into

It was very difficult to find and optimize a model so that real time detection does not take minutes to clear a feed, along with removing youtube posts, twitter posts, instagram comments, etc. It was also difficult to get fast llm readings (until groq!). Furthermore, using a CPU only build without CUDA made processing difficult.

Accomplishments that we're proud of

We are proud that we were able to accomplish real time detection and were able to scan and remove posts quickly for the user.

What we learned

Learned a lot about deepfakes, a lot of huggingface and their models, along with trying Groq for the first time. This was also the first time making an extension.

What's next for AI Deepfake Detector Extension

Next up is optimizing and going beyond an extension to hopefully detect explicit material of unconsented individuals and reporting it to authorities, we also want to improve on detection and confidence with finetuning (veo 3 is difficult to detect.) Along with utilizing our detection technology to prevent scammers and impersonators. But before all of that we need to deploy a server. We also want to improve our UI as it is very barebones and basically auto generated, we spent a lot of our time on the backend.

Built With

• fastapi
• groq
• huggingface
• javascript
• llama
• manifest
• opencv
• pillow
• python
• pytorch
• scikit-learn
• uvicorn