this project is a milestone for me on my ML learning journey. I’ve designed a model to detect AI generated faces and developed a tool called DeeFace from it for other people to use: deefaces.com
Congratulations on achieving such an important milestone in your journey with machine learning! It is really amazing that you were able to make a model to find AI-generated faces and a tool like DeeFace to share your work with other people. This project sounds like it took a lot of work and time from you. How has the process gone for you?
Hi @amalbabu , thanks so much for your interest in checking out my tool!
I’m sorry to hear about your experience. But one thing about this tool tho, is that it works best when the image is square, and the face is completely in the cropped image (there’s a section about this on the page, but since a lot of users have also skipped it, I’ll try to make this section clearer in the future).
I’m currently working on enhancing the performance of this model, as well as an auto-cropper so that the user doesn’t need to crop the image manually every time. Hope you will also be there to test this tool again when all these updates are done!
This project took a lot of effort to build, but it was worth every bit of it. 1. I was able to build a model that applies in the real world, and learnt more about ML along the way, 2. talked to a lot of same-minded people and shared my project with them, 3. encountered “the void” when trying to build an audience for my tool when I was practically invisible on the internet, but things have started to work out!
This would actually help were so many ai generated fake videos are spreading, can your tool detected even that too, I mean can your tool detect if a fake actress face is used to create a fake video?
So far I’ve only designed my model to detect images, so technically we can take a screenshot of the face in the video, and then detect that screenshot. I’ve tried this with a video of Taylor Swift teaching math, and it seems to work fine.
I might try to expand this to detecting videos in the future though, would be a really cool thing to do
It took me almost a month to complete this website, but it was surprisingly the API hosting and web development that took most of that time (since I was inexperienced in this area, and getting that cropper functionality to work also took quite some time)
The ML part took me only a couple of days actually, since I’ve done many similar projects before, so I was about to come up with the effective data pipeline and the model architecture fairly quickly.
Currently, I’m learning Javascript and its frameworks to develop a better and more serious project, as well as learning multimodel search and RAG thanks to @Deepti_Prasad’s suggestion to hopefully achieve more with this model
I don’t know if you know this or not, something which you have created could be a great model and investor would look forward to it as there are some governments who are creating laws to prevent AI-generated fake videos. So keep working on it, get as much accuracy you could, then try on video, image and text generated data and showcase your model. Wish you all the best @CheeseCheddar