Deep Fake Survey

Building the neural network

Inspiration

Deepfakes have generated a lot of traction as of late. There are many videos out there of celebrities having their face swapped or popular apps out that that age you. Despite the cool use cases of deepfakes, it poses are a large threat to people's well-being and personal dignity. A lot of people are unaware of the issue and most assume that the crude face swaps from previous years are what we have today.

Purpose of our Project

Our app web app not only aims to collect data on how well people can identify fakes vs reals, but it also brings into perspective how real some of the images can be.

What we used

We built it using a flask server that feeds information to the user. We collected the images from kaggle and used javascript to control the buttons to either change to the next image or upload the user's choices to our server. Within the server, we have a MongoDB server connected through atlas that stores information on each image such as how many people voted on real or fake.

Challenges

We had intended to use AWS or Google Cloud computing to build more complicated models but when the time came to set up the servers, we were either restricted by Amazon's CPU limit or Google's GPU limit. Also, a lot of the technology we were using was brand new to us. One a few of us had an experience working with any of these so it was a completely new learning experience.

What we are proud of

We are proud that we were able to generate a model and push it into production. We also think that our solutions for sending and receiving data, though it took a long time to set up, were very good.

What we learned

We learned that there is a difference between what a human sees and what a computer sees. It's sometimes easy for humans to detect deepfakes because we notice the large glaring signs but other times, we can't see the small pixel differences like the ai can. We also learned how to host a website and connect it with not only HTML but also MongoDB and have communication between the different technologies.

What's next for Deep Fake Survey

We would like to spend some more time working on the UI. It can be polished a lot more to be more pleasant. Next, we think that implementations of different algorithms and the usage of AWS or Google Cloud will help us improve our accuracy of the model. There are lots of papers out there and we spoke with a mentor about implementing the various different algorithms out there.