In today’s world, artificial intelligence can be used to create what’s called “deep fakes.”
YR Contributor Jayda Buckley interviewed Salma Mayorquin, an AI scientist and solutions architect based in San Francisco, California. Mayorquin used her machine learning skills to create a program that can spot deep fakes on the internet.
Mayorquin told us how she built her project and gave us some advice for breaking into the field of AI.
This interview has been edited for clarity and length.
Jayda Buckley: You have a degree in applied mathematics. How does that degree apply to your work detecting deep fakes?
Salma Mayorquin: Essentially, the technology behind some of that work is based on a bunch of math, so lots of calculus for our model that we train to be able to detect whether an image was a deepfake depends on being able to capture lots of information.
There’s lots of math involved in being able to capture that signal and differentiate between what is a fake image and what is a real image.
JB: What are the pros and cons of deep fakes?
SM: Deep fakes can be used in a good way. For example, there is one group that is trying to help, by using it for video streaming. Take Zoom, for example… you have a bad connection and you can't use your camera. Some groups are trying to incorporate deep fake technology to help still capture video, even though you might be having issues.
But deep fakes can also be dangerous. Maybe some people might try to make something that isn't true. And since they're very realistic looking, that can be dangerous. So like any technology; there's good ways to use it and there's bad ways to use it.
JB: How does the process of detecting deep fakes work?
SM: In the example that we have as a project, we try to take lots of images. So like when you have a camera, you know that your camera shoots a bunch of images to make a video. We take a little sample of time of a bunch of those images. And we use a model to extract the signal out of those images. So we kind of try to analyze what the face looks like. And after we have that numerical representation of what the face looks like to a computer, then we give it examples.
We give it a bunch of examples of what a real face looks like and then what a fake face looks like. Then we train our general model to be able to tell the difference.
JB: How long does it take to create a deep fake?
SM: So for a computer to understand how to create one of these can take maybe even a full day, 24, or 30 hours. Depends on what kind of computer you have. So the bigger computer you have, the faster the process goes, but it's still a lot of work.
JB: Your official job title is “Solutions Architect”. What does your day to day role look like as a solutions architect?
SM: My day to day job was actually something totally unexpected. I had no clue that this role existed until someone told me about it, and I gave it a shot. A solutions architect is essentially kind of like a software engineer as someone who knows about code, but they guide other engineers on how to build things, especially around machine learning or A.I.
J: What advice would you give to a young person interested in following your footsteps?
SM: I would say to try out lots of different things. Don't let any preconceived notion on what you're good at or what you are not good at stop you from trying something new.
When I first got into college, I didn't know what I wanted to study.
But after taking a bunch of classes and figuring out what I really liked, I just gave it a shot, and it turned out. Give yourself plenty of space and time to pick up anything new! You shouldn’t underestimate yourself!