Summer Games Interns Use AI to Spot Fake Pics

While it’s hard to imagine convincingly faked cat photos causing much harm, AI-generated human portraits could easily be used to conduct online espionage and spread misinformation. Enough to make you a little anxious, right? Raj Shah and Mike Ainsworth share your concern, and they’re taking action to improve the situation.

As Booz Allen Summer Games interns, they spent 10 weeks working on a project that they hope will make it much harder for would-be spies and propagandists to fool unsuspecting web surfers using Generative Adversarial Networks (GANs). Generative Adversarial Networks (GANs) are an (AI) technique that can generate photorealistic images of things that are not actually real—fictitious cars, cats, and even people who have never existed outside of a hard drive.

 “We looked at how the American citizen is threatened by GANs, and what we could create to help mitigate that in an accurate, user-friendly way,” Raj says.

The project, which uses AI to let users check whether digital images they encounter on social media are AI-generated or authentic, was Raj and Mike’s team’s entry into Booz Allen’s 2019 Summer Games. The Summer Games Challenge Cup, an annual competition, brought together close to 450 interns this year. Divided into teams, they competed to see who could develop the most compelling original product or capability in 10 weeks. 

For Raj, a rising senior studying computer science at the University of Florida, that team-based structure was a really important part of the experience. “At previous internships, I had a boss, they gave me tasks, I did them. Here, I was teamed up with other interns, and we took full responsibility for an entire project, from concept to prototype,” he says.

Being responsible for every aspect of the project meant that Raj, Mike, and their teammates had to stretch their skills and develop new ones. If there was something they were unsure about, they had to figure it out or ask for help.

Mike, a rising senior studying biomedical engineering at Johns Hopkins University, was surprised by how easy it was to find help when they needed it. “With other internships I’ve had, interacting with higher-ups was difficult,” he says. “At Booz Allen, whenever we ran into an issue, we could ask around the office. More often than not, someone who had dealt with something similar could help us right away.”

In the end, the project was a huge success. Titled “FakeBook: Leveraging Machine Learning to Defeat AI-Generated Social Media Manipulation,” it earned Raj, Mike, and their teammates—Fatima Mohamedali, Isaiah Gregory, and Sofia Sikander—a place among this year’s Summer Games finalists.

Like all Summer Games projects, “Fakebook” was developed to meet a need identified by the Booz Allen experts who serve the firm’s many diverse clients. Raj, Mike, and team hope that the work they did this summer will serve as a foundation for further development. “We think the ability to identify GANs images is an important first step toward tackling the bigger problem,” Mike says.

“You could take what we started and add the ability to uncover deep fake videos or AI-generated text—the possibilities are pretty endless as far as how this architecture can be used in different ways,” Raj adds.