DEEP’t is a team comprised of five Ewha students majoring in Cyber Security, bonded by their interest in digital forensics and artificial intelligence. For their graduation project in 2020, DEEP’t developed an automatic detection system of deep fake, which is a deep-learning technology that creates fake videos manipulated by computer graphics, without detailed human intervention.
Their cutting-edge system won the grand prize in both the “2020 Startup Storytelling Competition” hosted by the Korean International Trade Association (KITA), and the “SW Startup Competition” co-hosted by the Ewha Center for Entrepreneurship and Software-Oriented UniversityProject. They also appeared on the TV show “Unanswered Questions.”
The team members, Kang Mi- hyun, Kim Yeon-hee, Baek Hyun- jung, Yoo Ye-lyn, and Yoo Hyun-sun, shared with Ewha Voice the process of developing the deep fake detection system.
How the detection model by DEEP’t is constructed
The team experimented with various model structures. Yoo Ye-lyn explained that the model they used is a so-called supervised model. A supervised model refers to a model that is trained with a set of input and correctly classified answers.
Not only does the model detect whether the video is fake, it defends attacks. Baek Hyun-jung explained that by attacks, it means videos or images with slight alterations that confuse the deep fake detection model.
DEEP’t was able to develope a model that bypasses altered image attacks by adding in more training processes with different datasets.
How the deep fake detection module was developed
The members explained that it was not easy deciding the scope of their graduation project.
“Researching and developing the deep fake detection module required a lot of academic knowledge,” Yoo Hyun-sun said. “It was hard to predict the level of difficulty of this project, especially because the field was not being actively researched at the time we challenged it.”
Yoo stated that the team set as many intermediary goals as possible in order to break down the project into manageable steps. Then they quickly moved onto the check-up stage, validating each goal.
Once the framework of the project was in place, each member was assigned different roles, such as web development, deep-learning module development, and tuning the training process for module attack and defense.
Yoo was in charge of web development. At first, she felt rather lost because it was her first time working in this field.
“Before implementing detailed functions, I needed to check whether the overall technical implementation was even possible,” Yoo said. “Specifically, I remember feeling stuck when trying to implement video functions, because there were not many reference materials. When I succeeded in connecting all the functions and realized the web page development was possible, I felt overjoyed.”
DEEP’t hoped the module would be used to eradicate digital crimes such as deep fake crimes. They believe the module should be approachable by the public for it to develop at a faster pace, and hope companies and organizations will utilize the module. They look forward to it becoming a popular program available for civilian use in the future.
Appearing on “Unanswered Questions”
DEEP’t shared their deep fake detection system on the episode “A war with the fake - deep fake” of a famous TV show named “Unanswered Questions” on Feb. 27. On the show, they were featured as a student team who developed a deep fake detection model.
Yoo Hyun-sun predicted that the “Unanswered Questions” production team contacted DEEP’t after seeing the article of them winning the grand prize of “2020 Startup Storytelling Competition.”
In the episode, the writers masqueraded three actors with deep fake videos made by DEEP’t. Facial features were extracted from different subjects’ photographs, to videos of a person in a hospital room with a deep learning model. These videos were sent to the subjects’ parents, stating that the subjects were involved in an accident and needed money.
Although parents showed doubt that the videos were truly that of their child, the fake videos were natural enough for them to send the card numbers.
Project preparation and future aspirations
Before gathering as a team, members of DEEP’t had individually studied deep-learning in different ways.
Yoo Hyun-sun and Baek Hyun- jeong first found their interest via “Machine Learning and Artificial Intelligence” class provided at Ewha. Yoo stated she was thankful of her professor who advised her to accompany a dissertation study group. Kang Mi-hyun, who worked as an intern at Ewha’s Image Processing Lab, read through thesis papers by herself. Kim Yeon-hee too researched over diverse papers and topics.
Two of the team members graduated this March. Kim entered a graduate program at Ewha on Software Science, while Kang, also graduated, told she attempts to proceed developing her career as a software engineer.
Baek began to work as a second lieutenant. She explained that she hopes to continue developing software programs while undergoing her responsibilities.
Yoo Ye-lyn and Yoo Hyun-sun both showed their willingness to continue studying about deep fake. Whilst the former is currently an exchange student over at U.S., the latter is reviewing the knowledge utilized in her project and planning out her career path.
Lastly, the members of DEEP’t delivered messages for students looking forward to studying Cyber Security at Ewha. The common message was to try out as many diverse activities as possible in order to figure out about oneself, and experience various cyber security studies as well which will help narrow down to a specific field later.
“Facing graduation, I feel like our school supports students in multifaceted ways for them to develop their career.” Yoo Hyun-sun said. “ I hope the new-coming students try all they might within the protection of the school and experience an unregretful college life.”