How much would your favourite movie change if it starred another actor? A recent trend on social media involves editing clips of movies with actors who originally do not star in them. The most recent one I watched was of comedian/actor Jim Carrey in the horror film, The Shining. I was confused by what I saw at first, for I knew that he wasn’t originally in the movie, and yet, the actor’s facial movements and expressions seemed absolutely real, sans any obvious editing. It was only later that I found out that such videos are made using an Artificial Intelligence technology known as deepfake.
What is deepfake?
It’s a form of AI technology used to create videos or images of fake events. The process works by uploading numerous still images of one person, and video footage of another person. By morphing the former’s face with the latter’s, expressions and movements can be matched.
In a survey conducted by a cybersecurity platform, it was found that the number of deepfake videos are growing exponentially. The report suggests that such images/videos are doubling every six months, and they might not always be used for good. In September 2019, 15,000 deepfake videos were found online, and 96% of them had pornographic content.
In India, a video featuring a political party leader criticising another went viral on WhatsApp before the Delhi Assembly election in February. But after further speculation on the matter, the video was found to be made using deepfake software.
From amateur media enthusiasts to researchers, pornographic producers, and political parties, everyone is making content using deepfake technology. However, there are very few tools to identify such videos, and there are no strategies in place to prevent the circulation of such content.
Keeping in mind the rampant misinformation stemming from deepfakes, four students of Raisoni College of Engineering, Nagpur have come up with a solution that uses AI to identify manipulated videos, pictures, or audio.
“After months of research and development, we have developed AI and computational neural networks to detect deepfakes with 96% accuracy,” says Atharva Peshkar, a 3rd-year student of B.Tech Artificial Intelligence. Peskhar worked along with three other students on this project — Rishita Mishra (3rd year, Electronics and Telecommunication), Yash Moharir (3rd-year Computer Science) and Atharva Khedkar (2nd year, BTech AI) — and they named themselves ‘Team Detectd’.
Identifying small changes through a frame-by-frame analysis
In August 2020, the team began research on the kind of solutions that were already available. After speaking with a few cybersecurity experts, it was clear that there are advanced technologies to identify deepfake images, but not videos. The images are identified using softwares that pick up any distortion between the image’s background and foreground. This distortion can be anything; uneven shapes, lines, or facial features.
“So, we took it as a challenge to work on a model which can analyse any form of media (video, image or audio) and accurately predict whether they are fake. For this, we used the method of Spatial and Temporal Data analysis. This identifies small changes in a video by conducting a frame-by-frame analysis,” says Atharva Khedkar, adding that the method is considered to be accurate because it has a memory of previous frames.
After a few months of developing the machine learning model, the team tested a few real-life fake videos to test the accuracy. “It did not cross the 50% mark, because we could not access enough data sets (videos) to expand our machine learning model,” says Peshkar.
To perfect the system, Peshkar says he started reading research papers to understand other methods that are being used to process and analyse videos. “By using some existing methods with the new model we have innovated, we were able to process videos effectively. After we started seeing results, we started using data sets provided on a website named Kaggle, to analyse fake videos and images,” he says.
To date, the team has accurately processed over 7,000 videos with 96% accuracy. They are also partnering with Cyber Forensic Technologies (CFT), Nagpur to deploy their technology and help the organisation investigate cybercrimes. “But we are still developing the model, making changes, and updating the system to achieve 100% accuracy,” says Atharva Peshkar.
In January, Team Detectd enrolled in the Microsoft Imagine World Cup and submitted their research paper. They were selected as one of the four teams from India to participate in the World Cup. The team selected as the winner of the World Cup will also get a mentorship opportunity under Microsoft CEO Satya Nadella.
Edited by Divya Sethu