Facebook and Microsoft are coming together with academicians and Partner members of the AI Coalition to develop technology to detect deepfakes. The two tech giants have launched a ‘Deepfake Detection Challenge’ aimed at promoting detection research.
Facebook will also appoint researchers to create more realistic deepfake videos to test detection tools. The social media giant has assured users that the company will not utilize their data for its research. Deepfake videos which will be released in December will only feature paid actors.
A threat for upcoming US Presidential Elections
Social media companies are currently faced with the challenge to tackle the threat of deepfakes in the U.S. presidential elections to be held in November next year. The upcoming US Presidential elections are expected to witness AI-based hyper-realistic videos involving both common people and political leaders. (In the hyper-realistic videos, people can be seen saying things that they haven’t said.)
Democratic National Committee had earlier shown a deepfake video of its own Chairman Tom Perez in August. The audience was made to believe that the real Tom had Skyped into the conference. Another deepfake video of House Speaker Nancy Pelosi was also manipulated to demonstrate the possible havoc that these videos can create. The original video was manually slowed down to make it look like her speech was slurred.
Initiatives to Tackle the Threat
Facebook has partnered with researchers from several universities including MIT, UC Berkeley, University of Oxford, and Cornell University to tackle this challenge. One of these teams under the UC Berkeley Professor Hany Farid is reportedly building soft biometric models which will map real politicians’ facial quirks to detect if a new video is fake. As pers the reports, researchers are also planning to develop systems that can authenticate visuals using digital watermarks.
But despite numerous efforts to combat the situation, deep fake technology has also evolved consistently in recent months. It has almost become a race between people producing these deepfakes and the ones trying to detect them. Experts have also raised concerns over the increasing accessibility of this technology to less-skilled creators.
Online creators of deepfake content have also begun to explore the market via. offering easy-to-make deepfakes. Machine-learning enthusiasts from several countries including Japan and Poland are offering step-by-step tutorials on YouTube for AI-based impersonations. Some of them have been charging as much as $30 for 50 words of an AI-based Donald Trump’s voice impersonations.
Deepfake Detection Challenge Details
Deepfake Detection Challenge will begin by hiring a group of paid actors to create a dataset of videos. All the co-sponsors including Facebook and Microsoft of the contest will tamper these videos to generate deepfakes using artificial intelligence. Participants will then be required to create programs to detect the deepfakes in the dataset.
The contest is open for all and it is expected to begin next month. However, participants will need to go through a screening process and agree to certain terms on using the video data. Facebook and Microsoft will collaborate with academia to develop a testing mechanism to score the effectiveness of detection programs submitted by the participants.