Global challenge created to accelerate the development of technologies that stop the spread of manipulated deepfakes and media
NEW YORK, June 17, 2020 – ID R&D, the award-winning biometric solutions provider offering AI-based voice, face and behavioral user authentication and anti-spoofing capabilities, proudly announces that multiple individual and team entries by ID R&D data scientists ranked in the top 10 of the Deepfake Detection Challenge (DFDC), claiming gold medal placements in the leaderboards. More than two thousand teams from around the world participated in the global challenge, which was sponsored by Facebook, Amazon Web Services (AWS), Microsoft, the Partnership on AI and leading academic institutions. Google’s Kaggle data science and machine learning platform hosted the challenge and leaderboard.
AI-generated deepfakes portray real people doing and saying fake things. These manipulated videos, audio files and images are being used for nefarious purposes from swaying public opinion to undermining elections to blackmail. Subsequently, the goal of the challenge was to accelerate the development of technologies to identify and stop the spread of manipulated media. Entrants in the DFDC gained access to a new data set of over 100,000 videos for use in creating new and improved detection models.
The results of the Challenge created controversy as some DFDC critics have pointed out the relatively low accuracy of the detection results, especially on the private dataset. ID R&D sees these results differently. Based on a comparison of the results achieved on various publicly accessible datasets and ID R&D’s own data, our analysis suggests that participants were not tasked with real-life scenarios but rather were asked to test the boundaries of what is possible: reduced FPS (frames per second), decreased size of the video, and other types of hard augmentation.
In other words, the Challenge was likely designed to test the ability of the technology under a highly difficult set of circumstances and with a limited amount of information available to participants. The fact that leading submissions managed to identify the majority of deepfakes correctly proves that even under unnecessarily difficult conditions, deepfake detection can already be useful in protecting the public from manipulated media. Moreover, ID R&D sees a much higher accuracy of deepfake detection in a real-world setting where the algorithm can analyze videos with normal FPS and size. ID R&D is working on productizing these capabilities.
ID R&D had multiple team members in the Top 10 of the Challenge leaderboard with team “ID R&D” finishing #8 and ID R&D’s Chief Scientific Officer and co-founder, Konstantin Simonchik, finishing #6 on the private leaderboard. Additionally, ID R&D Lead Researcher Anton Pimenov reached a Top 10 position on the public leaderboard. The individuals leveraged their expertise in machine learning and unique insights gained from their work on ID R&D’s passive facial liveness detection – a product that is used to detect presentation attacks on face recognition systems.
“ID R&D is a company that was founded on a deep commitment to research, development and innovation. We are immensely proud of our employees who participated in the challenge and congratulate them on three Top 10 finishes out of thousands of participants,” said Alexey Khitrov, President and co-founder of ID R&D. “Ensuring the legitimacy of the information presented by news outlets and shared online is critical, and we thank all the sponsors for their efforts to accelerate this important initiative.”
View the final Deepfake Detection Leaderboard.