Unsuccessful, but Triumphant — Facebooks Battle Versus Deep Fakes
Facebook’s competition to find a possible algorithm to detect masterful deep fakes has ended mediocrely but I don’t think that’s to any surprise. The unsolved problem of discovering deep fakes is a daunting task that even some of the best researchers haven’t found solutions to yet. Facebook founded a competition to accelerate deepfake detection development called the DeepFake Detection Challenge (DFDC) where anyone could participate to help an open, global collaboration get closer to a solution for the deepfake issue.
The best news from this collaborative challenge is Facebook’s research team published that “the top model achieved an average precision of 65.18%” which is nothing short of impressive considering these were from community fed solutions. Facebook had an inside solution but it failed during run-time which prevented them from being included on the leaderboards. In addition to running the event, Facebook also sought after raw material to be used during the competition, hiring about 3000 actors to create videos. Facebook then used an abundance of different augmentations on the videos to be able to train the artificial intelligence to be more in tune with organic material.
The DFDC’s top-performing model was able to generate an 82.56% average precision among the deep fake materials that were publicly distributed during the challenge, but when it was time to bring the algorithms to a different “black box” dataset that the participants weren’t able to view prior: Selim Seferbekov took the crown and achieved his 65.18% average precision. Although the lack of other statistical figures for this data is worrisome I do believe that this is an important improvement. The participants’ innovation to find deepfake classification techniques will be monumental in finding a proper solution to fend off the wave of deep fakes that could occur.
Facebook notes that because none of the participants achieved 70 percent on average by using unseen deep fakes that this is still an unsolvable problem that needs work. Facebook’s research teams have always been leaders in new technologies and I hope the open collaborative approach to improving deep fake detection will continue to allow us to have an open and innovative technique to fight deep fakes.
Source : Facebook AI
