A year ago, Facebook announced the launch of the Deepfake Detection Challenge, a contest designed to combat the false images popularized deepfakes. These last, for example, can embed the face or voice of a person on another discourse, which suggests that the latter has uttered some words then that this was never the case.
This contest brings together other partners, such as AWS (Amazon Web Service) and Microsoft and is designed to ensure that images rigged and not able to be shared massively on social networks. Especially as the us presidential elections will be held in a few months and that fakes news has already wreaked havoc in 2016.
The model is best achieved 82%
The first edition of this contest took place in the month of December to the 31st of may. Facebook has unveiled the results, indicating that the challenge had gathered more than 2,100 participants from around the world. They have used more than 100 000 videos from a set of public data and the social network has also introduced exclusive videos of deepfakes carried out with the help of actors for the occasion. The social network has specified that no data belonging to the users of the platform had not been used in the context of the contest. More than 35,000 models of detection have competed during these next few months.
If the results are encouraging, they also show that the detection systems deepfakes are not yet to the point. In fact, the average success rate of the devices shown in the Deepfake Detection Challenge is 65%. In comparison, the best model was 82%.
In a statement, Facebook acknowledged : “ The identification of these common features will help researchers to improve their models, but the results of the Deepfake Detection Challenge also show that it is still a problem not solved “.
In the long term, Facebook is hoping to have a tool able to automatically detect the deepfakes present on the social platform in order to be able to prevent their dissemination. If these are to be used for pornographic purposes, social networks are preparing for an increase in the years to come, which would leave the field open to a whole new kind of fakes news even more complicated to identify for the users.
The technical director of Facebook, Mike Schroepfer confirmed, adding about deepfakes they” are not a big problem at the moment. But the lesson I hard learned these past two years is not to be taken unawares. I prefer to be well prepared for bad things which will never rather than the reverse “.