Working on the machine to detect deepfakes
The image and video manipulations by deep learning or also known as deepfakes can put a huge impact on the people. According to Facebook, if the team wants to fight with these deepfakes they should fight like fire with fire. Facebook along with Microsoft and several other platforms are collaborating together to create a machine learning capabilities to detect deepfakes on the instance.
This idea of machine learning to spot deepfakes is still quite new but we are living among the generation where every day a new method is used to produce manipulated content so this new artificial intelligence might be our only hope. On a daily basis, more convincing fakes appear and we often ignore them but there is nothing else we fear other than our face being used in some inappropriate videos or pictures.
Facebook along with Microsoft, The Partnership for AI, and also along with multiple universities like Oxford, Berkeley, and MIT will be working together for better deepfake detection strategies.
The data available with millions of images of ordinary objects help us easily detect the difference between real or fake objects but there is no kind of data available to detect or compare for deepfakes and this where we need the advances in AI.
According to Mike Schroepfer, Chief Technology Officer, Facebook is planning to donate around $10 million to resource Deepfake Detection Challenge. The creation of datasets of deepfakes can be pretty difficult because you’ll need to gather information with consent from every person participating in it but since the majority of deepfakes are made without the consent so this means that they might not be permissible for usage in an academic context.
So now Facebook and its partners will be making deepfake content of the source videos and also datasets of personalities that they may need to map onto that. Facebook along with other platforms will be spending money to resource implementing the advanced techniques to generate altered videos as part of the dataset. The data Facebook will be gathering for deepfakes won’t be its user data and instead, a variety of paid actors will do that.
The dataset of deepfakes will be then provided to multiple platforms that want to build solutions for deepfakes and put the results on a leaderboard. Although there no detailed insight on the cash prizes but according to our sources at some point cash prizes will be given out too which might also help create a serious competition among the researchers and academics.
The involvement of the research community in an open environment can help the platforms develop new methods and systems to easily detect and remove the effects of manipulated data from the platforms.
The large collection of genuine and manipulated data might help the community to research and find solutions for this new rising problem of deepfakes. Deepfakes can put a huge impact on the life of the victim so these new dataset collections might help minimize this threat in a new way possible.
The initial testing of the dataset is planned to proceed on the International Conference on Computer Vision in October and the full launch is expected to happen at NeurlPS in December and we are really excited for it.