Thought Leadership
Microsoft Research Asia recently presented an artificial intelligence called VASA-1, which is able to generate highly realistic deepfake videos from a single photo and an audio clip.
VASA-1 uses machine learning techniques to generate realistic videos. Trained on extensive data from the VoxCeleb2 dataset, the model can not only precisely match lip movements to audio recordings, but also simulate complex facial expressions and head movements.
The videos generated by VASA-1 have a resolution of 512×512 pixels at a frame rate of 40 frames per second, which ensures a smooth and natural display. The low start latency enables the videos to be processed and generated quickly, making the model suitable for real-time applications. Despite this high quality, there are currently still recognizable characteristics in the videos that indicate AI generation. But it is probably only a matter of time before even these small errors are eliminated.
The potential of VASA-1 is considerable – for better or worse. However, there are serious ethical concerns associated with the ability to create convincing deepfakes. Microsoft has made it clear that the model will not be made publicly available until it is assured that it can be used responsibly.
The regulation of deepfake technologies remains a global challenge. Various countries are already working on legislation to curb misuse and promote innovation at the same time. In the US and the UK, measures have already been taken to combat non-consensual deepfakes and protect public safety.
One case made it clear that this is already a real threat: in one case of fraud, around 25 million dollars were stolen from a multinational company in Hong Kong through the use of deepfake technology. Local media reported how fraudsters used the perfidious technology to deceive an employee in the company’s finance department.
