Scientists develop new way to detect deepfake videos

Irish Times

a day ago

Scientists
at the
Netherlands Forensic Institute
have developed a groundbreaking new method to identify
deepfake videos
by looking for subtle changes in facial colour caused by the human heartbeat.
'In video of a real person, one can detect blood flow around the eyes, forehead and jaw – and that's exactly what's missing in deepfakes', according to lead investigator Zeno Geradts, professor of forensic data science at Amsterdam University.
Deepfake videos –
artificial intelligence
-generated content in which real people appear to do or say things they never did or said – are seen as posing risks to society because victims can be inserted into sexually explicit content, for example, or misrepresented in fake news.
The new technique, called blood flow detection, uses advanced image analysis to identify tiny variations in skin tone created by a human subject's pulse – variations that are absent in manufactured or manipulated footage.
READ MORE
When the Dutch team first investigated the issue 13 years ago, poor video quality made such sophisticated analysis impossible.
Since then, video technology has evolved, and a combination of improved picture resolution – measured in terms of the number of pixels in a video frame – and rapidly developing artificial intelligence (AI) tools has allowed the experts to successfully revisit the problem.
'When we looked at this first in 2012, we were being asked by police to analyse so-called snuff videos showing extreme violence, including torture and murder, circulated via dark web platforms and on encrypted messaging apps.
'We were trying to find a way to establish scientifically that the victims being shown in those videos were real people.'
The fundamental crux was the inadequacy of video compression technology, said Prof Geradts.
'Large video files had to be reduced in size and during that compression process the colour differences per heartbeat were lost.
'But now, more than a decade later, compression methods have improved and we can detect even the slightest discolouration caused by pulsing blood flow.'
How to manage your pension in these volatile times
Listen |
37:00
In the most recent tests, volunteers were filmed wearing heart monitors and their heart rates were then matched against colour changes at 79 points on the face, under a range of lighting conditions and varieties of movement.
The project confirmed a strong correlation between visible colour changes and the heart rate measurements in all the different settings. The results, say the Dutch team, were 'very promising'.
'AI can do a lot but it still cannot generate a convincing pulse', said Prof Geradts.
Although it is used in individual investigations, blood flow detection is not yet in routine forensic use, mainly because it hasn't been fully validated for use in court cases – but that may be just a matter of time.
'In the world we live in, forensic research into deepfakes is more urgent than ever,' said Prof Geradts. 'I sometimes worry that soon everything will be regarded as potentially fake. Then how will we know what's real?'
According to Helpwanted.nl, a Dutch national helpline for victims of online abuse, reports of deepfake pornography and manipulated nude images increased in the Netherlands by 31 per cent in 2024.