Deepfakes are video, and/or audio that have been altered to misrepresent reality, most often by using the face of a famous person doing or saying something that they did not. While having roots in the very late 20th century, it is only recently that the sophistication of deepfakes has progressed to photorealistic levels. It was not until a 2017 report, in Vice magazine, did the wide mainstream media picked up on the phenomena.
In the very short amount of time since that initial report, the machine learning technology deepfakes employ has moved at breakneck speed. Algorithmic face-swaps which, in 2017, required hundreds of images and many days of processing has recently been accomplished with a single image and within seconds.
A recent development out of Samsung’s research department has developed a new technique where, in place of training the algorithm to fix one face onto another using a collection of expressions from one individual, the new algorithm uses facial structures that are shared by the majority of people to puppeteer a new face. In this fashion, the Samsung researchers have been able to animate pictures and paintings, such as the Mona Lisa.
Attempts to have a machine give an audibly realistic reading from text have been the goal of many. Numerous companies, across industries, are invested in this research, including tech giants, game companies and navigational software providers. Producing realistic audio clips that are possible of ‘bringing the dead to life’ or taking existing samples of real people and having them ‘read’ text, has been around since the start of the century and have only become more believable and more accurate.
Even with technology bounding ahead, low-tech techniques such as slowed-down video or splicing audio snippets, labelled as ‘cheapfake’ or even ‘dumbfake’, which are easy to make with real content being ‘doctored’, can still be found entering the media.
The danger deepfakes pose as a disinformation tool is a major concern to many and has yet to be addressed from a legislative standpoint. Deepfakes have become a pervasive, international sensation, but platform moderation and legislation have failed to keep up, instead relying on the interpretation of existing codes such as those present for identity theft, cyberstalking and revenge videos.
Now that the technology has moved along, the cost and ease of producing a deepfake video is trivial. The cost of recognising deepfake videos, on the other hand, grows with the increasing difficulty involved to identify them. Since humans are quite poor at being able to recognise forgeries, it becomes more imperative that digital solutions be developed. These programs can utilise a range of techniques, although what works in one instance may not work on the next. One of the most promising tools looks at the target’s mouth. While various parts of the film may or may not be faked, the mouth is nearly always altered, to fit the audio, and ‘tells’ may be more obvious. Unfortunately, this tool remains far from accurate and requires further development before it can be used with accuracy.
Into the future
While laws may come into place that target deepfakes specifically, they are more likely to fall under the umbrella of current legislation. It could be expected that with the technology to create deepfakes getting ever cheaper, more realistic and harder to detect by consumers, that media industries will embrace the techniques. ‘Hiring’ separate faces, voices and movement actors that fit a role more fully than that capable by a single individual will be too hard to pass up, especially if it works out to be increasingly lucrative. The higher the proliferation of deepfakes, the more cybersecurity measures, such as detection tools and algorithms, will need to be brought to bear on the issue. At this point in time there are limited technologies available, although that is unlikely to remain the case for long if an initiative from Facebook and Microsoft called the Deepfake Detection Challenge bears fruit.