Deepfakes - a meditation on the future of video as evidence
Until recently, video modification tools have been exceedingly expensive, and often custom developed to the tasks required of them. Historically, movie studios, and corporations with large video creation budgets dominated the video editing industry. It was not unheard of for video creation and 3D rendering tools to be created specifically for the demands of blockbuster videos, so that the movie studios could showcase the creation tools newest features, by fueling their creations with gigantic budgets. Peter Jackson’s Lord of the Rings Trilogy is a prime example of this. The tools used to create the CGI effects were designed and programmed to fulfill the demanding expectations of the developing studio. To be able to create video scenes that are believable, large amounts of money are often required to provide the skill, the software, and the computational power required.
Video editing tasks (such as cutting footage, splicing feeds, etc) are still expensive to do. The biggest distinction here, is that these tasks remove media, they do not create. When splicing feeds, media cannot be constructed to fill gaps, instead, media is cut away to create a more seamless splice. These tasks, even without creating new footage, are still expensive, because the digital tools required are expensive, and the skills required to use the tools are still specialized (albeit in a slightly different field than 3D modelling and media creation). These tools are often used by media companies, or consulting firms that contract their skills on a per-contract basis.
Outside of these two areas, video editing is fairly rudimentary. Home users don’t have access to sophisticated tools to edit video, and the tools that they do have access to are not designed to fool viewers with accuracy. Photographic tools, such as Adobe Photoshop, have become available to home users to modify static photographs.
Many home users have become proficient enough with this tool, and others, to be able to modify images with enough accuracy to fool the average viewer. Many citizens now are skeptical of static images, due to the ease with which the images can be manipulated. To submit a photograph as evidence to the court, a witness would have to testify under oath that they either took the image, or was responsible for its creation. A digital photograph is just too easy to tamper with. Most people would consider video footage to be different. Many citizens assume that video footage is too complicated to fake, or to manipulate, as it has been in the past.
But they are wrong.
Recently a new tool, named ‘Deepfakes’ after the programmer that wrote it, has been circulating the internet. This tool takes advantage of the advances in computing power available in home desktop computers, and improvements in the field of Artificial Intelligence, and Neural Networks. These advancements allow what was once impossible: the ability for a home user to convincingly modify video footage. Not just modify, but by modifying what was once viewed as the hardest thing to create or modify digitally: The human face. This tool allows a user to submit a photo library of facial images, which is “learned” by the neural network. Once learned, the neural network can then programatically modify video footage to resemble the “learned” face. Although this seems far-fetched, the results are shocking. The tool takes approximately 100-300 photos to “learn” someone’s face. Once learned, the neural network takes approximately 12 hours to swap a face in a 5 minute video. This can be done without any video editing skills or experience. When compared to thousands of hours of 3D modelling, video rendering, and production passes, a measly 12 hours seems to be an impossible feat.
Currently the resulting videos seem to be most influential in two fields, pornography and politics. Face swapping porn stars with video actors clearly demonstrates the tool’s abilities, while playing into the creator’s interests. Popular videos include Adolf Hitler giving a presidential speech to Argentina, and Scarlett Johansson as a porn star. Neither of these videos are real, but upon viewing the footage, it is hard to discern that they are not. It is entirely possible if you did not know what Adolf Hitler looked like, you would not notice that anything was out of the ordinary.
This creates a few interesting issues when it comes to the courts. If video footage is submitted as evidence, how, and how intensely to we scrutinize it? Do we assume that the footage is not tampered with, just because it is plausible and visually “clean”? Video surveillance footage (like those from security cameras and hand-held cameras), is often low quality, and therefore would be easily modified or created using this tool. It would not allow for drastic modification, such as replacing a car with a horse, but it would allow for replacing one suspect’s face with another.
Another question that will have to be addressed in the future, is whether or not this type of video modification and publication constitutes impersonation, or identity theft, by way of obtaining photos or media without consent, in order to publish or create a digital “persona”. The next question here is whether the courts would view this as use for personal gain, or if the requirement is needed, as publicity could be argued for the gain. If this is the case, then the impersonation or identity fraud provisions would clearly apply, and the publication would constitute a criminal act in and of itself.
Due to the prevalence of social media, obtaining digital images of people’s faces is easier than ever. Many people upload thousands of photographs of themselves, which makes them easy targets to use with this tool. This means that the legitimacy of video evidence is subject to the knowledge of its creation, and the assurance that it has not been modified. With tools that are easy to use, inexpensive or free, and publicly available, that are able to convincingly modify digital video footage, the court needs to approach any video footage presented to it with a level of doubt.