The threat of deepfakes – human image synthesis based on artificial intelligence – entered the public consciousness in 2017, when a video of President Barack Obama emerged with convincingly manipulated visuals and audio–the video was a fake. More recently, a new BBC series The Capture brings to life many of the fears about so-called ‘deepfakes’, in which the protagonist is accused of kidnapping a woman based on a video doctored by the British secret services. It may seem fantastical, but such technology already exists today.
Recent research on deepfakes from cybersecurity company Darktrace, found that the number of these types of videos published online has doubled in the last 9 months. While these findings may seem frightening, the hysteria around deep fakes has occluded the potential benefits of this type of AI.
What are ‘deepfakes’?
Let’s take a step back here. Deepfakes are the product of a type of artificial intelligence known as ‘generative AI’. In simple terms, generative AI is a style of machine learning program that learns from real data (audio, visual or textual information) to produce original content. A subset of this tech is the Generative Adversarial Network (GAN). GANs are a type of machine learning whereby a generative network learns to create original data which is then evaluated by a discriminative network, which learns to distinguish these from the real data–this process repeats until the generated data becomes indistinguishable from the real. Applied to deepfakes – these networks can be used to synthesise human images from pictures or videos of real people.
Tackling the problem
Unsurprisingly, there is a rising number of newly formed start-ups claiming to be able to tackle the problem of deepfakes. Truepic is one such company, marketing itself as a ‘video and photo verification platform’ and claims to be able to be up to date with the latest methods for verifying content.
There is a fundamental issue with using this type of software developed by startups. The more widely the algorithms designed to detect certain tell-tale characteristics are used, the quicker they become outdated, as developers creating deepfakes will always be able to find a way to adjust to changes. This creates an unsustainable virus/anti-virus dynamic–because, like antivirus software, it cannot guarantee permanent protection as new viruses are created every day. To stop deepfakes from proliferating, big tech companies need to use their resources to need to step up and address this problem.
We have seen some multi–organisational attempts to head off this ‘threat’. One such initiative is the DeepFake Detection Challenge led by Microsoft, Facebook and The Partnership on AI Coalition which seeks to crowdsource solutions to deepfakes. There are already methods for avoiding fakes – sourcing your news from reliable sources who are required to fact check reporting. Technical solutions could also help, such as watermarking images with digital signatures and recording their provenance on blockchains etc.
Changing the narrative
Generative AI has the potential to have huge benefits for business and society. Applications include human-quality content creation, style transfer, 3D object design/generation and text to image translation. As well as creating content, it can be used for classification, currently the workhorse application for AI.
The U.K. is uniquely positioned to make the most out of generative AI. In 2018, U.K. venture capital investment in AI was double that of both France and Germany and has increased six-fold in the past five years. The U.K. is home to a third of all of the AI companies in Europe, and the U.K. Government and tech industry recently announced the ‘2018 Artificial Intelligence Sector Deal’ (with £1 billion ($1.3 billion) in funding), which is absolutely a step in the right direction. However, the U.K. is probably behind the U.S. and China in this field, possibly because of the lack of understanding about how useful this technology could be. The reality is that the U.S. holds 31% of the machine learning talent pool, and China announced in 2017 a development plan to dominate the AI industry by 2030.
There is great potential for generative AI to support creative industries. Runway ML is at the forefront of the creative potential, offering a variety of technologies such as a storytelling machine that generates images as you write. Generative AI also presents an opportunity for companies to create synthetic data that is expensive or unethical to obtain. Having AI generate high quality ‘fake’ data could be incredibly useful, such as in research where it could be used to test models.
As a nation, we’re often more fearful than others about the negative impact of technology–in the early days of the internet, there was plenty of nervousness about impact on our lives–but in practice it is hugely beneficial to our society, whether it be a patient talking to a doctor from the comfort of their own home or someone finding a long-lost relative across the globe. The U.K. is certainly home to the skills to become world leaders in this field, we just need to see past the fear to the benefits this new technology can bring.