Deepfake: An existential threat to our information dissemination systems

Deepfake's proliferation poses huge challenges

The Bharatiya Janata Party (BJP) deployed a new artificial intelligence-based (AI-based) technology, Deepfake, during the campaign for the Delhi assembly elections. This is the first time Deepfake was officially deployed in a political campaign, although it has been used earlier to disseminate fake political-slanted news. Deepfake is a portmanteau word combining “Deep learning” and Fake. It can be used to create realistic videos of somebody doing, and saying, things they have not done or said. The technology has been around for three years. It is now being commoditised with many programmes easily available.

In the Delhi campaign, Manoj Tiwari was featured delivering short speeches in English and Haryanvi. The “base” video featured Mr Tiwari actually speaking Hindi. Deepfake was used to lip sync and dub new videos in what sounds like Mr Tiwari’s voice. A Chandigarh-based communications outfit, The Ideaz Factory, created the content and the BJP pushed it out to over 5,000 WhatsApp groups. In this instance, Deepfake was used with Mr Tiwari’s consent. But the realistic nature of such content makes it hard to fact-check and, thus, makes Deepfake a powerful tool of disinformation. The technology has often been misused for malicious and nefarious purposes. 

Apart from unique faces, voices, and figures, individuals have unique postures, expressions, speech patterns, and movements. AI can analyse video footage to extrapolate how an individual would say or do something. Most Deepfake programs use Generative Adversorial Networks (GAN) with two algorithms: One forges deepfakes while the other points out flaws in the forgery, which are then corrected. Related technologies like the DeepNude program can guess what a clothed person looks like without clothes. Ageing and de-ageing programs can accurately guess past or future appearances. It is even possible to create “synthetics” — full-body moving, talking images of non-existent people.

Hence, Deepfake videos can, for example, feature Donald Trump chanting in Sanskrit, or Lionel Messi playing cricket. Mark Zuckerberg, Barack Obama, and other well-known people have already been victims of Deepfake videos. Deepfakes have “embedded” Nicholas Cage, the late James Dean, and the late Carrie Fisher into films, and depicted a young Harrison Ford. There are now ongoing lawsuits by the estates of several dead actors to prevent Deepfaking.

Most Deepfake content has been banned by Facebook and Twitter, with only content labelled for humorous and satirical purposes allowed on these platforms. TikTok also claims to be against Deepfake deployment but it has an embedded feature, Face Swap, which makes Deepfake easier to produce. The state of California is legislating a ban on Deepfake pornography since it is often used to target celebs, and to create “revenge porn” to embarrass former partners. As of now, it’s estimated that over 95 per cent of publicly available Deepfake videos consist of pornography. But Deepfake is likely to see ever-wider deployment in political campaigns, as the technology catches on. It also provides a convenient excuse for politicians to disavow recorded statements they have actually made.

The technology presents a thorny problem for law enforcement, and it is also likely to be misused by authoritarian regimes to manufacture fake evidence against “anti-nationals”. At the same time, it can be used creatively for sports and linguistics training, and creating content with “synthetic” actors and musicians. This particular genie is out of the bottle and the technology cannot be controlled or legislated out of existence. 

Deepfakes’ GAN is specifically designed to help it pass forgery tests, which makes it very hard to detect. Facebook is investing $10 million in creating better recognition systems, hiring actors to make videos for training recognition algorithms. Google has released a publicly available dataset of 3,000-odd deepfake videos, to help train AI to identify deepfakes. But the technology for creating Deepfakes is also constantly improving. If recognition systems cannot keep pace with the new technology, Deepfake proliferation could absolutely destroy the credibility of our information dissemination systems.