Insights from our Synthetic Content and Deep Fake Technology Round Table

This week, DataSeries, an OpenOcean led initiative hosted another Virtual Roundtable about “Synthetic Content and Deep Fake technology”.

INSIGHTS GATHERED:

CHALLENGES:

The ability to recognise deep fake pictures by humans is very low, while the ability of an AI to detect them is fairly high. Facebook’s Deepfake Detection Challenge, in collaboration with Microsoft, Amazon Web Services, and the Partnership on AI, was run through Kaggle, a platform for coding contests that is owned by Google.
The best model to emerge from the contest detected deep fakes from Facebook’s collection just over 82 % of the time. It was argued that, at the current level, we won’t get to 90%.
On the other hand, the percentage of deep fakes currently circulating on Facebook is in the single digits and there are many other sources of misinformation.
We are easily fooled by video: Our subconsciousness has made a decision whether something is real or not before our conscious mind even starts processing the content.

Authenticity and security — A cat vs. mouse race: generation of new deep fakes technologies vs. the detector of deep fakes. A significant security challenge is also socially engineered cyber attacks, not only main stream deep fakes. Increased accessibility of this technology to the mainstream public will significantly accelerate this race.

A definition problem: There are various perspectives what a deep fake is and sometimes there is also confusion with terminology such as fake news. Historically, facial reenactment & face-swapping was the main deep fake use case. Now the term is used in a variety of situations and other developments such as, for example, voice synthesis add more layers to fool our senses.

The last 10% is hard: Many use cases break at a certain point, especially for more horizontal use cases where the technology is simply not sophisticated enough. For example, creating avatars/assistants for gaming, customer service and AdTech. Interactive use cases are typically challenging.

The quality in deep fakes, synthesis compared to other AI capabilities is not different. Even if you look at image classification — which has been democratised until now, if you try to generalise it — it will break. If you want to take an AI solution into production, it needs to be systematically structured and trained. The same is the case with synthetic data generation > if you can narrow down the use case and know the practicalities of what you are trying to produce, then it will likely work.

OPPORTUNITIES:

Lower costs of production with synthetic content is going to significantly accelerate high quality media production. This in turn will enable a wide set of application areas and use cases for commercial application — even by individuals.

Deep Fakes are fuelling the start of the third evolutionary stage of media.

Top Content: Movies, videos and post alterations: Individuals will be able to produce high quality content, even movies, with very limited resources. There will likely be a whole market for virtual actors to be customized for any purpose. Even after the final video is produced it could be altered for a different story with a new script.

Digital copies and avatars: Individuals such as celebrities could scale their presence by addressing people in their local language or have their team write new scripts for presentations, talks or even commercials. Another use case is customized avatars that guide us in virtual worlds. Related to content creation — editing the video component to a virtual assistant is an interesting area.

Synthetic product placement and fashion: Any products can be placed in media for more personal advertisement and essentially offer new marketing channels. Furthermore, clothing brands could use this technology for their advertisement and e-commerce sites.

Personalization or Anonymization: Consumers can choose to personalize avatars in virtual (e.g. gaming) environments or swap our looks with alternative versions to stay anonymous.

Workplace of the future: Immersive experience and interactive engagement mechanisms will enhance the accessibility of video calls, moving a step closer to semi-virtual reality. Advancement in video synthesis, where NVIDIA has recently shown a demo of how via face-tracking with a few floating-point variables, you are able to reconstruct it on the other side of a Zoom call for example. If we push the boundaries here we will be able to reach a 3-point video format that will allow us to look at people in a 3D format (for instance getting a step closer towards holograms).

Identity verification/protection and blockchain: With a vast increase of use-cases pinned on these technologies, it is ever more important have protection over your digital identity and online reputation. Especially for those who’s authenticity is important to their digital reputation (politicians, influencers, etc). Tools to let someone know whether your image has been stolen to create a deep fake is an appealing use case.
Single shared source of truth (SSOT) is one way to approach this. Meaning, one piece of information put on blockchain gets replicated 10,000 and stays there for eternity. Digital IP can be put on blockchain to track the authenticity of any content. Any modification to a picture gets shared on the blockchain, hence knowing if something is real or not. Some people argue that this is the (only) most scalable and sustainable way to approach the identity problem.