#AI The rise of DeepFakes and the pending attack on our social structure

(Update: Elon Musk publicly announced on 19 feb of his departure from the Open AI project which has created the Fake News program mentioned below.)

Our world is changing and we are not ready.

Inside of last 3 months, two important developments happened in the world of AI. Unless you are an enthused researcher, you would have missed them.

On 12 December 2018, Cornell university received and published a paper called “A Style-Based Generator Architecture for Generative Adversarial Networks” by Tero Karras, Samuli Laine, Timo Aila.

On Feb 14 2019, OpenAi.com published a blog article called “better language models and their implications”.

Both these happenstances don’t look exciting on the surface, but both of these pushed the AI capabilities that we had to new levels. While the cornell university paper dealt with creating fake human images, the OpenAi post talked about their record breaking scores in language generation. One machine could now create photograph of human faces which have never existed on this planet, another machine could now talk like human.

Problem of image

Machines have been capable of generating human faces for some time now. But the challenge was in the realism. Machines had not been able to create the right pose or the right amount of expression which made such faces look artificial.

Researchers at NVIDIA were working on this problem using something called Generative Adversarial Network (GAN). GAN is relatively new concept in machine learning wherein it doesn’t create a human face from scratch, but takes an existing face and using Machine Learning(ML) changes it to such a degree that it becomes a new face. So GAN had been able to create new faces, but faced challenges rearranging the facial hair or styling the hair or changing the pose.

They created a system called Pro-GAN which could mix various images to create a new image, but the challenge of the pose and styling remained.

Last December, the Nvidia researchers published a new paper which talked about mixing the input images in three distinct layers. Their contention was that through this layer by layer building of the composite, all forms of style in the new image could be changed. They aptly named it STYLEGAN.

The following video will give you a good overview of how it works.

The idea has already seen some quick traction.

Recently "This person Does not exist" Website came up using the same technology. Every face you see on the website, has never existed on this planet. It is artificially generated.

Similarly there are projects on GITHUB which have generated a smiling face of Donald trump among others

The problem of language

Social media as we know today has two elements. A profile pic and the posts. When you analyse a profile to be genuine, you take both into account. A genuine photo just posting retweets is a dead giveaway that the account might be run by machines.

Creating natural sounding language has been a challenge due to variation in style, subject matter knowledge and common logical flow. The team at open AI has recently published a blog post which claims to have modeled the next generation version of authentic writing generated by machines on a given input paragraph.

Not only the machine create a believable writeup, it also has capability to comprehend a given passage and answers the questions given on that passage correctly. Look at the below screengrab from the blog (link shared above) where the machine expands a snippet on a fake news item of unicorns found in south america.

Did it feel like machine writing? If you read the entire news item, you would believe it to be 10% true. 10 different outlets writing on it over a period, and you might actually believe it to be true and share it further.

The OpenAI team understands that the tech can be misused and has therefore only shared a small sample for research purposes. The original trained model was trained on over 40 GB Data.

Are we ready for its Impact?

Some time ago, i wrote about the rise of artificial models like Shudu which are taking over instagram and appearing on covers of Vogue. These models require a decent size team to render them and make them appear human. They are more art than science at this stage.

But millions of social media profiles created using a combination of above two technologies will be virtually indecipherable.

Are we ready for it?

Are we ready where a single billionaire or corporation can create such a massive public swing in favor or against a particular policy?

We live in a democracy. Very early in our evolution, we realized the power of numbers and therefore built systems where people together wielded the power more than a single person. Even kings ruling expanses had to bow down to the wishes of the populace. But the population then was genuine. Their views came from collective of reality being lived day in and day out.

This resulted in democracy where populations needs and wishes became the bedrock of government. When income equality rose, we still held on without dissent, because the poor still had their voice and numbers to get heard. There was belief in the system that if there was a grave injustice, people could rise and overthrow the government and the rich in a french revolution kind of way.

You will not believe how much solace has this thought kept our social structures from breaking down. This was the appeal and the promise of the left, which found its way to the poorest of the countries and the most oppressed masses.

But if this voice comes to be controlled by the masses, the social structure will fail.

The rise of artificial population and income inequality

Ever since we made machines, we made them in our own image. We created tools that were extension of our hands. We created wheels which were extension of our legs.

When we created factories, they were a representation of our entire bodies, which still needed minds to be run. But with creation of computers, now we have created a mind.

Our economic growth or GDP is actually a very simple number. It is collective work done by the population of a country. Long ago, the only way to grow GDP was to have more people and working hands. This is the reason that countries like India and China had majority of the world GDP, and this is why people made a beeline to the orient.

As machines grew, the west kind of added an artificial population to grow production and surged ahead in GDP. For agriculture, till there were good enough farm equipments, US depended on slave population. People who worked but were not counted as people.

The end of slavery had more to do with farming equipment than the vision of Abraham Lincoln. Because now there were machines to till the land, Lincoln could ask people to move away from slavery.

The politics of our world has closely followed the economics and the method of production. Now we have even more smarter machines which can produce equal to an entire countries population, but the democratic politics was people's.

Our politics will cease to be our own

With these two developments, the politics will also not be people's. Right now the entire world's sense of acceptable or not is largely guided by social media and this is increasing day by day.

People could trust social media because there was no way to control that (in theory).

But now there is. And that is a problem.

Few incentivised people could meddle with opinions of US elections via data generated by cambridge analytica. But what if there were 2 billion fake profiles spewing a particular branch of philosophy?

Influencing will become super-micro influencing. Every human in the world on social media can be covered by 3 artificial profiles working day in and day out to change the viewpoint by sharing fake news and views. We are social animals. We are prone to be influenced by perceived social behaviours.

Now that perception of the society can be changed by one malicious agent and that too very convincingly.

the age of tomorrow will not be of influencers. It will be of nano-influencers.

Conclusion

The only way to contain these changes is by policy and transparency. UK has recently announced that they are working on legislation to control facebook. I feel there should be more of such steps. The power of social cannot continue with safe harbor any more. UK plans to hold FB responsible for content posted on their platform. The challenge we saw due to fake news also will need to be addressed clearly.

Else every form of system built on human to human trust is at risk. Be it a simple review on amazon which influences you to buy a product to a post that influences you to not vote for a government. I genuinely feel that this time, our society and the societal cohesion is under attack.

Few forces are trying to own the entire society by controlling their minds. If you think that this is too far fetched an idea, read up on how British policy on education was brought over by Macaulay called Macaulayism and how chinese unified because of the Great Leap forward.

New colonialism is just around the corner if we don't act now. It will be a new matrix when we can't tell truth from falsehood. When that happens, our mind would have been imprisoned.

-------------------------------------------------------------------------------------

References:

1. Great Leap Forward : https://en.wikipedia.org/wiki/Great_Leap_Forward
2. Macaulayism : https://en.wikipedia.org/wiki/Macaulayism
3. Artificial Fashion Models: https://www.harpersbazaar.com/uk/fashion/fashion-news/a22722480/how-artificial-intelligence-models-are-taking-over-your-instagram-feed/
4. Cornell University paper on STYLEGAN : https://arxiv.org/abs/1812.04948
5. How does Style GAN work, Architecture : https://www.lyrn.ai/2018/12/26/a-style-based-generator-architecture-for-generative-adversarial-networks/?fbclid=IwAR0bngOpEfmesyWSOC4Y83ctJVwrdltZxCFN2Km0YJiAOL5qXFNGazLZnsE
6. GITHUB Project on STYLEGAN : https://github.com/Puzer/stylegan
7. Open AI blog post on language models: https://blog.openai.com/better-language-models/
8. Ars technica article on how scared OpenAI held back their own research : https://arstechnica.com/information-technology/2019/02/researchers-scared-by-their-own-work-hold-back-deepfakes-for-text-ai/
9. Pentagon trying to control the DeepFake menace : https://edition.cnn.com/interactive/2019/01/business/pentagons-race-against-deepfakes/
10. Axios on how DeepFakes are getting bigger : https://www.axios.com/newsletters/axios-am-08f3c088-1909-4140-9747-bb1e3804fa36.html?chunk=4&utm_source=linkedin&utm_medium=lisocialshare&utm_campaign=organic#story4
11. Deepfake crisis on democracy and privacy: https://www.lawfareblog.com/deep-fakes-looming-crisis-national-security-democracy-and-privacy

--------------------------

Disclaimer: The views expressed here are my own and don't represent the views of any of my employers, present or past.