As a kid, I loved impersonators — they were hilarious! But part of the fun was recognizing the faults in impersonation. Impersonation was never supposed to be perfect. But today, faking has evolved to a whole new level.
Last week @deeptomcruise ( a tiktok account that uses deep fake AI to create highly realistic Tom Cruise fake videos) posted four videos which gathered over 2M views each and had many wonder if Tom Cruise is finally on Tiktok. While the misinformation was verified quickly by media sources, the sophistication of AI to make this possible is alarming. In this post, let us understand how such fake videos are created, what their implications are and how can you detect them.
How to create a fake video? Cropping and swapping faces on images has been a problem right since the inception of photoshop. However, back then, videos were too sophisticated for these editing tools to reliably work on. So if you had someone say something on video, that was a strong enough evidence. Today, technology has grown such that even videos can easily be manipulated.
But this process isn’t as easy as you think. Firstly, you need a ton of images of the person (Tom Cruise in this case) showing different facial expressions and angles. Then you need a really talented impersonator (Fisher impersonated Tom Cruise). Lastly, you need a lot of processing power and patience. Processing multiple images and using deep neural networks to automatically identify and adapt the impersonator’s face to Tom Cruise’s face is not easy.
As a beginner, websites such as deepfakesweb or faceswap provide all the resources to create a fake video. However, these videos would not be as perfect as the Tom Cruise one above. The neural network model learns all the movements from the celebrity video and interprets the same for your target image. In some cases, there would be missing information for which the model makes a guess. At every frame, the model then assesses information lost. The closer the impersonator is to the celebrity, the better the video will be.
What are its implications? Fake videos have been circulating for a while, but most of us assume we are cynical when we see contradictory evidence. However, children, people with less education or people with poor internet access are more vulnerable than you think. Due to limited information, they may believe the contents of the fake videos and wouldn’t necessarily know how to validate the information.
Fake videos have especially been identified as a threat in politics. In today’s world where many elections are won by a narrow margin, misinforming even a few people can lead to a significant difference. Further, politicians are easy targets since they are almost always in front of a camera, and often have very similar postures and mannerisms. Other malicious uses could be tampering a person’s social image through inappropriate comments and activities.
While the creators of deeptomcruise stated their intentions were purely entertainment, not all creators may come to public. Further, sometimes it might be too late to act.
How can we detect fake videos? There are AI based software such as Sensity that are tailored to identify deepfakes. However, as deepfakes get better, these tools should too. For instance, this tweet highlights the results of running Sensity on each of the deeptomcruise videos, and none of those videos wer detected for face swapping. The detection technology needs to evolve as quickly so that misinformation is not spread. The creators of fake videos should be incentivized to counter them as well! Further, more foundations such as the News Literacy Project needs to evolve to ensure the most vulnerable population can effectively identify fake news.
As the creator of the videos, Chris Ume, said here “The genie can’t be put back in the bottle. Deep fakes are here to stay”, we need to adjust our expectations to not believe everything we see or hear. As the detection technology plays catch up, it is imperative that we act in our best judgement and be critical at examining evidence that contradicts our hypothesis.
