This seems like a good time to mention Captain Disillusion.<p>A filmmaker and video editor, Alan Melikdjanian, started a YouTube channel back in 2007. He debunks viral videos and adverts, whilst performing in character as the rather surreal "Captain Disillision".<p>Don't let the bizarre performance scare you off - his production values are simply outstanding, especially considering he does this all pretty-much singlehandedly, and has done so with hardly any recognition for the past 10 years. I think his channel only hit 100k subscribers last year, so it looks like he might finally be getting some well-deserved recognition.<p>Here's a few links to some of his best stuff. You'll probably be thinking "WTF is this?" when you first start watching, but trust me and stick with it:<p>- <a href="https://www.youtube.com/watch?v=H_slT5YpBok" rel="nofollow">https://www.youtube.com/watch?v=H_slT5YpBok</a><p>- <a href="https://www.youtube.com/watch?v=KbgvSi35n6o" rel="nofollow">https://www.youtube.com/watch?v=KbgvSi35n6o</a><p>- <a href="https://www.youtube.com/watch?v=NXCtgI2lkFw" rel="nofollow">https://www.youtube.com/watch?v=NXCtgI2lkFw</a>
As he mentions in the article, DARPA currently has a large media forensics project going on. They have split it up into teams making manipulated videos and images, and other teams trying to detect those manipulations. They had a workshop at CVPR-2017. I'm on one of the teams that is making manipulated data.<p><a href="https://www.darpa.mil/program/media-forensics" rel="nofollow">https://www.darpa.mil/program/media-forensics</a>
I had a lot of great college professors, but Hany Farid (the subject of this article) was always one of my absolute favorites.<p>He always placed a strong emphasis on strong mathematical fundamentals, once joking in class that we only think AI is hard because we're "bad at math." And to prove it, he showed us endless examples of how the right math makes problems far easier. For example, I remember one time when he showed us how correctly understanding something like the sinc function would improve directional derivatives of sampled data, and then turned this into a better video motion tracker.<p>He loved working with colleagues in other departments, solving all kinds of fun problems and publishing joint papers. At one point, I think the campus magazine was writing articles on him several times a year, because he always had something cool going on.
I wonder how the courts and elections will survive if this type of fake ever exceeds our bandwidth and investment to detect them. Maybe ultimately no one will believe video any more? Maybe people will ignore fake detection and go with whatever their biases tell them is true? Either way electing or convicting anyone might become too easy to manipulate.
Note that the article also talks about how to detect standards image manipulation as well. This could be useful for me when reading academic papers and questioning the visual representation of results.
We're already at the stage where one can't trust the video evidence, unless it's backed by the camera's signature/encryption.<p>Creating a very realistic fake is now trivial:<p><a href="https://www.youtube.com/watch?v=ohmajJTcpNk" rel="nofollow">https://www.youtube.com/watch?v=ohmajJTcpNk</a><p><a href="https://www.youtube.com/watch?v=nsuAQcvafCs" rel="nofollow">https://www.youtube.com/watch?v=nsuAQcvafCs</a><p>Google is very close to synthesizing realistic voice.<p>It's game over, as far as I can see.<p>It's a matter of time someone creates a fake video of someone famous saying something very outrageous, like nazi propaganda, and it will result in the destruction of that person's career and life.<p>We really need something like Secure Enclave in every camera.<p>EDIT:<p>Another related video:<p><a href="https://www.youtube.com/watch?v=hPksv1gJet4" rel="nofollow">https://www.youtube.com/watch?v=hPksv1gJet4</a>
FotoForensics [1] is a great tool for checking if images are airbrushed/manipulated, seems like they work by checking JPG compression and the RGB-average concept.<p>--
[1] <a href="http://fotoforensics.com/" rel="nofollow">http://fotoforensics.com/</a>
Recently listened to an episode of Radiolab includes an interview with Hany Farid: <a href="http://www.radiolab.org/story/breaking-news/" rel="nofollow">http://www.radiolab.org/story/breaking-news/</a><p>Pretty interesting thinking about the gap between what is possible now and where the human population is with understanding that.
It is intriguing to think about this in reverse. Normalisation of the error rates across an image, and the techniques to achieve that or detect that. Also, rebuild images according to specific camera profiles such that the per pixel statistics match.