Weird reason, there are huge amount of nsfw stuff on reddit.Reddit suspended the deepfakes subs due to the nsfw clips.
It's due to it being celebs faked in porn scenes.Weird reason, there are huge amount of nsfw stuff on reddit.
A colleague of mine were doing some programming while travelling between Stanford uni and San Francisco. The guy next to him asked him whether he was doing statistics or something. They got chatting and this bloke was working for a Silicon Valley startup who were trying to develop software which could make completely photorealistic pictures from nothing. My colleague naturally asked whether that wasn't a dangerous endeavour what with all the fake news. They hadn't thought about that apparently.Sam Harris has been talking about software like this for a while. Once it’s perfected we’re into a whole new world of Fake News.
Tweet
— Twitter API (@user) date
Data integrity would need to be key going forward. There can't be any potential of data alteration from capture through review.What are the implications for surveillance cams for instance?
Tweet
— Twitter API (@user) date
It is going to need strict regulation on social media’s responsibility to regulate the videos it allows shared on their platforms. I can’t imagine it would be impossible to create AI to analyse the videos and look for key signs they have been generated or doctored. It should be the platforms responsibility to either remove or flag as fake.Very scary to think what people will be able to claim as fake in the future. And scary to think about what kind of things will be faked in the future. The era of misinformation is about to get worse.
At the moment, it is relatively easy to know if something is fake or not. Problem is, it can become virtually impossible to know it soon enough *.Very scary to think what people will be able to claim as fake in the future. And scary to think about what kind of things will be faked in the future. The era of misinformation is about to get worse.
Hard to say. The current methods are still quite primitive and so relatively easy to be detected.We’ve already entered the era of unreality and it’s going to get worse.
@Revan
Would there be no way of having software that could determine what is or isn’t a deepfake?
Cheers. Fascinating stuff. So it will be like a game of cat and mouse.Hard to say. The current methods are still quite primitive and so relatively easy to be detected.
There is a type of neural networks called Generative Adversarial Networks (GANs) introduced in 2014, initially for image generation, but in the last few years they have been used for many other things (still mostly image manipulation). In layman's terms, those networks generate fake images that look quite real and have a very similar distribution to real images. Nowadays, you can generate beautiful high resolution images which are hard to be discriminated from real ones. They still don't work as good in manipulation as in just generating images, and still don't work as good in videos, but we are getting there. The problem is that when GANs become reliable to do DeepFake videos, there is feck all you can do to discriminate those videos from real ones (again in layman's terms, GANs are actually 2 networks competing against each other, where one tries to generate stuff that looks real, while the other tries to discriminate between fake and real. If the generator wins, then there is no way on knowing that those images are fake, cause the generator is so good at fooling the discriminator. Again, this is simplified, but hopefully you got the idea).
Anyway, we are probably a few years from it happening, and many people are working on this, so someone might find new ways of being able to detect fake from real. It is both exciting and scary.
Kind of. The generator actually never sees the real data, it is just trained to do well vs discriminator (who in turn is trained to defeat the generator). So when it works, it really works.Cheers. Fascinating stuff. So it will be like a game of cat and mouse.
for me the abuse is ripe for targetting eg pensioners who only subscribe to certain channels. there's loads of right wing BS facebook groups that could just send through the deep fake videos they want to. it's alright if you're savvy enough but judging by the BS my mum posts on FB i'm sure there's a lot that aren't.I think at worst, depending on how the technology evolves, videos will quicky lose their "evidence" value in the minds of people.
Sure, someone can fake a perfect video of Sanders giving a speech in one of those weird paedophile conventions, but 10 seconds later someone will have replaced his head with Trump's and it will have gone viral too. Even the most dimwitted will realize it's worthless, because everyone watches videos online. Soon people will get bored with it, and only the most creative and funny will get some attention.
I don't see much harm coming out of this, there are far bigger concerns on how social media can influence society. Some areas like forensics and security might have to update some features for digital security, but I think the impact will not go further than that.
the solution is to not let olds vote or handle moneyfor me the abuse is ripe for targetting eg pensioners who only subscribe to certain channels. there's loads of right wing BS facebook groups that could just send through the deep fake videos they want to. it's alright if you're savvy enough but judging by the BS my mum posts on FB i'm sure there's a lot that aren't.
Imagine what Brexiters would believeit's alright if you're savvy enough but judging by the BS my mum posts on FB i'm sure there's a lot that aren't.
you got my votethe solution is to not let olds vote or handle money
I'm not sure I agree, though obviously this is all conjecture. I know a lot of people like you describe, but even someone like your mother probably knows photoshoping exists and would recognize a picture of you on Mars as fake. I just think "videos can be fake" will become very common knowledge, except for a very marginal segment of nutcases/very low IQ, and most of these were beyond salvation anyway.for me the abuse is ripe for targetting eg pensioners who only subscribe to certain channels. there's loads of right wing BS facebook groups that could just send through the deep fake videos they want to. it's alright if you're savvy enough but judging by the BS my mum posts on FB i'm sure there's a lot that aren't.
the danger isn't necessarily obviously overt stuff like me on Mars, but the (potential for) more subtle, pernicious stuff.I'm not sure I agree, though obviously this is all conjecture. I know a lot of people like you describe, but even someone like your mother probably knows photoshoping exists and would recognize a picture of you on Mars as fake. I just think "videos can be fake" will become very common knowledge, except for a very marginal segment of nutcases/very low IQ, and most of these were beyond salvation anyway.
What about the effect that has on real videos then? Say a genuine video is released of a politician in a compromising situation. He can easily just pass it off as fake and if people already distrust videos who will question it?I'm not sure I agree, though obviously this is all conjecture. I know a lot of people like you describe, but even someone like your mother probably knows photoshoping exists and would recognize a picture of you on Mars as fake. I just think "videos can be fake" will become very common knowledge, except for a very marginal segment of nutcases/very low IQ, and most of these were beyond salvation anyway.
I'm on this train of thought. Been thinking about it for a while with this kind of technology being inevitable.I think at worst, depending on how the technology evolves, videos will quicky lose their "evidence" value in the minds of people.
Sure, someone can fake a perfect video of Sanders giving a speech in one of those weird paedophile conventions, but 10 seconds later someone will have replaced his head with Trump's and it will have gone viral too. Even the most dimwitted will realize it's worthless, because everyone watches videos online. Soon people will get bored with it, and only the most creative and funny will get some attention.
I don't see much harm coming out of this, there are far bigger concerns on how social media can influence society. Some areas like forensics and security might have to update some features for digital security, but I think the impact will not go further than that.
If you want to hear something genuinely terrifying, here's a recurrent neural network learning to talk:Recently a channel has been uploading speech synthesises to YouTube. I see this being related to deepfakes in the sense that you can use somebody's voice to say whatever you want it to say. Although there are some funny ones like:
You can essentially create a song with their voice which would lead to all sorts of problems with being able to make music without the artist being there.
Moreover, couple these with deepfakes and you can make somebody say whatever you want. The truth becomes much murkier everyday
Nice. It is quite interesting why the progress on speech/audio has lagged far behind the progress on language, which in turn has lagged behind the progress on vision. The basic techniques are quite similar to each other.If you want to hear something genuinely terrifying, here's a recurrent neural network learning to talk:
How do you mean lagged behind? Language processing is still the most difficult thing out there. Speech and voice recognition has come a long way. Vision and speach are easier to deciper than NLP.Nice. It is quite interesting why the progress on speech/audio has lagged far behind the progress on language, which in turn has lagged behind the progress on vision. The basic techniques are quite similar to each other.
I think he’s referring synthesising language rather than interpreting it.How do you mean lagged behind? Language processing is still the most difficult thing out there. Speech and voice recognition has come a long way. Vision and speach are easier to deciper than NLP.
You’re right. Lapsus from my part, language has lagged behind speech which in turn has lagged behind vision.How do you mean lagged behind? Language processing is still the most difficult thing out there. Speech and voice recognition has come a long way. Vision and speach are easier to deciper than NLP.
That's true, they are very different tasks esp. image generation from NLP and speech generation. Generating images is just about modelling distributions usually with hard boundaries between them but for generating NLP and speech generation you have to model those distributions (usually lot more than image) with added temporal/sequential dependencies. I don't think we have yet figured out that perfect architecture and algorithm to deal with such sequential dependencies like we have conv networks for images. Add to that the importance of pre-processing in NLP and speech makes it even more difficult.You’re right. Lapsus from my part, language has lagged behind speech which in turn has lagged behind vision.
As interestingly to me is that despite that the basic tools are the same (backprop and SGD) at the end, speech and NLP needed quite different tools to succeed (attention and Wavenet). Also, I am a bit surprised why the main tool in image generation (and now manipulating, including DeepFakes) has been from very unsuccessful to totally useless when it comes to NLP and speech (talking about GANs).
At the end, probably not surprising in hindsight, all three problems seem to be fundamentally different from each other and seem to require very different solution. Still everything is NN based though, which is nice.
Yep. Thing is, in images there is temporal dependency too (for example, videos), but the solutions there have been relatively straightforward. I would say that tracking is on a better place than anything NLP/Speech.That's true, they are very different tasks esp. image generation from NLP and speech generation. Generating images is just about modelling distributions usually with hard boundaries between them but for generating NLP and speech generation you have to model those distributions (usually lot more than image) with added temporal/sequential dependencies. I don't think we have yet figured out that perfect architecture and algorithm to deal with such sequential dependencies like we have conv networks for images. Add to that the importance of pre-processing in NLP and speech makes it even more difficult.
I agree with you in terms of tracking in videos but that is still a lot easier task than video generation. I'm not really up-to-date in terms of video generation, but just googling recent works on that, it seems it is not even up to NLP or Speech generation.Yep. Thing is, in images there is temporal dependency too (for example, videos), but the solutions there have been relatively straightforward. I would say that tracking is on a better place than anything NLP/Speech.
Temporal dependencies in vision seem to be much easier to model than when it comes to the two other fields IMO. In fact, often you do not need to make them explicit at all, just treat each frame like an image (with maybe some trajectory smoothing) and works relatively well.
It has come a long way in the last couple of years. For example:I agree with you in terms of tracking in videos but that is still a lot easier task than video generation. I'm not really up-to-date in terms of video generation, but just googling recent works on that, it seems it is not even up to NLP or Speech generation.
