Jordan Peele as Barack Obama: This is a dangerous time. Moving forward, we need to be more vigilant with what we trust from the internet. It's a time when we need to rely on trusted news sources. It may sound basic, but how we move forward in the age of information is going to be the difference between whether we survive or whether we become some kind of [BLEEP] up dystopia. [ROCK MUSIC] CHARLIE BENNETT: You are listening to WREK Atlanta, and this is Lost in the Stacks, the research library rock and roll radio show. I'm Charlie. I'm in the studio with Matthew, Cody, a blast from the past, Fred, and Wendy. Oh, and he's laughing on mic now. Each week on Lost in the Stacks, we pick a theme and then use it to create a mix of music and library talk. Whichever you're here for, we hope you dig it. FRED RASCOE: Our show today is called deep, deep, deepfake, "Deepfake." WENDY HAGENMAIER: New technology has unnerved privacy advocates, policy makers, and the general public. CHARLIE BENNETT: We'll let our guest expert explain the details of deepfake. But let's just say that there is mounting evidence that the internet is a terrible place, filled with scallywags, fiends, miscreants, incorrigible rascals, no good rapscallions, and all around scoundrels. FRED RASCOE: So it's not about the technology? CHARLIE BENNETT: No, it's about scumbags. It never really was about the technology. But it does make me wonder what an AI might do with deepfake. Nothing to see. Everything is fine here in 2019. WENDY HAGENMAIER: If you want to join the conversation, the hashtag for this show is #lits421 for Lost in the Stacks, episode 421. Feel free to tweet your thoughts, questions, or unsettling conjectures. FRED RASCOE: Our songs today are about deception, fakery, false claims, and body parts. CHARLIE BENNETT: This isn't going to get gross, is it? FRED RASCOE: Charlie, the online world is gross on so many different levels. CHARLIE BENNETT: You make a good point. FRED RASCOE: And I'm growing to hate it as much as our colleague, Ameet, really. CHARLIE BENNETT: Fred hates the internet. FRED RASCOE: Deepfakes are just the latest outrage. So let's go ahead and start our show with "Fake" by Oxford Remedy, right here on Lost in the Stacks. [ROCK MUSIC] AMEET DOSHI: This is Lost in the Stacks. And joining us is Mark Riedl from Georgia Tech's College of Computing. Is that right? MARK RIEDL: You got it. AMEET DOSHI: Oh, yes. MARK RIEDL: Great to be here. AMEET DOSHI: We are starting off well. Mark is a associate professor in the school and Director of the Entertainment Intelligence Lab. Mark, welcome to the show. MARK RIEDL: It's great to be here. I'm so glad. AMEET DOSHI: So in previous episodes, we've covered the concept and the phenomenon of fake news. I think it's getting out there. But there's this emerging concern that is characterized as this word deepfake. And it's come across our radar here at Lost in the Stacks. And we thought we'd invite you in to give us a sense of what deepfake is and, why is it concerning? MARK RIEDL: Yeah, so deepfakes is a term that's come to refer to a technology or a set of technologies that uses artificial intelligence and machine learning technologies to splice people's faces over other people's faces into videos. AMEET DOSHI: So my dad, every holiday, sends these elves that dance around. But they have my face on the elf's face instead of the elf face. Is it kind of like that? MARK RIEDL: Yeah, it's kind of like that in the sense that it's photoshopping in someone's face. But I think what's different is that photoshopping a person's face into an image is really hard because you have to get the angles and the camera angles and the view angles and all that, to make it match up very nicely. In video, it's even harder because if the person is moving around or their head is moving around in conversational, the way we bop our heads and all that sort of thing, it becomes much harder to frame-by-frame, match someone's face specifically and continuously and seamlessly. And what these neural nets are able to do is basically synthesize new images, so it's able to find the exact angle of the face and blend it into an existing face to the point where it's becoming much, much more difficult to identify whether it's been a Photoshop job. AMEET DOSHI: And so this is where the machine learning comes into play because let's say you've got a 20-second video clip with 25 frames per second. That's 500 images that would need to be hand edited and which is conceivable. But now, that can be done very quickly. And I'm assuming that this technology is cloistered away at the National Security Agency, and no one can access it, only those that have good intent, right? Yes. Please say yes. MARK RIEDL: Yes, that's exactly right. And the interview is over. No, so you're exactly right in that professionals have been able to do these very sophisticated Photoshop or video jobs for a long time in movies and other propaganda sorts of things. The issue right now is that deepfakes has brought this technology to the point where it's virtually point and click. Show me a video. Show me another video of a person's face that you want spliced in there, and press a button. Let it process that data for a few hours or maybe a day, and it's ready to go. And it looks pretty good. And this technology has been released on the internet. You can find it. You can grab it. And anyone with some basic Python programming skills, it's pretty much at the point where you can set it up and let it go. AMEET DOSHI: So my understanding is that, at present, anyway, it's really been famous people, like notable personalities that have a lot of video already on them, that these algorithms would require to learn on who have been affected. Is that still the fact, the case, and do you conceive a point where it might just be anyone that could get conceivably deepfaked? MARK RIEDL: Yeah, well, things are moving fast. The original uses of this were in situations where you had a lot of video of your target person, politicians, actors, actresses, musicians, those sorts of things. But if you think about it, we have a lot of video of ourselves now because of cell phones. And we take a lot of photos of each other. We have a lot of pictures and video of each other in casual sorts of situations. That is enough to use this on anyone. And a very disturbing and emerging trend is using this for revenge porn, where you take a picture of your ex and splice that into pornographic images, and then you share it widely for purposes of humiliation and shame. AMEET DOSHI: Yeah, and it seems like this is where law and policy start to step in, because already, there's been movement in some state legislatures to address the wider issue. But I guess I'm curious about how we got to this point. What technologies had to come together to create this moment? MARK RIEDL: Yeah, so as I mentioned before, it's machine learning and artificial intelligence driving this technology. And about four or five years ago, we really had an explosion of this particular technology that we call neural networks or deep neural networks, which was driven by an explosion in data we had readily available on the internet, as well as better computing power. Some people now talk about GPUs. These graphics cards are now used to do artificial intelligence. So what this meant is we can get more data, we can collect it up easier, and we can process it faster. And this technology-- I'm not going to get into how it works-- but this technology called artificial and deep neural networks came along. And what they're very good at is mimicking data. And if you think about it, video and images are just data. And so deep neural networks came along. And one of the first things they worked really, really well on was things like image identification, facial identification. Does this image have a face? Who is this face of? Et cetera, et cetera, et cetera. And then along came a second type of neural network that's called a generative adversarial network. Now, that sounds very complicated, but the way to think about that is two systems working side by side. One tries to learn to detect forgeries. So is that a real image, or is it a fake image? And another one trying to fool the forger, learning to fool. And they play off against each other, kind of doing this tit for tat until the generator network generates things that are such high quality that the forgery detector can no longer detect forgeries. AMEET DOSHI: Oh, that is it's fascinating to think about. And I think we want to definitely dig into the kind of cat and mouse game that is starting to come into focus here, between the forgers and the forger-- the people trying to stop the forgers, the good guys and the bad guys. If you're just joining us, we're speaking with Dr. Mark Riedl from Georgia Tech's College of Computing. And we'll be back with more about deep fakes after a music set. MATTHEW: File this set under TR148.B78. [ROCK MUSIC] CODY: That was "Imitation Me" by the Brazen Hussies. Before that was "The Camera Loves Me" by the Would Be Goods. And before that was "Have you Seen Me" by In Case, songs-- ah, see, you got to put them at the top, man, not the right order. Those were songs about deception and misrepresentation. [ROCK MUSIC] AMEET DOSHI: Today's Lost in the Stacks is deepfake or deepfakes. And we're talking to Dr. Mark Riedl from Georgia Tech's College of Computing about this technological phenomenon. So in the last segment, we talked about the technology that was needed to create these very authentically seeming fake videos. Do you see any kinds of beneficial potential uses? There's clearly a huge downside. But might there be some other advances that would benefit from this type of technology of machine learning and photosynthesis? MARK RIEDL: Yeah, well, actually, I want to talk a little bit more about the nefarious uses first, because I don't think we've covered it a lot. Obviously, the biggest use has been in pornography, first, putting famous people into pornographic videos as kind of fantasy fulfillment sorts of things, also just to attack people, and then moving into everyday people being spliced into videos, these things. A lot of people say, well, what's the harm in this? And really, the big harm, from my perspective, is it's something that you're doing without the person's consent. So while it's not a physical rape, you are in some sense portraying a very realistic and aggressive attack against the person and using a person's face without their consent. So to the extent that people are considering laws against this, we really have to address the question of, what right do you have to your own image? And what right do you have to your own image in uses or in situations that you are not physically present? AMEET DOSHI: And YouTube and other big internet companies, are they trying to identify and deal with this issue? Or are they basically throwing up their hands, as sometimes they do, and say it's just too difficult to catch everyone? MARK RIEDL: Yeah, well, some communities have come out very strongly against deepfakes. So when the technology first came out, there were communities developed around this technology on Reddit. And Reddit came out very strongly against it and shut down those communities and are very aggressive to make sure that when it does come up, they try to mediate it and bring it back down. I don't remember how YouTube has addressed this, though, obviously, they have mechanisms for which you can challenge and have things brought down. But I don't think they're-- well, I don't want to say for sure how-- it's very hard to detect whether these things have come up in the first place. So often cases, it's in the hands of the victim to self-police. And of course, that means scouring the internet. Sometimes you're not even aware that these things are out there. AMEET DOSHI: Yeah, sure. So of course, there's got to be many researchers that are looking at how to combat this. And what are their strategies? Is it akin to the virus, antivirus approach, or is there something different about a person's face? MARK RIEDL: Oh no, there's definitely an arms race going on between the people who are making these generative synthesis technologies better and the people who are trying to detect these. Right now, I think the generators have the advantage. But there are a few tells that you can look for. As a person, you can look at a face very closely, and there's two giveaways. One is blinking. So for whatever reason, these technologies haven't really mastered the fact that people blink in very particular ways when they're talking and moving about. So blinking can look very unnatural or not happen at all. Sometimes it gets the mouth incorrect, so lip syncing or asymmetries in the mouth. And then there's MPEG-like artifacts, where you just get something that kind of looks blocky or very unnatural about that. That's what humans can do to do that. People are now working on technologies that can also detect those artifacts and identify when things are generated, versus not. AMEET DOSHI: So the blinking is fascinating. So is the word stochastic, that it's kind of-- MARK RIEDL: Yeah. AMEET DOSHI: It's not something that's logical or easily deduced, to the point where even these kind of advanced algorithms can't really mimic that. So in a way, it's this imitation game for your face, like how do you tell It's a computer, or an AI, or a human? I guess I'm trying to draw the analogy between fake news. And with fake news, we advise students, for example, to look at the source, try to track down the provenance of whatever they're looking at. Are there similar approaches that could be used for deepfakes? Or is it purely a technological advance that's required? MARK RIEDL: Well, we've talked about the technological advance. But the standard techniques of identifying fake news of any source, which are mostly textual, also apply to deepfakes. So you look at the source, and you look at the social network. What's the sharing pattern? Is this an entity that tends to not interact with many people until they put something out? And then there's a particular pattern of sharing and resharing around that. That's how fake news and Twitter bots are identified. To some extent that works when deepfakes are political. But the patterns for, let's say, revenge porn sharing are going to be a little bit different. These are real accounts that start the sharing and picking it up. So I think you need a combination of both of these because there are many different uses for deepfakes. AMEET DOSHI: So clearly, there's really negative kinds of things happening with this technology. Is there anything good about a deepfake technology that could be used for medical imaging? Or am I just grasping at straws here? MARK RIEDL: Well, generative adversarial networks have been used to make training of medical image, like diagnosis recognizers, work better because you can generate images of other things other than faces. For deepfakes, which is specifically about putting people's faces into videos, this is a little bit harder to defend. It synthesizes or generates videos. So you do find truly kind of creative artistic uses of this. My favorite right now is an artist, a musician, Charli XCX, who spliced her own face into Spice Girls videos to make a new music video that was kind of a mash up of other things. Someone spliced in Harrison Ford's face into the new Solo movie to make it look like the real Han Solo. There's such a thing. And of course, everyone will splice Nick Cage into everything. AMEET DOSHI: Yes. MARK RIEDL: But other than that, I have a hard time finding things other than these artistic expression sorts of things. AMEET DOSHI: Well, this is fascinating conversation. And we're going to come back and finish it off. And I imagine we'll get into some of the politics of because there is a political dimension, you might say, to this whole technology. You're listening to Lost in the Stacks. And we'll talk more with Mark on the left side of the hour. [ROCK MUSIC] STRONG BAD: This is Strong Bad, and I'm pouring some yam sauce on my stacks because you're Lost in the Stacks on WREK Atlanta, the hard FM. [ROCK MUSIC] CHARLIE BENNETT: Today's Lost in the Stacks is called "Deepfake." F for Fake was a 1974 fake documentary, directed by Orson Welles, about a faker named Elmyr de Hory, a professional art forger. The following passage is part of Welles's opening monologue to the audience. I am not going to do an impression. Ladies and gentlemen, by way of introduction, this is a film about trickery and fraud, about lies. Tell it by the fireside, or in a marketplace, or in a movie. Almost any story is almost certainly some kind of lie, but not this time. No, this is a promise. During the next hour, everything you'll hear from us is really true and based on solid facts. Well, while you consider that, truth teller, the medium and the message, file this set under HV6695.S53. [ROCK MUSIC] CODY: That was "If it's not You" by the Language of Flowers. Before that-- what? Now you're telling me-- oh, man, guys. CHARLIE BENNETT: It's almost like we planned it. CODY: Yeah, that's written here in the script too. How did you-- that was "I Thought my Hair Was my Girlfriend" by the Velvet Cactus Society. Before that was "Your Mouth" by ThunderBunny. And before that was "In your Eyes" by the Kravitz, songs about paying close attention to how body parts look. They can fool you. MATTHEW: So can scripts. [ROCK MUSIC] AMEET DOSHI: We're talking about deepfakes on Lost in the Stacks. And our show today is all about AI and image manipulation, but also the political and policy implications as well. So, Mark, we've talked about some of the challenges of identifying these kinds of, really, like encroachments on personal privacy, and really, criminal encroachments. So there's a policy dimension here. Certainly for those that have been personally impacted, this can be incredibly traumatic. But there was an op-ed written by a senator, Senator Ben Sasse from Nebraska, Washington Post, and he claimed that this technology could actually take advantage of existing political divisions in the US. So we're talking at the scale of millions and millions of people and the kind of thing that could really roil our politics even further, if that were possible. Do you think that this technology could have such a negative effect? MARK RIEDL: Well, he's definitely not wrong. The political dimensions are a little bit disturbing, because now, we can make videos of any politician saying anything we want and having that look good enough for at least passing inspection. Now, I think, for at least the short term, the concern is a little bit overblown because you need a target video and a target a source video. And so it's very easy to disprove some of these videos by finding the source and showing how it's been altered. But really, that's not the thing. So the reason why we even have fake news in the first place is because there's a narrative out there that people want to believe, and they're not going to do a lot of deep inspection to disprove all their sources. Just having it out there long enough, even though it gets disproven, it propagates, and it becomes rumor mill and so on and so forth. And that is what politicians are afraid of. And I think it's completely justifiable. Videos are easier to disprove and combat against because you can identify that source video. The audio equivalent of deepfakes, which is also out there and a thing that we haven't really talked about, you can fake people's voices more and more realistically every day. And in that case, there is no background video that you can have. You just have the audio source. And to that extent, those are a little bit harder to disprove and also easier to share. AMEET DOSHI: Yeah, and maybe more, in a way, nefarious because just the power of sound rattling around your headphones, a soundbite, for example, does it use the same technology as the machine in terms of the machine learning and AI kinds of-- MARK RIEDL: Yeah, very much based on these deep neural networks. Obviously, the deep neural networks are set up in a very different way then the video of neural networks that use deepfakes. But, they all have their genesis in this technology, called deep neural networks. AMEET DOSHI: Yeah. So it seems like this is not a passing fad, that it's just another kind of channel, media channel, that is adding fuel to the fire of-- we'll call it fake news or what you want to believe. When do we stop talking about deepfakes? Or is this a more permanent part of our media landscape? MARK RIEDL: Well, I think it's here to stay. Between both the political uses and the public shaming uses, these are very powerful and very negative sorts of things, and those small few benefits that we talked about in terms of artistic performance. So I don't think that this is a passing fad. It's a thing that we're going to have to deal with more and more. The ways we can combat it, we talked about identifying these things. But really, the broader issues are we have to address how fake news propagates. What are the underlying incentives for sharing and disseminating fake news? And I think, frankly, our media and the people who are on social media, they need to become more aware of these issues and more media or technology savvy and just kind of generally more skeptical. And until society changes in those ways, this is going to keep coming up again and again. The other thing that we really have to worry about is even if this technology doesn't get used-- and it really hasn't been used in the political arena yet, in any serious way-- just the fact that the technology exists means politicians can deny things that they have said for real. So oh, I said something. Right now, there's a lot of retraction. But you can say, well, I didn't really say that. Someone must have deepfaked me. AMEET DOSHI: Yeah, I didn't say that. Or did I? MARK RIEDL: Or did I? AMEET DOSHI: Or did I not? MARK RIEDL: So just the existence of this gives people a plausible deniability of things that have actually happened in reality and are being reported on. AMEET DOSHI: Oh, that's very troubling. MARK RIEDL: It is. AMEET DOSHI: Well, you've given us a lot to be concerned about. But I would have to say, this is very much in the realm of information literacy. And I think that skepticism that you note could-- perhaps this is a propellant for more media and information literacy to be taught in schools and universities and just a part of everyday discourse. Everyone's a fact checker. That's the dream of every librarian. I'm not sure if we're quite there yet. MARK RIEDL: Well, yeah. And I think that's the way society matures itself. So technology outpaces human social evolution to some extent. We have our ways, our culture, and our society, and our norms. And then they get challenged. And we resist making those changes as well. And then change eventually does happen over time. AMEET DOSHI: We've been speaking today with Dr. Mark Riedl from Georgia Tech's College of Computing about deepfakes. It's been a fascinating conversation. Thank you so much for coming on the show. MARK RIEDL: My pleasure. Great to be here. [ROCK MUSIC] MATTHEW: File this set under BF637.D42S85. [ROCK MUSIC] CODY: That one was "If it's Not You" by The Language of Flowers. Before that, it's "We're not What we Appear to Be" by The Bunch. And before that was "Watch Out" by Modus, songs about trusting in something that's not real or claiming that something real is fake. [ROCK MUSIC] CHARLIE BENNETT: I'm so glad Cody visited us today. Our show today was called "Deepfake." FRED RASCOE: And, Charlie, we've talked about deepfake as the technological phenomenon. But here at Georgia Tech, you've brought that into the classroom of first year students. CHARLIE BENNETT: Yeah, I did a GT 1000 session, talking about information literacy. And a big part of it was, hey, you can't trust everything you see on the internet. FRED RASCOE: What did they think? Was this like a new concept to them? CHARLIE BENNETT: It was not a new concept, but the deepfake was a new concept. So you can imagine a librarian telling a bunch of 19-year-olds, you can't trust everything you find on the internet, kids, and the reaction of, yeah, dad. But then I showed them three videos. I showed them a video of President Obama, of President Putin, and of a random citizen, just talking. And each of them was a fake, stuff that Obama had never said, stuff that Putin had never said, and then the final person did not exist. They were a created face, talking. And I think it was the created face that made everybody kind of lose it. FRED RASCOE: Did you prep them? Did you say in advance, these are fake? Or were they surprised? CHARLIE BENNETT: What I did was I said, let's look at the way the video can be manipulated and then to be able to say, that was not manipulated. That was completely created. And they were like, oh, I can see how Obama and Putin, they were sort of changed. You took faces that exist. But then in the third one, that person never existed. That face is not a real face. Those words came out of a fake. Then I could see the lights go on in people's eyes. FRED RASCOE: So they're primed and ready for the coming dystopia. CHARLIE BENNETT: My class is not what did that. Roll those credits. [ROCK MUSIC] MATTHEW: Lost in the Stacks is a collaboration between WREK Atlanta and the Georgia Tech Library, produced by Charlie Bennett, Ameet Doshi, Wendy Hagenmaier, and Fred Rascoe. WENDY HAGENMAIER: Matthew was our engineer today. And the show was brought to you by The White Hats, trying to fight those internet scallywags. CHARLIE BENNETT: Legal counsel, stochastic eyeblinks, and funny walks were provided by the Burrus Intellectual Property Law Group in Atlanta, Georgia. FRED RASCOE: Special thanks to Mark for being on the show, and thanks, as always, to each and every one of you for listening. WENDY HAGENMAIER: Find us online at lostinthestacks.org, and you can subscribe to our podcast on Apple Podcasts, Google Play, and plenty of other fake and real places we don't know about. CHARLIE BENNETT: Next week on Lost in the Stacks, we explore catastrophes of literacy with Daniel Kalder. FRED RASCOE: It's time for our last song today. And I don't know about you, all of you, but I am terrified of deepfakes, fake news, identity theft, basically, AI. I'm afraid of anything that has anything to do with the internet. CHARLIE BENNETT: Do you really hate the internet as much as Ameet hates the internet? FRED RASCOE: Maybe, but I still think that this particular tune is kind of catchy. CHARLIE BENNETT: What tune? [LIVELY TUNE] Oh, god. [MEOWING] You know what? I think maybe I hate the internet as much as Ameet now. Can I just have some reality? Can I just have something real, something I can touch? FRED RASCOE: That's our last song, "Reality" by The Five Americans right here on Lost in the Stacks. Have a great weekend, everyone. CHARLIE BENNETT: And watch out for the cat, everybody. [ROCK MUSIC]