The technical, societal, and cultural challenges that come with the rise of fake media
The. Welcome to the Riley data show. I'm your host bend Oregon before we jump into today's episode. I want to remind our listeners that we do have to event series that they can go and attend. And learn more about the topics covered in despite cast the first one is called the strata data conference, which you can find at strata conce dot com. The second one is the artificial intelligence conference, which Yuban find FDA icons dot com. And this episode of the day's show, I sat down with see lien. He is an associate professor of computer science at the university at Albany, which is part of the state university of New York seaway is a leading expert in digital media forensics, a field of research into the tools and techniques needed for analyzing the authenticity of defiles. So over the past year, of course, has been heightened awareness about the REIs of tools for creating fake Mead. So this is mainly images video and even sound finals. But not as well known as that there has been a barrel effort on the deduction side. So researchers in digital image for insects have been taking advantage of advances of machine learning and deep learning, and they are also making progress as far as building for detecting, which files are thanked or not so one note seaway as far of a very strong set of speakers at our awry conference in New York, discarding April, so we have strong roster speaker speaking on computer vision as well as on ethics privacy and security. So I hope you enjoyed this episode. See rail you an associate professor at the department of computer science at the university of Albany in New York. Welcome to the data show. I think you anger ban. So first off your part of this field that I think people are beginning to hear about called, basically digital image forensics. And I think seaway another term that I hear trolling around this image, synthesis. So at a high level described these two fields. All right. So actually when I started working our fifteen years ago, it was known as digital image frenzy. But now, I think the sculptors broadens to be. So the new name is these a'mia friends IX because we're inches eating bird fighting the authenticity of all kinds of these including all deals, videos and images. So this is the IRA we're talking about you know, work and then. The otter term you mentioned is image, synthesis. And this is also very recent research area in machine learning and computer vision where the goal is through ultra massively synthesize or because say Holocene age images that looks realistic. But but this is company generated by autism. So, you know, always doesn't really understand does have the concept of all jocks or environment. But it just based on causal puzzle treating images that we had autism. And we tell the hours that this Xishan real for emerges and always of will wiz to synthesize generate images looks like then I see. So you met sue. So you true out fifteen years. So Dan, the natural response. Someone listening to this podcast would be so also this field. Proceeds deep learning. So what were you folks using back in the day? While back in the day, the learning was numb existence. All our biggest enemy by the time is for the shop. So both actually exists more than I think more of that Hayes and house a very handy tool for people images, and the time the most important has Sern is about the authenticity of this kind of images into place. I'll say, you know, images controllers coming from film cameras. And what are we use as as as the Avidan, for instance, becomes an issue because for fuel refund graphs, come from fume cameras is much more difficult who manipulate them to you know, I didn't something movie something or changing something the image about for digital images because the existence fuller shop, you know, people can do that much easier. So a bike back done. We you've mostly relied on this was so because when people actually fades images, they usually. White you high the trees of wrong Shuman ouija inspections? But when they actually do this kind of manipulation on digital images. They are actually unintentionally changed the underlying statiscal property of those images in our way. If you look through say know, make -nology just like looking at those images, we cut with the kind of natural skull. But you just this is typical macro scope, which is based on a site of status on another vehicle truce or our 'sms. We look through those kind of quote, untold mackerel. Scopes, you be able to see some differences that you know, human is when AVI see and the is the word the basic huddled. Choose were using to tell the difference between an authentic image, and image has been manipulating so fast forward to today how much domain knowledge or expertise and computer vision, the do you need. So in other words are. D stool and five lines automated. Because before you said you had to have some awareness of how Photoshop works and stuff like that. So since you are a faculty member training graduate students. So what is your advice to graduate students? So how much machine learning versus how much of the other stuff? Should they know before the for doing the forensic work? Yes. Yeah. Yeah. Yeah. Yeah. I will say, you know, they're a lose the importance of. So on one side, we need, you know, the the subtle cheeks the father is house. So, you know, following all the current literature on images, synthesis and image relation way need you have a thirty good a working knowledge of machine learning, actually because part of my auto part of resources actually, Indian focus on the confusion. So for me, I have more experience in those areas than you know, the bare minimum for doing friends IX. But for my students, I do request and most of them to get a good grasping of what the central knowledge of machine learning deep learning and confusion on the other hand, I think for two specific of work were working on did talking is kind of image foundries. We do need to understand from a different perspective. Which is the angle that most of all jerus- probably will now pay much attention to are the whole image. Synthesizing rose is and the city so analysis all photographic images. Some producer countries all those kind of images and the potential models were for those six. So was we have the part is our man Biden. And you know, we we are able to use those states models to tell if something's the found in Longueuil Nells await we also use the nacelle all were using are also common sense, right because the fundamental difference. Beaching all anything that is complicated or related is not coming from. Real photographs dot his captured by camera over real seen happen the physical world. So there's a fundamental generation differences. So anything that can tell us this difference will be helpful for detecting the fake media. So I talk about stages was also use other tools like physics like shadow lies of by students. Actually. Oh, so it's a before we're going to the Ashra detectors, actually, let's kind of describe for the audience at a at a high level what some of the tools are for synthesizing these images, and by the way, actually, you already alluded to this. So even though we thought about it in terms of forgeries. And that's there's why tat applications for this for distinct as well. Right. So like game designers, or whatever people who make movies and things like that. So here's a few things seaway listed in. Maybe they're kind of too basic, but feel free to ask. Onto some of these some of the things I thought about our someone can cut and paste inaugurated image. Someone can remove an object from seen swap faces or objects maybe the audience inconsistent with the video and things like that. So so what else are some of the Haman things? Oh, and then the other thing I read the other day seaway is a sometimes there's like a mirror in a photograph. And then the reflection isn't consistent. Right. You talk about the even hype menu relations. Yeah. Yeah. Right. And some images be completely generated. Trump's rash. So we thought about all of us. You know, we have basically every computer used as opening system has a hassle tassels speech synthesizer so called GS. Okay. So dot is the species designers, go, you're basically can create our speech rewards of from abuse attacks by so those things can also be used maliciously. Four maliciously as Audrey, and we had personalized this basically can generating suiciding speeches of one pretty person's voice than not is human the more of a problem. So you said earlier that in many ways to the toolkit for detecting, these things rely a lot on common sense. So what are some of the common sense approaches that you folks us while we being using commas basically as a measure the a referred to our fundamental difference between the generation pro size. All you know, digital fabricated reminded lives in media from the actual real harsher media Joe for his one risen work. We have done is talking this video debate, we deals which are designed using all resumes. Learn with a large number of images and small faces between different people are making the facial expression. So the. Resolves looks thirty realistic. But one thing we realize that can be used who attacked out is actually losing odds Aibling, okay? Because of the wheels of synthesized by fate a lot of those kind of deals do not have a good realistic. I blinking motion. And the problem is does images is a mission. This this model is trained using a lodge amount of images, and if the person making those wheels doesn't do this very carefully just grabbing images from the internet's say, Google image. Search you've got to images mostly with somebody's is old because for Porsche images. We don't usually include an image with Ikwo by. So that is a bias the training beta. And this is gonna call says, we know you the real person possibly because there's a physiological study you saying that a stadium dots four normal healthy person. We blink about was every Tuesday were through ten seconds. So if we see a wheel that. Is longer than say thirty seconds were one minute video with the person that we never blink. That is a sign that this this. We do have some problem. I saw this is the kind of commonsense measuring Dodd, you know, just looking back at what we'll Hunan real person or will Sanders to possess, and what is then like from the image or video where examining so using Kanda differences. You talked to tell Dell's all those media. So if I'm trying to do like kind of a world class fake media. It seems like there's many things I have to take care of Dan. Right. So there's like you said, so there's a link gang breathing. But maybe even there's like digital fingerprints that signal processing can pick up. So my the surface of things I have to fake as quite big. So would you say see way at this point detectors are further ahead than these generators of of fake thing? I. Would not be so optimistic about that. Because there are some reasons, you know, Curie looks like difficult sensitize media wishes tool, but our hand Madore some factors also on the side of the Faulders. The first one is I think is completely non-party article issue is the fact that in the most you'll doing this threatening world's working like myself working in Vania. So every time we while the always on dot can detach of fake deter us for as a for forensic technique. We usually just write a paper and polish. So the the details of our did Huchon always him are not becomes available immediately. You know was find them. So on hand, you know, people making fake reveals they are usually Yunessi craziest of position don't usually publish their the details of their resumes. So there's a kind of imbalance between inflammation on the phone frantic side and Afam resign. So this is this kind of a running model. So far. So that the property issue. The issue is also a little bit not not directly that attacking usual. But he's kind of user habit on the fact that, you know, even though some the face are very low policy when because we have internat- once somebody actually bras, those fake media Aldering turnouts as a typical user open up the cell phone check. It just four spill second. And we see something that is kind of interesting exciting syndication. Noel uninspected way. We have this habit of distributed without Ashi checking all busy. So this actually helps caliphate me as broke pro very Faust. Even though we have always have done probably is no Fassi now. So she stopped the Trump to earth thing is I think again, all the touch autism. We put them all together all the shoes that is our bases for that action are still, you know, if somebody do this, well 'cause you. Or all this Hashemi's IRS. So here's my producer is really unnatural effort and restored if somebody actually with a lot of time with along route source appointed on and just focusing on making, you know, perfect one pretty her fate media to be undetectable. I think that's complete these days. So huddle of thanks. I I think there is also divall Eum of media being generated an uploaded, and I think my understanding seaway is that maybe detectors still the latency of detectors still slow enough that maybe it's not practical implemented of YouTube or Facebook or the volume of thanks. They have to deal with right? The bottom. The bottom of data is is clearly a very unusual. And I think these algorithms are roughly speaking still slower, Dan, the Dedi Wally were seeing on those or up one as you mentioned. But I think there's another more important issue that is kind of slowing down the friends detach. In is the. Politics Mozell diseases hours day work. Well, configured actually to fake media for the problem is also have a high pollen Royds was me as if you actually send a real media, you know, with no money. Malaysian, these hours sometimes made me stakes unflagging them as fake media. Now, the problem is is the EMA we talk about because you mentioned the volume of Mia told the ammonium of media were seeing you and like one person or two point one percent of false positive rate me as we have to look, you know, someone has a look the actual, you know, hanzel follow the navy menas number of a media, and that actually is probably steal auto wrath. Lax, and those are awesome be false, and that is really causing the cannata disbelieve of the always I think one of the technical barrier for all the friends method will been working all so far. Is this height falls? Paulsen ray? So we have all working through style. Dots. To a lower level. There's also a couple of things that PUM into play around kind of the information imbalance. I would say, so I I what's his name Harry for read who's like one of the fathers of this field who've I think you probably studied with right views you buy he he was my PG of the LEGO. There you go. So so he he wrote an essay last year, which basically as I understand it. He said, well, if we if we publish everything that we know about defecting than that will give the forger technical edge. So maybe we should start holding. Right. Some of what we know our theories have like a fine between when we actually reveal what we know in. We should. Yeah. I think that's something. I've heard people UNIFIL Haga Delta's because you know, fifteen years ago west started in this era. People don't really take this very seriously. So, you know, we just most of the work we're doing this is some hall of them Corrales. You know, just play challenging games to see if you know somebody fake media. We talked to us. But now, I think the game becomes real because the the power of the truth that people have a hands, and as you mentioned early on it all kind of knowledge that required to use generating. Fake media become much the threshold on us lower. So the volume of fake media, and the level real is that realism of those they Mia are increasing. And were we need the cash hop, but under the current model, we will right into this intrinsic in Balas, which causing trouble in our and the nature of machine learning and AI in particular right now, or at least over the last ten years, it's all about openness, right? So publish on archive. Right. The shore code on Guetta as expectation this different. Whereas on the other side, they don't have to do that. And then the other thing is I guess as far as information information has been weaponized. And also as even if for the sake of argument, we saved the detectors are better than the generators, but the bad after still have the power of information propagation unsocial meteorite. So by the time you actually detected maybe millions of people of ready exactly what one one loss factor. Singh is also the actual use unto or for this vote for the truest lines the work, right? So for Francis. We don't, you know, we don't got even cash of PISA fate media. It doesn't really we don't get anything. But for people making those fake media big on a log more financial, or you know, some other forms on Sunday with us. So I think that that's also another form of imbalance as playing here. Interesting one thing that caught my eye away assist. New DARPA program on knee difference, which I'm sure your part of. But one of the things they're trying to it's not just a question of. Okay. So we know distinct fake they wanna find understand. How was fake in the why it was faked? Right. Yes. So the upper one is food name of the preliminary call the media runs IX, and he started in two thousand sixteen back. The debate is still is not unusual. Nobody does exist at all. So when the program started smo-, see. Targeting traditional kind of media falls Ray techniques. I talk about like, although shop using fullish off or you know, talks to speech soon as is our or real splicing, those kind of money relations of so we were part of our this dollar meal for one, and I think the goal of the program is to ask her wide of you'd Hegg Reiter, the platform dot can summarize all of our friends like media friends park needs and make it about -able tool industry, or, you know, even like audience user, so I remember can have a have a wage authenticate is a piece of media is authentic. So so that's the goal of the program. We are entering the third year actually, the second half year of the program regional third year on mar and then the last year will be focusing on ins, you raging and packaging to bars. So I am highly hopeful know often for one company. Don't be something provided a lot of the poblic on. And we do have a sort of choice of size or just kind of Murphy. So when you look at kind of cybersecurity field, there's almost like a parallel industry like an underground economy where people sell attacks or malware and things like that. So I imagine there will arise parallel economy in media as well. Yeah, I believe so I think there is there's a not here that somebody will try to, you know, get benefit from out. So everything we've talked about mostly focus on the US. But one of the hunters time interested in 'cause we do have a conference in in Beijing swell as China in their video is massive, maybe even more massive than the US. Right. So the on the phones to the short video to do you detect people talking about fake media in China? I heard Assan talk about Dodd because I'm athlete. Chinese. I I grew up from China. So I know there is a concern there. But Ungava hand, you know, based on my understanding, the the platform company, China has a stricter regulations or user upload media. So I think even though they have a lot of Walloon. There are some kind of Kansas future, all all trades contents your political your from onerous Bax. So in doing all, the I think the also have some form of often, the Asian or fact checking all those radio by I'm not sure if that's actually was done in in an hour is not a way. I I don't know that the hills said there's no headlines like here, I think that there has been a lot of headlines here and concern about the potential, but in many ways, I think actually in Ohio. How do you feel about the seaway? Maybe right now at you. We haven't had like a really bad series of episodes of fake ni in the west. But. I'm glad actually we're ready talking about it. Right. Right. Actually. Yeah. I mean, what the I was interviewed last year April. May we have a lot of lot of people are concerned about the potential misuse of this college -i and their potential impact for me from the actions of Las year. I think we did see a huge wave of those state media for a couple of reasons the first one is this imp- increase ask lated awareness as a federal problem actually helps us I think the media did a very well job of this preventive and proactive action, I do catching the general hobbling the existence of this kind of fake media. And also, you know, highlighting some works in frenzies and giving people some kind of I was intimidating fact there that people making fake media Nolde out there exist techniques, second talk them show. I think I think has facts Donner thing is I think the technology is self is. Now. They're yes. Four year. So the away focus on debate. We of xyzal that debate technology. I follow dodge line of work very closely. So, you know, most of the videos where seeing are still no wis- on kind of artifacts. You know that can be detected. Even though you is is is hardly see them with a quick glimpse. But if you ask you spend some time, you know, five minutes later, you'll find those those videos are are not an are real. So that's I think that analogy is not mature back then to have this kind of realistic fake media yet by NAS vehicle future. Because I I'm seeing technology information, quite Faust. And there are lot of exchange of ideas, I'll be hauled to improve and increase of house, y'all to the generator video. So, you know, I think we're HAMAs which is a good thing. But that cannot save for the future. You know, who knows when when the Rudy could faintly. Will shot and the other thing here see way of courses as we mentioned, the the detectors and digital media. For instance, experts like you do all the research in the open in many ways, many of the fake nedia generators. We hear about are the ones that are being done by researchers like yourself. So there might be group forking that we don't know of Frinta on on the generation size. Exactly. Maybe even state bat groups ransom. Exactly, I think that's the worst is not real world that has the most concerned because as I mentioned this fake media, if they're in sufficient Amal forces an effort put into the size of those media to be made with a high level of realism soul and is vastly. If by Gaba by you know, a state where common level agency soul. I think you know, that that's why we need to work faster. All the frantic side to be able to stop this is way. So in closing, so. Reminder to our listener seaway is going to give at our AI conference in New York this coming April. But I wanted to close by giving you a chance to talk about a give us a sneak peek seaway of some of the things coming out of your computer vision and machine learning lab over the next year. Okay. So we have hog about a friend's IX allowed about you know, taxing those fake media. And I think not only these days with the fast pace of technology ends fall jury choose. I choose of we need to be was that butter. So, you know, so as we are working on the action tack Niks in parallel where now also focusing on for actively protecting us from this tax. So one idea is, you know, we because the social network of the solo warm Facebook treater is Dan Brown, I'll stab child. We basically upload our images. We deal. Those platforms, and this is actually providing data training data for people that as we can using the abbess technologies for free. The tree has rate. They had grabbed the harvest our faces from the images Alito, high resolution image radios for Treanor evening based on your network malls. So one of focusing now is hall. Can we stop that? How can we slow to Deng how come it? Make it a Inc. Inconvenient. Or this kind of harvesting all the phase images and Trina big model. So one thing would be working on is proactive for action was essentially was down to a scheme where we actually in bide some kind of perturbation or you can call it annoys into an image. We applaud to the social network, and then all of this base. Always them. Gambles phases by running automatic face. Hotter and faced on Mark locator to crop all the phases. And then prepared a phase as Trina day of. Ordos David Muir, not for models. What are we trying to do with the purples of inviting bills noises is great the facing hackers and lend Markle haters so that they will not easily finding the phases or a line of phases. But the voice we put there are almost imperceptible to human system. So when a user look those images, they will now see I need problem whistles phase with whistles pigs. Ours was images, but this is only aimed to solve sold on as always. So it was because for tweeting a on your network cheaply, you need at least thousands of images. So if we have the phasing trope heart somebody probably have to go in tacking of the images and probably all manually Dowell slow down the whole overall production full-size, and hopefully will will will deter, you know, them to to use that can tell tag -nology as as an attack. So this is something way believe will be very useful. In the future, and we're spending along dot idea. So I would have about a little bit. Some some free results we have in that area June mytalk the conference so won't won't half of your colleagues. Be mad at you. I mean, we I know I don't fervor model people working in the the size. So use time. We just don't wiz the world arch enemies, guess he's haver. But but indeed, I think, you know, visit a cat and mouse game both side are, you know, improve based on the other side is over vision. So I think in the long run we need to have an hand on the detection side when each you control this problem because you know, in came in a circle this is this is just for stages. Wore kinda intellectual challenging each haver, but forty actual society. I think that the impact the potentially negative impact. Just so emails that we need to be serious about it. Well, seaway Lou this has been great and look forward to seeing you in person at the conference in New York this coming apple. Yes, sang. If you thank you very much. Just a reminder seaway is part of a strong lineup of speakers that are upcoming AI conference in New York City this April. So we have great speakers many topics including computer vision and ethics security and privacy. Euler follow see way on Twitter at you Albany CAS. Thanks for joining us. If you like to show piece of scribe at rate us on I tunes or Stitcher or tune in dot com or cloud or Spotify. And never miss an episode.