Texas A&M Engineering: SoundBytes

Engineer This!: Dive into detecting deepfakes (Featuring Dr. Freddie Witherden)

February 11, 2020 Texas A&M Engineering Communications Season 1 Episode 24
Texas A&M Engineering: SoundBytes
Engineer This!: Dive into detecting deepfakes (Featuring Dr. Freddie Witherden)
Chapters
Texas A&M Engineering: SoundBytes
Engineer This!: Dive into detecting deepfakes (Featuring Dr. Freddie Witherden)
Feb 11, 2020 Season 1 Episode 24
Texas A&M Engineering Communications

Deepfakes have made headlines more and more in recent years, with many discussing way to detect and combat these computer-generated images. At Texas A&M University, an ocean engineering faculty member is working with his students to develop better ways to tell the difference between real and fake images. We sat down with Dr. Freddie Witherden, assistant professor, to learn more.

Show Notes Transcript

Deepfakes have made headlines more and more in recent years, with many discussing way to detect and combat these computer-generated images. At Texas A&M University, an ocean engineering faculty member is working with his students to develop better ways to tell the difference between real and fake images. We sat down with Dr. Freddie Witherden, assistant professor, to learn more.

Hannah Conrad:

Can you tell the difference between a real and fake profile picture? This week, we're diving into deepfakes. Welcome to Engineer This!

:

Intro Music

Steve Kuhlmann:

This is SoundBytes. I'm Steve Kuhlmann, and with me today is my cohost Hannah Conrad.

Hannah Conrad:

Hi there.

Steve Kuhlmann:

And we're joined by assistant professor in the Department of Ocean Engineering, Dr. Freddie Witherden. Alright, well, Dr. Witherden, thanks for joining us.

Dr. Witherden:

Thank you.

Steve Kuhlmann:

So we wanted to start off today as basic as we could go with this research. Uh, could you define for us what a deepfake is?

Dr. Witherden:

So a deepfake typically in the terms of a deepfake image is an image normally of a person, but it can be other objects, that's completely synthetic. So it's been generated by computer algorithm to look and seem as lifelike as possible, but it doesn't actually exist. And the term deep comes from, it's a reference to the algorithms that generate these images. So how they're generated on a computer normally uses a technology called a deep convolutional neural network, hence the term deepfake.

Steve Kuhlmann:

Okay.

Hannah Conrad:

I feel like deepfake technology has become kind of a big thing nowadays. It's kind of a hot topic. How did you get interested in this research?

Dr. Witherden:

So I've known about deepfakes, I guess the same way that kind of everyone else found out about them and heard about them mostly with the media. But my research interest mostly comes from my grad student who kind of proposed this idea that he had for a way of detecting these deepfake images. And I said, okay, let's, let's give it a go. Let's see how well it works. It seems initially like something vastly different to what an ocean engineer would normally do day to day, but quite a bit of my day to day research involves machine learning, applications of machine learning, and in the same way that deepfakes try and synthesize realistic looking pictures of people, some of my research involves using the same technology to simulate fluid flows or generating fluid flows without having to do a full simulation. The idea being that this may be less expensive than a traditional simulation. Instead we can just get one of these relatively compact algorithms to generate fluid flows. So there is application of this technology within ocean engineering, aerospace engineering and a range of other fields that are heavily dependent on computationally intensive simulations.

Steve Kuhlmann:

That's really cool.

Hannah Conrad:

That is really cool. I think when I think of deepfakes, I always have the idea of like, I dunno the worst case scenario, so like identity theft or fraud and, and things like that. Why is it so important to identify what's real and what's not regarding deepfakes?

Dr. Witherden:

So I think things like identity theft are initially what you might think of as some of the worst applications of it, but on a more practical basis and how a deepfake's likely to affect us in the next year, the next six months. I think it's mostly going to be around social media, social media profiles. One of the traditional ways that people try and tell is a social media profile on Twitter, on Facebook real is we say, "Okay, how long has this account been active? Does it interact with other content? But more importantly, is there a profile picture and is that profile picture of a real person or is it just stolen somewhere off the internet?" On social networks like Twitter and Facebook, a lot of fake profiles have been found effectively just using stolen images as profile pictures to try and give them legitimacy and then those accounts do things like post fake news and everything kind of around that. And so deepfakes are quite worrying in this regard as you can use them to generate a realistic looking profile picture that will be found nowhere else on the internet that hasn't been stolen from anywhere, hasn't been taken from anywhere, looks realistic, and then from that, it gives that profile legitimacy. People can then engage with it and it can do whatever nefarious things its creators want it to do.

Steve Kuhlmann:

Wow. Well the last thing that we need out there is more social media accounts. Just saying there are a lot out there already.

Hannah Conrad:

All right then.

Dr. Witherden:

These are the worst kind, completely algorithmically generated. There's not a real person really controlling them; it's all completely programmed. Some of the interactions are not a physical person who's controlling that account. A lot of it just being done off the back of algorithm. They push pre-programmed tweets, they like certain pages, et cetera, et cetera. And this is done on mass in very large numbers to try and get as much engagement as possible while having this veneer of being real.

Steve Kuhlmann:

So would that be like whenever you hear in the news, when they say like there are bots out there?

Dr. Witherden:

Yes.

Steve Kuhlmann:

Is that the same thing?

Dr. Witherden:

Yeah. So this would all be bots and deepfake pictures are a way of giving those bots more legitimacy, as these bots can now, say, generate photos on Instagram of them eating their lunch or the kinds of things that people post on Instagram, right, that a bot wouldn't do because how would you be able to get a picture of something that doesn't exist eating lunch. But with deepfake technology, you can generate these images ad infinitum, and they'll look at a cursory glance to be real. And the bot's goal in life is to make its behavior indistinguishable from that of a real person. The idea being that if Twitter wants to delete these accounts, there becomes a high probability they'll also delete accounts of real people. And that's always the objective with this kind of thing is to make it difficult. Even if you have a, suspect of 50, 60% chance, you're still wrong half the time. And so if you delete that account, you're going to impact a real user who clearly is going to be very unhappy.

Steve Kuhlmann:

So how good have deepfakes gotten now?

Dr. Witherden:

They're exceptionally good, I would say in the, if you put the effort in and you're willing to analyze the images computationally, you can probably pick up quite a few things about the image that may lead you to think it's a deepfake, but remember, there are billions of pictures online and if you're just an average person looking at a profile, how quickly can you tell just by doing a cursory glance? They're very good at just fooling you at that level, which is really all they need to do to be effective.

Hannah Conrad:

That's terrifying.

Steve Kuhlmann:

Yeah

Hannah Conrad:

Where do deepfakes come from? Uh, like what, what agencies or what sort of organizations make deepfakes?

Dr. Witherden:

If you want to make the deep fake, you can download some, a program onto your computer that will generate deep, fake images to your heart's content. So pretty much anyone can make deepfake images with very little effort. You don't need to be a data scientist or have any real training or expertise. The technology really is at the point where it's almost shrink wrapped.

Steve Kuhlmann:

That is terrifying.

Hannah Conrad:

This is like 1984.

Hannah Conrad:

So how do you spot a deepfake?

Dr. Witherden:

It's incredibly difficult as the way in which these algorithms learn to generate fakes is generally said to be adversarial, as in it's not just one algorithm that's trying to generate fake images. There's normally another algorithm that's learning to detect fake images, and the two are kind of in this training process competing against one another. It's kind of like the best way to learn to play a game is to play it against someone else at your skill level, and over time, as you play each other, you both gradually get better. If you teach someone chess and then put them up against a Grandmaster, they're not going to learn much. They're not going to have much fun, they're never going to perform well. But if you put them up against someone at their own ability level and they both play each other, they'll eventually get very good. And so you've got one, one algorithm whose objective is to generate fake images given some random data, say turn this bunch of numbers into a fake image, and you've got another algorithm whose job is, given a bunch of, some real, some fake. To label them as real or fake. And so part of the difficulty is as soon as you come up with a mechanism for detecting deep fakes, that can be incorporated into the training process and very quickly that method loses its potency. So it's definitely something of an arms race. There's not going to be a silver bullet.

Steve Kuhlmann:

Wow. So how is your team adapting your research to try to address this problem?

Dr. Witherden:

So our approach is a little bit different to what we've seen in the literature. The idea is that if you look at a real photo of a person, of an object, that will typically be taken with a camera and so that photo will have some degree of noise associated with it. If you zoom in and look at the pixels of a high resolution photograph, there's typically some kind of noise pattern. Now this noise pattern doesn't bother us when we look at the photos and so it's usually not much attention is paid to it. And similarly when these algorithms are being trained to generate deepfake images, they also don't pay much attention to the noise as it's not really a relevant feature. They're more concerned about, you know, does this picture that I'm generating have recognizable say eyes, nose, mouth you know, the high level features that define what a face is in a realistic environment. They're not concerned about the minutia associated with the noise and so they don't make any effort into trying to get that noise profile correct. And so deepfake images that you see online at the moment generally have a different noise profile to those of real images that were captured with the camera. And so by doing some very simple signal processing, we can identify these discrepancies and thus discriminate between real images or fake images.

Steve Kuhlmann:

So like the fake images are just too good.

Dr. Witherden:

I wouldn't say too good is it's kind of the pattern of the noise is different.

Steve Kuhlmann:

Okay.

Dr. Witherden:

It's not so much good noise, bad noise, it's just, it doesn't have the right properties.

Steve Kuhlmann:

All right.

Jenn Reiley:

Howdy. It's your producer Jenn Reiley here.

Steve Kuhlmann:

Jenn, are you a deep fake?

Speaker 4:

Haha, that's funny Steve.

Steve Kuhlmann:

No. Oh really? Just, just say "I am not a deep fake."

Jenn Reiley:

Hahaha. Anyway, I'm here this week to let you all know that something is going to be different in Engineer This! moving forward. We're going to take the Ask an Engineer segment and turn it into its own series. Moving forward. we'll have a variety of professors who we ask different questions to, ones that you might be searching for, just kind of like the everyday, "Hey, why does this happen?" or, "Hey, what does this do?" and hopefully these will be able to answer some of those. But for now we're going to get back into this episode and enjoy the rest of the podcast.

Steve Kuhlmann:

Whenever you're talking about this research, whenever you explain it to people, what is it that makes you excited about its potential?

Dr. Witherden:

I think it's nice at some level as it's combining more traditional signal processing techniques into this field of detecting deepfakes, whereas a lot of the existing algorithms to date have been quite elaborate, quite fancy, which there's, there's nothing wrong with that. I mean, we're engineers, we're scientists. We like to use flashy new toys to try and solve these problems. Whereas this is definitely a far more traditional approach. It's looking at something you might consider quite mundane, but it was surprisingly successful and I think that is, you know, a nice paradigm in some sense. In some of these more traditional signal processing techniques they do have applications to detecting deepfakes. It doesn't need to be some incredibly elaborate algorithm that one needs to devise to trying to detect these things. Instead, very simple approaches can be surprisingly effective. Now of course once this approach gets out there, those generating the deepfakes will retrain their algorithms and this loophole, this discrepancy, will evaporate pretty quickly. Just becomes a game of cat and mouse.

Hannah Conrad:

Is that game of cat and mouse something you're going to pursue?

Dr. Witherden:

I think at some level, depending on if we can come up with any other simple ideas that can work against these deepfakes. Um, I think there are other, far more traditional methods that can be applied here.

Steve Kuhlmann:

Where does this research go from here?

Dr. Witherden:

I think the next thing to take a look at with this kind of baseline method is to try and apply it to video files, which are vastly more expensive from a computational standpoint, but they also give us more opportunities at some level to identify discrepancies.

Steve Kuhlmann:

So we know that there are dangers to deepfakes. Are there any benefits to the technology behind them?

Dr. Witherden:

So I think the technology for deepfakes as deepfake images, it's probably not a net benefit for society, but the underlying ideas of, given some representative examples of something, a personal fluid flow or how a certain physical phenomena evolves, and then being able to generate representative samples of that phenomena computationally, very inexpensively, is exciting from a scientific standpoint. And so the underlying technology I think has lots of application but deepfakes themselves. Maybe not so much.

Steve Kuhlmann:

Wonderful. Well, Dr. Witherden, thank you so much for joining us today, um scaring us and letting us know that you're doing what you can to try and help us figure out what's real and what's not.

Dr. Witherden:

Thank you for having me.

Steve Kuhlmann:

Thanks so much for tuning in to Texas A&M Engineering presents SoundBytes, the podcast where students, faculty and staff share their passion, experience, and expertise. What'd you think? Do you have any questions for us? Reach out at engineeringsoundbytes@tamu.edu. That's bytes with a Y. Finally, the views and opinions expressed in this podcast are those of the hosts and guests, and do not necessarily reflect the official policy or position of the Texas A&M University System. Make sure to tune in next week and until then, from everyone on the Pod Squad, sounding off. Thanks and gig 'em.