PG - Disinform your Surroundings: AI and disinformation campaigns

Howdy Folks um well this is interesting this is technically my first be it's uh Las Vegas at least so um nice to see everybody um of course like you see we're going to go ahead and talk about Ai and disinformation campaigns today uh quick disclaimer this talk will have political stuff duh um oh my gosh you can't do that in a talk I don't make any statement about my own personal views or whatever in this um the other thing that there's going to be a very short section speaking about non-consensual image generation not graphic not explicit there will be a warning all right let's get started so yeah Howdy Folks um that would be me

uh if you see any of those those little stickers around con those are mine and they're now going to be yours you can pick them up um I uh I love animals all forms of Fiber Arts I raise fiber rabbits and goats um I run the main picking meet up right now so if you're ever in Maine come say hi you can get there from here and uh yeah you can also come up to me after for some stickers and stuff if we have time uh I've got some little 3D pry bunnies too all right so that being said uh this idea came from an off-hand discussion with a friend on AI and disinformation uh we chat out a bit about how much crap

is flying around the internet that's just generated it's just there it's an absolute swarm of content it's difficult to see through in many ways it is cyclical though as with any new technology the initial Title Wave comes with trials and tribulations of cross boundaries uh ethical concerns application proof of Concepts you know the deal uh so that's where this talk was born we'll go a little bit into some common tools for each generative technology how it's done on a high level and some examples of it being used in disinformation campaigns um we'll also talk a little bit on the long-term effects of these campaigns because there are actually some and there's a lot more than you

might think um and then finally what you can do about it I'm a very strong believer in talks with a call to action uh otherwise I've given you a problem and not a solution uh this is by no means exhaustive uh it might not contain your favorite tools and tricks uh most of these are going to be versioned more user friendly so that people can just kind of get into them some that are going to be web- based Etc um and I'd be remiss in not pointing out Pope trip over here uh some of you probably recognize this this image came out and it was passed around with a lot of folks inventing their own narrative for it um

it was created in mid Journey uh posted on Reddit and lots of people started posting it as if it was a real image and commenting on it as if it was a real image so very good representation of the talk to come okay differentiating these is not going to be what the talk at hand is actually going to be about today not 100% but let let's get this out of the way uh artificial intelligence is a set of technologies that enables computers to perform a variety of advanced functions think your ocrs Etc um it's technically more of a concept it's not a direct application of the concept machine learning is a subset of AI it's

a subset this is more the application of AI Concepts into very specific pieces right so feeding in data and having an average in into a training result or whatever all right so let's get started on audio and video deep fakes audio deep fakes use multiple iterations over very large quantities of data to train their output things like accent tone speaking Styles and emotions can even be introduced into the model which is shaped over time as with all Ai and machine learning applications this gets better over time and with more Hands-On management of training data so you make sure there's no crap getting in there um think of it like carving a mold one that you've

taken 5 minutes to carve won't be as fine and accurate as one that you've taken 15 minutes to carve and the final result of your mold is going to look a little bit better so rnn's recurrent neural networks and CNN's concurrent neural networks are two types of generation they have the equivalent of memory and analyze sequential data one after another in the case of recurrent like text and videos and spatial data which is more in-depth analysis of a single object or conceptual design as in with convolutional and with image generation uh deep fakes currently run into a lot of problems that we use to identify them by usually this is stuff like weird background noise mismatched

features confusing Dimensions uh or even just general problems with sequencing in the case of rnns or videos now we have some tools here that are commonly used in this sort of generation these slides will be available on my site um and presumably on the conference site soon I think they said that I'm not going to play this it's too long um so long story short that is an impressionist who's doing are you kidding me there that is an impressionist who is doing his impressions as his face is changing as he's adding on deep fake you know um it's not good are you kidding me what is going on okay no it's just loading um so

it's not just information though uh it's a performance and it's a pretty darn good one at that uh here we are on the strip in Vegas we're surrounded by lookalikes some of them good some of them really not uh what's the difference what makes disinformation it's partially context uh if we don't have control over context and narrative things start to spiral out of control almost immediately it used to be that talking the talk and acting the part was the way that you were able to get information across if you faked it you made it uh nowadays there are a lot of ways to control a narrative in addition to that using fake accounts to support your own

views acting like you're in a different position than you are when you're making claims and of course the creation of convincing media that applies directly towards your own goals now this video is a great way to showcase the capabilities of generative Technologies and how seamless they can sometimes be um in parts of this video If you go online afterwards and watch it um you'll see that there's some tiny mistakes problems with clipping but usually stuff like this in any deep fake is is not going to be super noticeable it's not going to be super easy to find so let's get the elephant in the room out of the way immediately which is elections when I started drafting this

talk a year and a half ago I didn't expect that things would move this rapidly but here we are uh you may have seen some discussion on the use of generative technology in rooc calls are similar if you haven't that's one of the things we'll be looking at briefly um so attached to this is a clip of the New Hampshire audio uh so some of you will recognize this but there was audio sent to New Hampshire voters voters in the form of robocalls um long story short it was supposed to be Biden saying don't don't vote just don't vote don't vote don't worry about it um and if you hear this some of you will be thinking well

that sounds like garbage uh obviously it's fake it doesn't sound 100% right it sounds litted and funky but that's not the point um as some of the social engineers in the room may know it doesn't take Perfection it doesn't take pure believability usually all it takes for information like this to be the intended action is the right scenario a robocall during a fevered election targeted at voters won't catch everyone but it will catch some people and it did um 11 Labs voice cloning software was used for that sample most of the tools used in this presentation and for these samples you'll find are extremely easy to find or use many of them being free cheap open source available in a

web browser okay so here's the other problem determining fakes is still difficult some of the factors used in determinations can also be seen in untampered media uh as a result we've been seeing cases of real videos being claimed as deep fakes uh this causes other problems with court proceedings getting stopped up or the line otherwise being blurry an example of this was when Elon Musk went on stage at a Los Angeles Tech conference and claimed self-driving capabilities for certain models of Tesla cars in 2016 there weren't those capabilities so Tesla later rebutted these claims as being deep faked footage something that did not amuse the judge as much as I think they wanted it to and that isn't even the first

instance uh the defense in a capital riot case attempted to Discount prosecution arguments by pointing to video and image evidence claiming that they had been altered and referencing offhand to the possibility of them being deep deep faked so with that being said let's get a little bit more into generative text this has been around a lot longer than some of the previous stuff and had a history of being just a neat thing to play with but it always had problems with memory and the ability to keep up a conversation with modern llms better data sets more money to throw at the problem uh generative text has become an extremely useful tool for a lot of

applications including disinformation these are built on data sets and work by calculating frequency or likelihood of characters uh it's why it can be so easy to algorithmically pick up on if you calculate the likelihood of the next characters words sentences then you can easily figure out how likely it is to be from a machine it also suffers from feedback Loops of ingesting its own Creations causing way higher frequency of some words than is actually

normal okay let's talk about an example that some of you may have seen and that may re raise a few more flags for the way that this technology can be used in disinformation campaigns there's hallucinations of course you may have search something in your browser only to find a ridiculous claim being bumped on to the top of the search like the idea that Kenya does not not exist in Africa and does not start with a k sound but in fact is spelled with a k sound that's not all though while hallucinations that drag up confusing results and features to the top are bad even worse is AI gaining liability for its Generations like in a defamation

case uh from Mark Walters against open AI for false and harmful information about him embezzling money there are also dumb applications of generated text like the lawyer who fake citations generated by chat GPT in an actual court case but worst are the intentionally sneaky ones like fake websites being generated by Russian influence networks or the potential for its use in fake product reviews these things have actual effects on Politics on industry on Law and General Social trust in ways that we may not be able to measure given their current widespread and hidden nature

okay with that here comes our warning about non-consensual sexual materials if this is something you don't want to hear I suggest you leave uh it isn't explicit it isn't anything particularly big or graphic but it can be upsetting so image models tend to utilize diffusion this is a probabilistic generative model that makes use of noise injections and learnable transformations to generate realistic images from random noise generative images commonly have problems averaging backgrounds and Tiny details outside of the focus of the image images that are given greater iterations on diffusion models tend to not have this problem as badly but things like image format Focus connection of object image uh edges and color or lighting inconsistencies are really really common

problems as some of you may know so now we're going to get on to the internet's favorite thing and I'm not talking about Taylor Swift but I also am so this is a very specific set of incidents that tangentially involve Taylor Swift some of you may remember that in January of this year uh there was a massive onslaught of not safe for work images that were going onto Twitter SLX and unsurprisingly these were generated while at this point most tools have some basic safeguards to try and prevent this type of generation users were quick to do what they do best which is circumvent every single attempt to contain them thankfully for Taylor she's got enough of a following that it's really

easy to dismiss these photos and say h of course that's fake but what about everybody else what about your average Joe for those working in jobs that are significantly less forgiving one erron allegation can destroy an entire career or an entire life a Creator offered on Discord to make a f minute deep fake of was known as a personal girl meaning anybody with fewer than 2 million Instagram followers the amount that they were requesting for this was $65 this concept unfortunately doesn't stop there either back in May a man was arrested for possession of thousands of generated images of child sexual abuse material and he was even sending these images to Children one of the problems

that comes from this isn't just the material itself self it's the training material that had to be used to create these images that means that random scraped photos of children were used in the non-consensual creation of child pornography and in the case of any generated images many of the materials used in training data were likely used without full permission of the creators or uploaders sometimes no permission at all so now that that's over with in March of 2023 an image was widely shared across the internet of smoke billowing from the Pentagon of course as the title of this talk was suggest it wasn't real uh but even so it was shared by numerous news outlets including Russia

Today and a faux Bloomberg account on Twitter and yes there are a lot of those um reports of an explosion near the Pentagon in Washington DC was the headline now in and of itself that seems like it wouldn't be all that concerning but the timing of the Images release perfectly matched with the opening bell of the stock market causing a 0.3% plummet in the S&P 500 gold prices climbed briefly General Panic ensued and yet here we are today no Pentagon issues a mostly recovered if otherwise generally not very good stock market larger surveys have been finding that this phenomenon of political and Market effects from generated images was worse post 2016 and that elections seem

to cause a surge in popular use go figure humans inherently want to believe and talk about things that feel Larger than Life that surprise them or that invoke a ton of other emotions it's why rage bait is a thing or Doom scrolling we pick out errant behavior from normal patterns which leads me to the next part of this talk don't panic this isn't the end of the World while it's really lame that we have to do this we have to be thoughtful of the things we choose to share and review interactions and engagement Empower all disinformation campaigns not just AI based ones with that trust-based processing and emotional Consciousness are important ask yourself who's promoting

this why how do I feel about this are there any indicators of deception we inherently desire for things that feel good or agree with our views to be the truth and that desire for the truth should be something that we train ourselves to acknowledge and maintain a healthy skepticism towards so something that you can do is use resources available to you like the AI incident database or the True Media project contribute by adding entries of observed activity to these projects you can also create an ethical note ethical proof of concept it it's already quite easy a lot of people do it but show people how easy it is to make generated content that's believable and ensure

that you have the follow through to properly educate them on it legal efforts are also starting to gain traction for instance the Federal Election Commission held a procedural vote that's received a lot of petitions against the use of generative media in election campaigns this is a great step forward it's a show of awareness and an active attempt to do something about it the final thing you can do is teach friends and family how to identify generated Technologies and how dangerous they can actually be I know it's easier said than done one day though we may see a time when a generated piece of media causes irreparable damage to larger society not just in a single incident or a small

group of people that's already bad enough so let's mitigate or avoid that as much as possible [Applause]

questions anyone is that kind of like a s it's pretty similar yeah so the question was is the incident database uh kind of similar to Snopes but for disinformation and Ai and yeah it is there's even a tag that you can sort by that is disinformation uh let's start over

here so question is is there any technology out there that can help identify deep fakes um and the answer is yes there are a few um more of the commercial options are gaining ground than the op Source ones I didn't really find all that many good open source ones um but yeah they're out there um so we should start seeing more of those soon great job today thank you you tell us why you got interested

in so the question oh I go you uh so the question is how did you get interested in this and it's just it's fascinating to me I started in forensics so I have an inherent desire to kind of look at things that are doing bad um and I think we have time for one more so the oh acon and PC uh proof of concept basically just showing that you can do it some of those links if we go to your get yes you will so I have these slides already up on my website um my website is well my Twitter handle wait do I have it on here wait wait wait I think I think I have it on

there there it is uh so I posted it on Twitter as well but it's also on my website which is that there is the domain for it so um yeah uh thank you guys so much for coming I really appreciate it