Transcript
h_YzuXQlIXI • AI-Generated Fakes: How to spot them, how they're made and how they have been used to mislead
/home/itcorpmy/itcorp.my.id/harry/yt_channel/out/Bendobrown/.shards/text-0001.zst#text/0031_h_YzuXQlIXI.txt
Kind: captions Language: en computer generated video to show apocalyptic what if scenario threat to National Security posed by Deep fakes artificial intelligence artificial intelligence manipulated by AI deep fake photos welcome to the digital era in a world where artificial intelligence can conjure up images so real they're virtually indistinguishable from reality how do we discern truth from fiction hi everyone I'm Ben and today we're diving into a topic that's not just fascinating but also crucial in our information saturated World we're talking about AI generated images they're stunning realistic and completely fabricated and while they showcase the marvels of modern technology they also pose a significant challenge with the rise of disinformation in recent years we've witnessed the power of disinformation it's impacted elections news events and human rights issues around the world these aren't just lines of code creating fun pictures they're tools that in the wrong hands can have real world impact shaping opinions and altering narratives but here's where we can make a difference by learning to spot these AI generated fakes we arm ourselves with knowledge becoming more critical consumers of information this isn't just a skill for teeken through enthusiasts or digital experts it's a critical tool for everyone because in the fight against disinformation every one of us plays a role in this tutorial we're going to look at platforms like this person does not exist mid Journey Dolly and more we'll learn how they work we'll see how they've been misused and most importantly we'll go through how to spot these fakes whether you're a social media buff a concerned citizen or just curious about the digital world this tutorial will be relevant for you before we jump into each of the more popular platforms it might be helpful to cover some of the main tips on how to detect irregularities inaccuracies or mistruths depicted in potentially generated or edited images there are five common tips to look out for first you should always look closely at the image when looking at an image focus on things like hands or faces or positions of eyes AI generated images struggle to replicate hands realistically often resulting in unnatural or awkward positioning they also have an unnatural gaze inconsistent lighting and sometimes even stiff facial expressions second look at the background look for blurred or pixelated elements sometimes there may be Distortion in the reflections repeating patterns which may indicate AI generation third do an image RSE search and we've gone through some of those image RSE search platforms through my ENT at home tutorials where we've used things like Google Google Lens Yandex Tini or others to check if the image appears elsewhere online particularly on AI related forums or websites the fourth tip is to verify the source sharing the content does The credibility of the account actually sharing it show that it's an authentic account or that it's believable and the last one is to verify the event or the claim itself if there's a photo or a video of a fire or an emergency event have there been other reputable resources posting about it sharing ing images about it are there other people talking about the event or the claim so let's get stuck into the first platform which is this person does not exist this person does not exist is a freely available website that applies user generative adversarial networks to create highly realistic human faces when you click ref refesh on the website you can see that it makes a new face each time if you were to image reverse search these faces you can see that they don't exist on the internet because these are faces that are newly created faces and are not real people as you go through the faces you can also notice some similarities and get an understanding of how the generation works and how these faces are generated while I'm clicking through some of these faces you can also see some of the obvious signs that they might be fake faces for example the first thing is the fact that the eyes are always in the middle of the actual image if you were to place your two fingers up to the screen as I click refresh here which you could even do now you can see that the eyes are always in the exact same location even when the face or the composition of the face changes what you can also notice outside of the eyes is the failure of images to actually blur properly each one of these images has issues with blurring or minute Clues as to it being a generated image for example if we have a look at the uh bottom left of this image we can see the blur bluring of this fabric down here we can go through more and we can start to see the odd complexion of some of the faces and some of the poor levels of blurring when you look and pay more attention to the detail of these images you can start to spot irregularities throughout these images for example the glasses go into the head here but seem to cut through the the hairline over this side when there's glasses there's often uh mismatch blurring as well for example here we can see the guy the side of the guy's face we can also see the outline here of where there may have been glasses in the first place where there is potential jewelry we can also see that often the jewelry is mismatched sometimes glasses Disappear Completely as you can see with the frame here other times there appears to be many odd things going on in an image sometimes you will also see body parts or other people in images and you can see that there appears to be some morphing on the side of the image like what you can see with what is potentially a hand so what are some examples of how generative adversarial Network faces or Gan faces have been used in campaigns before in an investigation I did in the past with bellingcat we identified a influence operation or a disinformation campaign operating within Indonesia targeting human rights activists and the information space within West Papa and around West Papa what I was able to do was to identify numerous accounts on Twitter that were using these similar styles of Facebook es which were likely from this person does not exist or another Gan style website and again you can see that centrality of eyes meaning the central location of the eyes in these images which you can often match up with straight lines but also for that further verification looking at for example the failure of the blending of this cap and you can see some of the images that we're able to identif ify just for a further follow-up verification to show that these are indeed poorly Blended Gan images by working with a researcher called Elise Thomas we're able to trace back the origin of this campaign to a marketing firm operating out of Jakarta which built these campaigns as part of a professional package for digital marketing Services these types of faces used used as profile pictures for social media accounts have been widely used in information campaigns and we've seen other networks use these for profile pictures for example in this piece we worked on with the BBC to identify numerous accounts promoting Chinese government narratives around certain issues and again these same identification features were used such as the central location of the eyes and those mismatch features of the blending and blurring an analyst that regularly identifies networks using some of these AI generated faces is someone called conspiradores runs a substack called conspirator zero who I'll link to in the description below and they are often identifying information operation and campaigns where these campaigns use those images as their profile pictures for the accounts that push out certain narratives disinformation or attempt to influence specific subjects while we've been looking at the site this person does not exist there are also two other websites uh that are quite commonly used for this kind of work there is another version of this person does not exist this one allows you to select categories of faces so for example allows you to generate male faces of a specific age group of a specific ethnicity and generate faces in that means another website which is called generated. photos does this on a much larger scale and allows you to look at specific ethnicities and create faces based upon your selection there as well the second platform we're going to look at is dolly or doll e however you want to pronounce it and you can easily access that through using chat GPT uh which is found at open Ai and this is a specific AI program that's capable of generating detailed images from text based descriptions uh so you can generate any kind of image based on what you write in this text box down here so for example we could ask for an image of a dog sitting on the eye of London wearing a cape as you can see I've created something that is obviously not real and obviously doesn't look real but essentially is a way to create an image based on a simple string of text that I put in there if I was to try and create a photo of the London Eye On Fire you can see that the policies of GPT and Dolly stop that from being created so that you can see that the policies are already in place uh and are continuously growing to stop images being created that may be for nefarious or harmful purposes if I was to create a more practical example to try and create a realistic photo we can start to see the limitations of what Dolly has so say for example I want to create an image of a dog in front of London ey to make it look a little bit more realistic except this time I'm going to include a prompt to make this more realistic as you can see even asking Dolly to create a realistic image it still includes this kind of filter over the image similar to this image as well and many of Dolly's cre Creations or images have these kinds of almost cartoonish kind of filters added to give them that that kind of unnatural look as though it's not a real photo so what's a real world example where this platform's image has been used well if we have a look at news outlets where they may not have images to includ in their news articles about certain things then we may be able to find purposes there for example in this article from next shark which is about a former Facebook employee saying that Facebook's CEO Mark Zuckerberg was waving a katana sword around the office there's no imagery of that however these two images which if you look at them they're very poorly generated were used as the kind of illustration or imagery for this article because it didn't have any images but what we can see is that at the bottom of the article the featured image was created using Dolly mini by using Dolly we can also generate faces as well so similar to where we have this person does not exist which would be used for a profile picture that can also be created through Dolly however that would look completely different in comparison to what this person does not exist have so looking back on this person does not exist we can see that these are very realistic faces we have the centrality of the eyes for dolly when we generate faces again you can see what I mentioned with those previous images we have this cartoon style appearance which looks like a animation filter that's added over the top of many of doll's images moving on to the third platform we're going to look at Mid journey and mid journey is similar to Dolly in that it's an AI platform that generates or creates images from textual description the same as what we've just done before and it's often used for artistic purposes or generating interesting images editing images as well changing images adding color or or colorful Creations mid journey is impressive because it makes far more realistic images than what dolly is capable of uh and really showcases that diversity an AI generated image based on prompts so where you would add for example what we did in the dolly one where we asked it to create photo realistic or realistic images this would actually do that and would really listen or does really listen to the prompts that you provide or the commands that you give it with mid Journey the platform that it uses or that it that it communicates through is Discord specifically what you do is you communicate with the mid Journey bot in providing prompts and commands to that bot so in giving an example I'll hit my forward slash and you can see the the kind of things that you can ask the bot for example you can ask it to blend images uh you can ask it to write a prompt based on an image so you can drop an image in here and get it to write a prompt for you I'm going to hit imagine and I'm going to work on a prompt similarly to what we did with Dolly we're going to imagine a dog standing on the London Eye in London with a cape or wearing a cape except this time I'm I'm going to take it a step further and ask it to be a realistic image okay so we have our images that have been generated and we can see through these that they look extremely realistic in comparison to the ones that we created in Dolly it almost looks as if I've taken my dog popped that cape on it and taken a photo of it within the London Eye Which is far more realistic than the images that we created in Dolly there so you've seen how the platform Works what are some real world examples where some of these images have been used in disinformation campaigns in one case last year an image allegedly from mid Journey was reportedly showing a explosion near the Pentagon while there was quick verification to show that this indeed was not the case Emergency Services were still alarmed and individuals were still of the belief that there was indeed an explosion or a fire near the Pentagon however analysts quickly looked to see that in the background of the image some of the blurring was failed as we've seen in previous AI generated images as well and if we look at Elliot higgins's Twitter posts where he used mid journey to make pictures of trump getting arrested while waiting for Trump's arrest again this was not a real event we can see some of those details there around the hand the number of fingers as well as the faces in the background if you count some of the the fingers on the hands and if you look at the details of some of the hands like this person's hand right here if you look at some of the faces in the background you can see that they have failed to blur properly these are all common attributes of mid Journey images where hands really do not work too well unfortunately with images like this so when Elliot put these out even these images were then used by other accounts to allege or say that there was an arrest of trump however as you can see in the notes under here from the community notes people again reminded individuals that these are fake generated images while I've created a quick example in mid journey and you you can see these ones of Donald Trump there have been far more advanced developments by using mid journey and specifically harnessing the use of prompts to make photorealistic images one good place to find these is through Reddit where people are creating photo realistic images using mid journey and very much showing the full extent of how well they work however if we look at these images again we can see the subtle flaws that I spoke about before such as the hands and some of the faces so we can see this person's face this person's face but definitely the hands there where we have mismatch fingers we have mismatch hands and they appear to be in different shapes so while this image looks completely very close to realistic we can see those subtle flaws still being present there another way to identify the flaw within a mid Journey image is is the wording used in these images so this uh very realistic looking image is flawed by the fact that you can see the wording and the generation of this image has failed to actually make sense with the wording so you can see the the vehicles here the number plates the wording on these patches the wording on the side of this vehicle and even the street signs over here are not of any kind of language that's on this Earth and they they mismatch and and really don't make much sense and that comes with a lot of mid Journey images that the wording just never seems to work out moving on to the next one is the face app editor and this is actually an application for phone devices you can use this on both Apple and Android devices I have this on my mobile phone which I'll run through uh on the screen here and you can see the ways that these uh essentially allow you to edit faces to demonstrate how this works on the phone I'm using a iPhone and I'm going to use the example of one of the faces that we created using this person does not exist to edit this person's image so I can change the complexion of this person uh but I can also change their hairstyle as well to give her a bob cut I can add a level of makeup I can add dimples and then there are further things such as changing and adding the addition of a smile on that face as well you can also change the age of someone um to make them appear more youthful uh or or indeed to make them appear a little bit older so how have applications like this where real faces can be edited how have they been used in real world situations well we have an example from India which you can see on this BBC report here where the faces of two athletes who were detained after a protest where they were looking quite sad in their original image and then how this was uploaded to make all of the people that were facing the camera appear as smiling in order to make out that this was indeed a joke or or that they were happy and not actually serious about what they were protesting about the implications of this indicates that there is an attempt to mislead the public and by changing someone's facial expressions shows that they're conveying a different emotion in comparison to what they were conveying in this original image that they posted there you can spot these by having a look at the original context so if you have the original image again as the five tips that we went through at the start consider considering the context of the event which was considered here doing an image reverse search of that image and also questioning the account that originally uploaded uh both the original or the manipulated version and what sort of accuracy and Trust does that account have as you can see each of these platforms highlight different aspects of AI generated imagery from creating entirely new images to alter existing ones and by covering these this tutorial provides a comprehensive overview of just some of the platforms operating in the current landscape of AI image generation and manipulation as well as their implications in the realm of this information to go back through those five initial tips that we went through when we're looking at these images we first look closely at the image to have a look at things like hands face pieces then we look at second things in the background looking for blurred or pixelated elements or distorted Reflections third we do an image reverse search using things like Google or Yandex or teni fourth we verify the source sharing the content so checking the credibility of the account sharing it and then five and one of the most important ones actually verifying the event and the claim itself I hope you enjoyed this session and I hope you found it useful and I look forward to seeing you in the next [Music] session