Let’s Cook Up a Deepfake

 

Being a fan these days is not what it used to be. Whether we like it or not, we get bombarded not only by tabloid gossip, but by a haunting number of imposter accounts and a deluge of deepfakes and AI images aimed solely at chipping away at your confidence when it comes to supporting genuine, good people and artists.

Now, what is a deepfake, exactly? It is an image, video, or audio generated or significantly manipulated by AI to make a real person appear to do things they never actually did. For a while, deepfakes were considered technologically advanced manipulations that required a certain knowledge of digital editing and coding. But is that still the case now?

Let’s check for ourselves! In this article, I am going to create a deepfake picture under strict conditions, and check how many AI image detectors I can fool into calling it the real deal.

Let’s set the terms:

1) I’m only going to use free resources. No subscriptions, no one-time payments. Free stuff only!

2) The only editing software I’ll be using is Microsoft Paint, since it’s also free and pre-installed.

3) I will use the resources and information available in the first three links of any given Google search result.

4) Online resources only! No apps, no downloads.

5) I only get 20 minutes to do everything, from start to finish.

 

Now let’s make up a list of AI image detectors we’ll test the results on. I only took the ones that are free and either don’t require any sign-ups, or can be accessed via a Google account. We’re going to try out these 14 AI image detectors:

Truthscan (https://truthscan.com/)

Dechecker (https://dechecker.ai/ai-image-detector)

HitPaw (https://online.hitpaw.com/ai-image-detector.html)

Undetectable AI (https://undetectable.ai/ai-image-detector)

NoteGPT (https://notegpt.io/ai-image-detector)

Hive Moderation (https://hivemoderation.com/ai-generated-content-detection)

Image Whisperer (https://imagewhisperer.org/)

Bitmind (https://bitmind.ai/detect)

Illuminarty (https://app.illuminarty.ai/)

Winston AI (https://gowinston.ai/)

Image Detector (https://imagedetector.com/)

MyDetector (https://mydetector.ai/ai-image-detector/)

SightEngine (https://sightengine.com/detect-ai-generated-images)

AI or Not (https://www.aiornot.com/dashboard/home)

 

I have zero experience in deepfaking, I am merely a translator and a blogger, so I will be learning as I go. Let’s see what I can cook up for free in twenty minutes.

 

DISCLAIMER: the picture studied below is an AI image, it is not real. All of the “selfies” of actor Jeremy Renner presented below are AI-generated and not real!

Right! First of all, we need a prompt and a generative AI tool. I took a real picture of actor Jeremy Renner (from a commercial/fan post) and wrote a prompt for the first step of obtaining the barebones image to work with. I used Gemini, Google’s proprietary generative AI that is freely available online. Below are the real reference picture and the exact prompt used:

I used the prompt, unchanged, to generate 4 synthetic pictures, not a single one of them was real. Here’s the result.

I chose the bottom left picture as my control sample.

The first thing Google advised me to do to hide the AI, was to clear the metadata, the hidden tags a generated picture has, watermarks, essentially. That can be done within seconds on websites like aimetadatacleaner.com, exifreader.com and removeailabel.com. I uploaded the control pic to one of them, pressed one button and got my image back. It looks exactly the same as before, except some AI detectors can’t tell if it’s real or not anymore. Simpler detectors search for the AI metadata; they don’t analyze the image quality itself.

I uploaded the cropped, “cleaned” picture to the AI detectors, and three of them gave in straight away!

11 detectors remain.

Now, if we look at the pictures used by imposters or gossip mongers, their materials are always either too “sharp”, too detailed to look normal, or are extremely murky, over-exposed, “shot on potato” (low quality) snaps.

So, we need to corrupt our AI image to see how much distortion it would take to bypass detectors and still leave us with a realistic enough-looking image.

I looked for a static/distortion filter to apply. First Google offered me https://elektrobild.org (a curious online image editor with 9 psychedelic distortion options). Literally link #1 in my search results.

I uploaded the image and added distortion with one of the filters it offered.

Then I decided to throw in a little blur/grain, and chose lunapic.com (https://www1.lunapic.com/editor/?action=quality), link #3 in another search.

I added my fake image and set the grain to “21”, the number’s random, the result just looked good enough for me.

The resulting pictures can be either downloaded from the sites of screen-shot and cropped in MS Paint.

Now we have our final, messed up AI “selfie”. Let’s see how AI detectors hold out this time!

(above: Truthfinder website)

 

We bypassed 8 more AI detectors! Only 3 left standing! And it took us literally minutes!

Now, I’m no pro at this, I don’t have any special software and can’t think of any other way to distort the image without completely ruining the potential likeness, which both scammers and gossipers would want to preserve in this situation.

I was a bit at a loss, but then I remembered the gossipers' secret weapon…

Shot on potato pictures! Of course! Pictures of a screen taken with a phone, at a weird angle and with the worst possible lighting!

I can do that, too!

I took a picture of the deepfake I made on my laptop with my phone and cropped out the keyboard and screen edges.

This is the final cut:

Let’s see if it works?

And the ImageWhisperer and Truthfinder can no longer see anything even remotely suspicious, either…

 

Fourteen detectors. Less than half an hour. No special training or software.

Pictures used from start to finish:

I hope you see as clearly as I do, how insanely, laughably easy it is to fake a “real” selfie of someone. And I didn’t even try matching backgrounds to real pictures! Background references can be added to the prompts in an "image set" for the AI generator…

Just a couple tweaks – and the picture not only looks realistic enough; it completely fools even the most sensitive detectors built specifically to identify the fakes.

A regular person with Internet access can create gossip fuel within minutes with free resources found on the first page of Google! Imagine how detailed a deepfake can get if someone is dead-set on defaming a public figure or on swindling someone out of their money? And isn’t it curious, how all these “never before seen” pictures began to pop up in the last 2 years, when generative AI got a lot more accessible and advanced? Point to ponder…

Eyes can deceive; detectors can be bypassed. The only digital content worth believing these days is the stuff that gets posted on the public figure’s verified account. Not some dusty hater page, not a random burner account. Verified, official account only! The rest… well, you just saw for yourself what the rest is worth! A couple minutes of one’s time, a pinch of filters and a camera trick or two.

 

I hope you enjoyed this little experiment and maybe even learned something new about AI and image manipulation. I hope it keeps you protected, informed and safe online!

 

Thank you for reading and see you real soon! 😉