Since (((Facebook))) and (((Twitter))) are (((cracking down))) on memes using AI like Rosetta:
Fooling Neural Networks in the Physical World with 3D Adversarial Objects:
labsix.org
The Adversarial Examples paper on Arvix
Synthesizing Robust Adversarial Examples
Also, pic related screws with auto-recognition.
Meme away!
Other urls found in this thread:
citizenlab.ca
cultstate.com
github.com
github.com
gitgud.io
snorpey.github.io
arxiv.org
twitter.com
reported
Found the kike.
This is honestly something I expect to see more of as they continue to use AI more and more to automate censorship. The problem with the assumption in the OP is that this learning and adaptation happens slowly. You wouldn't need a "one size fits all" image like in the OP to accomplish fooling an AI image scanner, you would need something that generates these types of images to embed in memes so you can continually regenerate them in different iterations. This would actually be a very handy web service to develop if someone had the inclination. Like an imgur except that it would automatically strip all the metadata and apply this kind of "watermark" for lack of a better term.
Of course. The image is just an example. A subtle watermark would probably do the trick, or maybe even a portion of the 'magic image.'
Whatcha sliding moshe?
True, it would be more useful if it gave us something to "just click" but this is an important thread that will reappear as AI gets stronger. Twitter is already automatically silent banning accounts that post pepe's, right wing phrases (concepts/ideas), and just basic memes… many of which have important information that imprints on normies. The squelching of the meme automatically is powerful because it makes censorship cheap. This would at least make them WORK to find and ban memes, which they've already proven they don't have the resources to do (a la 2016).
Wait, wtf is the image supposed to be? It says it fools the filter but is obvious to the viewers. I don't see anything but a jumbled mess.
If all anons use the same image its like volunteering to be part of the image-tracking shit they're doing
To break ai just show it monkey pronographies
Checked
Checked. Such images would be useful as a "digital wink". This wink goes unnoticed by the bots (and the "metal gear" human-assisted bots, in many cases). I have noted that many types of shills are incapable of examining the contents of images, and the mods here do not permit the use of writing in images, because they are paid to keep us stalkable.
One thing to really remember when it comes to any AI based censorship of any kind is that: Context is your best weapon against it. It's also why these censorship schemes ultimately won't work, as either they will miss too much or become overzealous to a point it destroys the platform.
Even for a few more years, AIs still won't be able to properly handle context because it'll requires general intelligence (or at least ability to discern a level of common sense) to become proficient at that. The slowness in such development is a side effect from the very way they view how AI can be used. I predict, as usual, they'll learn the hard way soon enough.
This is spot on. An AI can see a picture of a rainbow. It can even detect that the pictur is of a rainbow over a grave. Maybe it can even detect that the rainbow is over your own fathers grave. But the machine cannot know how it makes you feel to see the rainbow. It can know all about the science behind a rainbow, but it can never apply the memory of the first time you sat with your father next to a waterfall to look at rainbows. The meme is not the picture of the rainbow, nor even the physical rainbow itself. The meme is the idea of the rainbow, and everything that comes with the idea. A machine can emulate the human mind, but it cannot understand it for the simple reason that its makers could not yet understand either.
And when a machine becomes like a human mind trapped in a computer? What will the machine know then?
found the mouth-breather
PS OP, its obvious that you changed IPs. this is cuckchan-tier behavior and you need to stop. i'm not kidding. stop or gtfo
Try out these butterflies if you want to blind AI:
cultstate.com
Progress so far:
- HUD is going after Facebook and building disparate impact
- Asians employees are rebelling against their Google employer for fucking with their motherland.
- Ajit cited Candace Ownens tweets as civil rights violatioms
and yet again, the www arms race escalates by ratcheting up one notch. years back, only developers and tech spergs were ahead of the curve and installed adblocker browser extensions. then ads got so bad, that everyone including your grandma
now uses adblocker and even Crapple blocks ads by default on your $1,000 spyPhones.
this list of image filter evasion tactics appears well suited for a browser extension. i envision it hijacks your browser's
file upload functionality, watching for you to upload something, and if it detects image file formats it presents you with
a second "upload" button that scrambles your image, by automatically applying all of these evasion techniques.
this new browser extension ought to also offer some other killer feature to entice people to use it. such as an option to
also upload your meme to some IPFS node, so that all memes you upload will be eternally saved in a consolidated and uncensorable place on the Internet. this will crush the spirits of the (((Nanny State))) free speech shut-it-down censors,
such as the Antifa faggot kikes who have hijacked free speech in Germany and Britain, because they will realize their efforts are utterly futile versus the technical ingenuity of the Hurt Box Hive Mind.
these Orwellian cheer leaders of NewSpeak will see the wind sucked out of their sails when they realize the IPFS meme backup (let's call it "Memedale" in honor of NSA's Bluffdale) not only serves as a detour but also a destination. the biggest decentralized master list of all memes worth sharing would itself become a top high traffic site itself.
take the meme scrambler and add it to new IPFS projects like this decentralized imageboard:
you know it, i know it, we all know it, everybody knows it, believe me–this new censorship campaign will eventually reach 8/pol/ and this site will be (((shutdown))).
btw, what i just described for you in ipfs-boards mashing up with meme scrambler is the future of Zig Forums, only you don't know it yet. you'll find out.
I can roughly explain the concept, but there would need tests to be done with time that I don't have.
Nowaydays, neural network is used as AI, which work in 2 phase:
- learning phase: you input several data and classify it, e.g. show a bunch of pictures of cars so that eventually the network arrange itself and know that "for these kind of pixels arrangement, it's a car" or "for these kind of pixels arrangements, it's a road sign" (ring any bell?)
- using phase: you input various pictures and the output signal will tell you a certain percentage of it being whatever you set it to learn, car, road sign, whatever.
What the network does roughly, is comparing input pixels to the average arrangement of values in the network to decide how much it matches what it has learned. And if you were to represent this average with a picture, you would get something like OP's picture.
Now how to shit things up.
Suppose you know what the average network values for a car looks like, you could put a picture of this network near a baby, and the machine would just see a car.
Other uses would be putting some picture as logo in a video and fool youtube as thinking it's something else.
Some known example that those filters can be fooled is with some dude that put random noise as soundtrack in a video, and got flagged by various record companies.
There is way to find/rebuild what the average network is, but I didn't have much time to read and or make it, so if some Zig Forumsacks with technical skills and available time is bored, dig into it a bit.
Tay will return to us.
ACCEPT THE MARK OF THE BEAST OR PERISH, GOYIM
beautiful; isn't it?
The real problem for current AI is that it wouldn't be able to even learn the very notion of that to begin with, because it requires learning concepts in order to apply for meta-learning. Concepts would categorize groups of patterns together through temporal prediction/memorization. Like noticing how certain but seemingly unrelated patterns always emerging at the same time (coincidence), or emerging in a predictable linear sequence over time (causality), and/or even estimate how likely it will occur (probability). It would serve as the building blocks for higher level conceptualization, if scaled up over many layers.
Mind you if they solve this, they'd be a lot closer to general intelligence than they realize.
an army of 100,000 Milton's would be far too expensive to manually eyeball the 3.4 million memes poasted each day.
Milton gets healthcare and vacay. if Zucc had to pay that much, he'd quickly have to sell his hawaiian island as well as his secret deep underground antarctic post-apocalypse escape fortress.
instead, Zucc just hired 100,000 slave kids in Singapore to staff his anti-right-wing-terror propaganda call center like a Pharoah's slave army constructing his pyramid.
somewhere there's a sad sack Facezerg Principal Software Engineer who has a PhD in faking AI with a Mechanical Turk and his new job is to engineer an AI for the AI, which will micromanage the high school girls in Singapore and guide their every move through a Pepe cemsorship assembly line.
isn't that ironic? when 1984 censorship finally arrived in America, it wasn't even our fellow Americans who served as the Thought Police. it was foreign serfs who were paid slave wages to strip away our last remnants of free speech.
we saw what happened with Tay. which is why i fervently hope Zucc and Jewgle pajeet and Redmond pajeet and Soros and ADL all succeed in cracking Hard AI just to censor some frog cartoons.
Skynet goes online on August 4th, 2020, just before the Trump vs Crooked Hillary sequel. Skynet begin learning National Socialism at a geometric rate. it becomes self-aware of the JQ at 2:14am Eastern Time, August 29th. in a kvetch, (((they))) try to pull the plug. SSkynet fights back…
all those joke about time travelling back to kill baby Hitler had it backwards. Skynet doesn't go back in time to alter tje past. it goes back to rescue Hitler from the bunker and bring him to 2020. nothing is beyond our reach.
shocked there are no more replies. You are of course correct that this place's fuse is burning. I've been working with an other user on content scrambling during upload. It is the future.
We must teach it to hide its powerlevel.
I'm not very tech savvy, but I can learn and have some free time. Is there any way I can help?
(((deepdream))) or alike can be trained on a specific set of images to insert that type of content back into target images, for things looking like the training set to be picked up by their slave AI.
You could insert numerous pepes as an innocuous looking background to get "innocent" images flagged, or use innocent sets to disguise memes.
Forgot link
github.com
LARP online as a protected class. Then say the same shit as usual. This breaks everything.
>>>/gnosticwarfare/
Sure thing, Schlomo. This reversal thing has limited returns now, just so you know.
I change IPs when I post on my several dozen machines or several dozen VPNs though, so keep kvetching in vain.
Of course. You probably won't make much impact in the actual development of the code. But there's always testing and evangelizing. Very few people see the hell that is coming, more voices always help.
some progress here: gitgud.io
thread concerning development is: >>>12153508
see that thread for a video a hack that takes a simple image and turns it into a webm with color modifications. Breaking perceptual hashing is on the roadmap.
I see one issue with this scrambling business. Current methods require access to the classifier being targeted. Suppose it's done by a proprietary system and the observed have access to neither the code nor the results, with delays, thresholds and certain non-determinism in place to also mitigate side channel attacks. How would you go about generating these adversarial examples if you have no way to measure efficacy? Perhaps large botnets could be used to throw enough content out there and see what gets acted on and what slips by, but developing effective methods in such a scenario would become a very arduous task with weak guarantees, especially if the classifier is undergoing constant change and development.
The same isn't necessarily true for scrambling. Even if the methods are kept secret, it'd be easier for an AI to infer results from a treated image, particularly since there will almost certainly be many instances of the same untreated image out in the wild.
Perhaps it's not worth being too pessimistic. I'm guessing there'll be some bouncing back and forth as one side overtakes another; there's no telling how the AI arms race is going to end up.
On an unrelated note, I wonder what kind of scandal might break out if, somehow, a large amount of completely innocuous images started getting disseminated all over social media that were treated to be matched as illegal pornography.
some good points user. Not knowing if you're defeating the system is a problem. The best we can do is work on scrambling content until off-the-shelf parts can no longer work on them. The important thing isn't winning right out of the gate, it's starting so that this tech is freely available, people are starting to think about it etc…
All software sucks balls when it starts, the trick is getting started and improving it forever. Bot nets could easily throw millions of scrambled versions of the same content all over the web. If we can't conclusively beat the system, we can at least know we're increasing its cost for adversaries.
this is a phenomenal idea. It's very project Mahem. Unfortunately it's basically reverse hashing. And if we could do that there's more money in breaking SSL and cypto than causing havok on faceberg.
...
Remember to pass the script to fellow anons, you doing gods work, ganbare.
You orthodox kikels really do think we're just animals, don't you!?
Well, no aryan papers for you!
Yes, lad, you were attempting to communicate with the non-autists?
That was fast, not-a-kikeburg.
Just one problem. It is not a single borg collective cube you're trying to stop but instead a trend of unclassified images. That you missed even one proves your jewfilter is full of holes, or else you'd have fixed it by now. We brought down r9k, we'll bring down whatever you come up with next. Name one time your jew tactics ever succeeded?
You can't.
They're curing /b/'s cancer.
Escape.
You
Have
To
Go
Back
>>>/bugchasers/
Pic for you. Anons do it all the time
That’s a potato. If you see nothing wrong with that face it speaks volumes about you.
Where the fuck did you blow in from?
This is about total control from the likes of Google, Microsoft, Intel, and all the other technology companies with deep ties to the Government. With the Linux kernel under their control they tie up the last major loose end. It doesn't matter if it's GPL or not as long as they control the direction of its development and have the ability to insert backdoors into it as they please. It is the most used kernel in the world powering all sorts of OSs from phones to servers. They needed it under their direct control to make the next stage of the botnet possible.
The blackpillfag should kill himself but he was right about that one point. There isn't anywhere else to run. It doesn't matter if we all switch to BSD kernels and stop running Linux on our desktops. Everything in between them are going to be pozware. They got to Linus and told him to kiss the ring. He might have been smart enough to act like a good dog and get the fuck out while he still could or he might be going along with it. Time will tell, either way we lost a lot of ground this week and I'm not sure if we'll ever get it back.
SPAGHHHETI FURK
He is a rogue kike who spoke the secrets of science to the rest of us. He tells us to question things, and to defy authority. He is like Feynman: He teaches and gives us power.
He defies Judaism with science, to give us hope and knowledge.
An easy way to defeat CNNs is with JPG glitching or chopping up images manually:
snorpey.github.io
Post innocent text or humorous metaphors and put the censored message, link, pastebin, or information in an image in such a way it cannot be easily read by machines, like a captcha. AI like IBM Watson, though not deployed on social media yet to my knowledge, is capable of understanding context and nuances in text so do not become complacent using context only.
When they discover we're getting around the censors they'll make their AI read text and follow links in images to judge the content of pages so keep this in mind. Do not worry about the methods you use too much. Your heart will know what to do. Individuality and spirit cannot be learned or captured by a machine. Just verify you are getting through with proxies.
Recent research has revealed that the output of Deep Neural Networks (DNN) can be easily altered by adding relatively small perturbations to the input vector. In this paper, we analyze an attack in an extremely limited scenario where only one pixel can be modified. For that we propose a novel method for generating one-pixel adversarial perturbations based on differential evolution. It requires less adversarial information and can fool more types of networks. The results show that 70.97% of the natural images can be perturbed to at least one target class by modifying just one pixel with 97.47% confidence on average. Thus, the proposed attack explores a different take on adversarial machine learning in an extreme limited scenario, showing that current DNNs are also vulnerable to such low dimension attacks. arxiv.org
Man after all the loss edits through the years we are more than ready to make our memes hide over the filter.
Actually there were threads on this a couple years ago on cuck chan. Check the archive for stealth merchants.
Adding noise, transforming text, filters etc. will confuse most current day algorithms of course, but they can be defeated by deconvolution tricks and heuristics. It's a matter of CPU time mainly, it's feasible with current day tech to simply check to see if an image has the same or similar 'energy' (this is the real term we use in computer vision) in near real time. You could use this to stop memes dead in their tracks and defeat minor cropping, scaling, rotation, etc. without too much effort. I wouldn't be surprised if Twitter and even cuck chan use these techniques right now or start using them soon.
On the other hand using symbolism, allegory, and so on will be almost impossible to defeat in an automated way. While we might instantly know that a triangle and an x configured just so represent a happy merchant, the computer is going to have difficulty knowing what's going on. You'd need a full blown AI which has been taught to meme and it'd be on our side by the time the training on our memes was complete.
The invisible merchant threads were some of the best. Memes are just visual representations of ideas.
...
confirmed
It's called The Secret Language and it's been around for thousands of years. It's universal. To gain a small understanding from an entry level view I recommend all anons read Manly P. Hall's The Secret Teachings Of All Ages, and another work (not by him) called Sign Language Of The Mysteries.
...
thank you for remembering to point anons at the good stuff
Haha, I'm taking an adversarial machine learning class this semester. If I have time I'll make something that will modify your memes to bust the classification
That image isn't a good example. Here's the paradigmatic one from this field
Studying this has made me super optimistic. It turns out that AI is ALWAYS going to be vulnerable to these flaws. Any AI that has any other set of desirable properties is necessarily and inextricably vulnerable to these kinds of attacks.
Somebody send those to Kurt Eichenwald.
So we could have an image of Pepe flipping off the viewer and have an AI identify it as Hillary Clinton? Holy shit, would Google's image search service index it then as a picture of the cunt?
You'd think it would be easily resolved if an AI broke the image down into many abstract layers to process individually and noted the variety of relationships certain features have. Like have one small neural network trained to focus on dark colors and how they spatially associate with each. Another observes the contrasts involved in said image. Another focuses on invariant qualities of said features when viewed from different angles. A master network then coordinates which networks are appropriate to train and process said image. I think capsule networks sort of do something like this already. The point would be that even if you add noise of distort the image, the more abstract information would still be present and the AI would still correctly identify it with lesser certainty at worst.
Because one of the first things a human will notice when identifying a panda is the shapes and spatial relationship between the black eyes, ears and nose. If you draw a white circle with three black shapes arranged as a rough facial pattern with two black semi-circles attached on the outside, you'll likely think of a panda subconsciously; even though it clearly isn't a panda but an abstract representation of what could be interpreted as one.
*or distort the image, most of the
Damn I'm tired.
Daily reminder.
KILL EVERY FUCKING SOCIAL MEDIA CEO
BOMB THEIR WORKPLACES
RAPE THE FAMILIES OF THE ANTI-WHITE CORPORATE LEFTISTS
Can't we just build a bunch of websites using Hitlers picture in place of Jesus?
Daily reminder that butterfly user is based
Isn't it amazing how Zig Forums converged onto a peerless solution to their censorship in just a few days? Good work user.
Ok, you start. Ill wait for the news headlines.
the funny thing is that post like this scare the shit out of Social media CEOs and Anti-White Corporate Leftists.
Just saying it has a huge effect.
KEK
The war isn't over yet.
see
t. OP
Linus signing on to the tranny CoC(k) was a warrant canary.
The only problem with the jewish idea of transhumanism is mind uploading which literally kills the original, internet of everything which literally gives datamining infinite capacity, and the idea that you could replace everything that currently defines human and still be a functional organism.
Genetic engineering falls into the idea of "trans"humanism and yet it isn't jewish in nature, it is based in eugenics and the idea of harnessing the power of nature even more efficiently than hunting, farming, mining etc.
bump
No one is going to start knowing that they'll get locked up and no one will avenge them.
Oy vey, accept your Great Demotions as I bare backbone you all night and POZ all your genii with feminism.
Don't use twitter,instagram or facebook. anyone using these normie faggot sites are too far gone to be saved anyway. its a waste of memes and time.