>>10598632

I suspect DARPA, like other alphabet orgs, has rogue elements. That said, it is extremely concerning tech companies have classified tech. DARPA is OCA, someone there chose to give them it, legally or illegally. If they never terminated connection with fb/LL, made it unack, that could be a conduit.

Is current leadership there at least clean? From the DARPA twitter, average tenure was something like 4 years. That is suspiciously low for DoD. Are they grooming clowns to send out to tech companies? Hard to believe their employee retention is weaker than Chinese corncrete!

>>10598710

Sounds like this clown contracted for them to develop tools for a different purpose, and absconded with them. The devil is in the details with tech; once the lessons are learned, it is far easier to recreate so long as you bring over the same team. Hence why China steals everything, and poaches talent. The hard part is already done.

>>10598827

That could work, but it might help to make it non-rectangular. If I had to guess, they are using hough transform like algorithms over areas, calculating residues and doing feature extraction, then running that set through ML toolset with the criteria Q mentioned, tuned to a specific Pd/Pfa. ML tools work generally in feature space. 4 days is also a reasonable estimate to retrain such a ML tool over the vast dataset they must have.

Perhaps skew the image? Make it polygonal? More than 4 sides. Add general waves and distortions to it spatially and with color map.

Am I close Q, to how they're doing it? Does general nonlinearity break their feature calculation?

>>10598913

Addendum:

https://en.m.wikipedia.org/wiki/Feature_(machine_learning)

Identifying structural points = feature?

Do they have a hierarchy of ML tools using these 32 features to narrow images down to buckets?

Do they then do binary classification against all members in a bucket, if it matches over 80% (threshold), it is filtered?

Do they crop windows from the original image and compare those? I'm assuming they can detect compound images, borders, etc.

If they're using ML tools, the more we stay outside of the training area (sparse region), the more it screws with binary classification. More importantly, we need to develop a continuously evolving set of anti-meme-filter filters and keep our usage diverse and ever changing. Can't hit a moving target. I'll work on some tonight.

>>10598675

>>10599016

Brilliant! Many of these would surely screw up their algorithms. Can they be combined in an additive manner?

>>10599155

Looking into it now, thanks fren. A meme summary about these techniques to avoid censorship should be made. Any seasoned meme makers up for this? The more the ability to use a wide variety of techniques is spread, the less effective their clown tech will be.

>>10599208

I figured something like this would work. The more nonlinear the altered meme is from the original, the harder classification will be. It's probably the shear transform that had the biggest effect. Good work, fren.