>>10691359 (LB)
The Mickey Mouse clock my simply be too distinctive and too easy to detect. It's a circle on a white field and with a distinctive blob shape in the middle.
We should test the memescramble (and other tricks) with realistic memes.
>so we are currenty at the point of diminishing returns to keep coming up with MORE ideas that defeat Twitters sniffer
i disagree.
Me, a novice level coder, can bang out a line of Imagemagick commands in 5 minutes that beat their system. It probably took a small team of real coders and image recognition experts many many hours to develop and deploy a countermeasure.
It's like a state-of-the-art CWIS system to shoot down a soviet-era rocket. It is defiantly possible, bot only at much greater expense. Similarly, they need much greater effort and sophistication than we need.
In this cat and mouse game, the advantage is ours.
I think Twitter converts everything to jpeg.
Nope, nevermind. I just found a PNG.
I added random block size and random rotation to make it harder for their sniffer to look for color-block transition at predictable locations.
https://pastebin.com/mTRfeaJ8
>'stochastic dithering'.
That would require much smaller blocks. All they have to do is filter out the high-frequency component (undither it) and it's back close to the original image.
I'm trying to add both high and low frequency variations via transition edges to wreck the algo while still being easy for the human brain to look past it.