just from my reading of papers, optimized patterns make use of the fact that detectors (even modern transformer based architectures) are capital S Sensitive to the sillouette of an object.
in my other comment i linked a paper where they fool modern detectors with 80+% success rate. the clothing looks like a tiedye shirt to me. not fashionable, but not drawing attention either. if there is a way that you can make this fashionable is not a question i can answer :p
from my reading combining detectors (ensembling) does not help. this leaves spies with a problem, how to detect these noisy people without losing performance on normal nonnoisy people. there are tricks to this, but they are limited at best, for the same reason the noise worked in the first place. here is the reason.
assume noise makers create noise with methods A,B,C. you train on a dataset with images from the different noise makers (you don’t know which method they used). each of these noisy groups will have a distict sillouette you can detect, but together, you are building a function like:
pineapple: fruit bannana: fruit apple: fruit
this is a solvable problem for ml, so how do the noise makers win ultimately? randomization in the process. if each noise is distinct there is no training, you only have one image per type of fruit, you might as well use traditional CV to detect these noisy people (good luck).
My opinion is i want a product where the customer provides a random seed which generates the process of noise generation for their shirt. the product is the transparency into how the seed effected the noise making, and an evaluation of the noise on modern off the shelf detectors.
obviously we don’t see that here, but it seems doable, maybe a business idea for me? haha
Feel like it may not be if it’s pointed straight up at all times. but if the plane is doing an aileron roll it would not take minimal damage, so star fox would be effected.