Read on Twitter
Your eyes can deceive you, but hash functions won't lie.
While working on the background image classifier for the http://thehashnames.com , I noticed an anomaly in the data.
Most images were classified and put into buckets correctly. But there were also many outliers that created unique fingerprints.
So I took a look at the data...
Most images were classified and put into buckets correctly. But there were also many outliers that created unique fingerprints.
So I took a look at the data...
Something caused images with the exact same background to create completely unique fingerprints. So I rendered them for debugging. They all had one thing in common: they belonged to rare special images with hidden traits or unique masks
The image classifier is very basic: it will take a sample region from the image and hash all rgba pixel values in the data buffer to derive a fingerprint.
But somehow, the mere presence of a phoenix or halo in the image had an effect on all the pixel values in the image
But somehow, the mere presence of a phoenix or halo in the image had an effect on all the pixel values in the image
Looking at the individual pixel values confirmed the theory.
It seemed as if beside the image there was a separate hidden data layer that was placed as a bit-MASK on the pixels
It seemed as if beside the image there was a separate hidden data layer that was placed as a bit-MASK on the pixels
This can have two main reasons:
1.) the image uses some kind of steganography to insert a payload
2.) the hidden item image layer occupies the whole image region
1.) the image uses some kind of steganography to insert a payload
2.) the hidden item image layer occupies the whole image region
Another thing that became apparent was that egyptian hieroglyphs, greek letters and some other symbols would not cause this effect. Which means that either the process for placing them on the image was different or that they indeed create a complete different item category.
Hope we can crack this puzzle together. Happy hunting!
