https://twitter.com/TonightHere/status/1302016158049202176

More tips on how to defeat their image matching algo.

According to Q, DARPA's algorithm currently analyze 32 structural points on the image.

Let's make a few assumptions here:

It could mean something simple, such as:

32 structural points comes from an image evenly divided into 32 parts, to trigger a 20% mismatch, you need 32 * 0.2 = 6.4 area of mismatch sections.

Or it may be something more complex, such as:

32 structural points come from 1 + 2 + 4 + 9 + 16, where:

1 = A whole image structural match.

2 = half image structural match (split image into 2)

4 = split image into a 2x2 grid and compare each pieces.

9 = split image into a 3x3 grid and compare each pieces.

16 = split image into a 4x4 grid and compare each pieces.

Together they form 32 data points.

By distorting the 4 corners with small boxes, you get:

4 out of 16 mismatch on the 4x4 grid.

4 out of 9 mismatch on the 3x3 grid

The whole image, half image, and 2x2 grid does not have enough differences to trigger a mismatch due to boxes being too small.

In this case, you get 8 out of 32 structural points

A good 25%, above the 20% trigger, but it'll be defeated in a few days.

Generally you compare more than just patterns, you compare colors as well, but Q specifically mention "identifying structural point" so I am going with it.

TBC