... Of course the counter-counter attack would be ...
I'm not even thinking about the political side of all of this (and maybe I should). I just like to think about solving cool problems, such as how to detect these images if software that modifies pixels is used.
Thus far the discussion has been about one-way hash functions designed primarily for cryptographic applications.
You're making it unnecessarily hard on the detector, who, after all, gets to design the detection system.
The more likely candidate would be a hash function designed for the purpose, such as a geometric hash: "Used in visual recognition for classifying parameter objects within an associative container such as a hash-table."
This class of functions has been defined in the DRM context; see: Efficient Near-duplicate Detection and Sub-image Retrieval
On Hash Functions