There is a series of lectures on stego at Oxford [1]. The lecturer uses simple machine learning (average perceptrons I think) and achieves 99% accuracy with little effort although that was for LSBR[2]. Later he discusses problems with the graph approach [3]. Simply put whilst it preserves 1st order histograms it will probably mess with 2nd or 3rd order histograms.
[1]https://www.cs.ox.ac.uk/teaching/courses/advsec/ [2]https://www.cs.ox.ac.uk/teaching/materials15-16/advsec/advse... p48 [3] ^ p58