The privacy aspect does seem novel, and that paper evaluates the various aspects of privacy and the quality of perceptual hashes. However, anything concerning privacy (of client or server data) is orthogonal to the vision methods underpinning the system, which are well understood and are the primary cause of the post. For that matter, I don’t see how the paper’s proofs of privacy and/or formalization of the problem help with hash collisions. And that particular paper doesn’t use deep learning methods which are probably easier to attack with off-the-shelf methods.