A robust hashing of ID photos

The paper proposes an image hashing method that is highly robust to a series of currently encountered attacks on digital images. It is dedicated to images representing ID photos. The extraction of hash value is based on a global method for face recognition. Such approaches need a registration stage which is often obtained by means of face characteristic points. The detection of the characteristic points easily fails in the case of attacks like filtering, noise addition, random line deletion or jpg compression, mainly when they overcome certain limits. Our solution is to attach the characteristic points to the final hash, in order to obtain increased robustness to such attacks. The final hash is obtained by stabilizing an intermediary hash extracted by using Independent Component Analysis. The stabilization is done by clustering. The final hash is given by the cluster label having attached the coordinates of three characteristic points. We show that intermediate hashes of the attacked image approximate a hyper-sphere centered on the original image hash. Due to this fact, the minimization of a cost function representing the proximity to original hashes is sufficient to stabilize the intermediate hashes. We also evaluate the system capacity by estimating the number of freedom degrees for the intermediate hashes of the non-attacked ID photos.