Privacy preserving neural networks in iris signature feature extraction

The iris signature is used in "Iridology", for person identification and calculation of torsional eye movements in video-oculography. Iris data are expensive and difficult to be acquired and their amount plays an important role in recognition accuracy when data mining methods are used. However, privacy issues often restrict open exchange of data between stakeholders. The present article presents a privacy-preserving neural network protocol, for horizontally-partitioned datasets, i.e. datasets that share common attributes but contain different records at each party. The proposed protocol assumes a malicious user model and does not use homomorphic cryptographic methods which are inherently only suited for a semi-trusted user environment. The performance analysis shows that the communication overhead is low enough to warrant its use while the computational complexity is identical in most cases with the centralized computation scenario (e.g. a trusted third party). The accuracy of the output model is only marginally subpar to a centralized computation on the union of all datasets. Another important aim of this work is to search proper choice of the part of the iris signature for person identification and calculating torsional eye movements. Also, estimate changes of the iris contour, sections of the iris and elements of the iris signature. The mathematical model of formation of the iris image on the plane was compared with the real image of the iris.