Secure-design artificial intelligence: the advantages of replacing classical artificial neurons with neurons of geometric mean and harmonic mean

Abstract

Background. Responsible artificial intelligence applications must run in a secure execution. One of the directions for creating such applications is the use of large artificial neural networks, for example, consisting of perceptrons automatically trained in accordance with State Standart R 52633.5. The purpose of the work is to replace classical perceptrons with artificial neurons of the geometric mean and harmonic mean. Materials and methods. Perceptrons enrich biometric data through their accumulation in linear space, that is, they can be considered as artificial neurons of the arithmetic mean. Artificial chi-square neurons and Mahalonobis neurons accumulate data in the space of the standard deviation. The transition to other classes of artificial neurons by formal replacement of the data accumulation space with the geometric mean and harmonic mean is considered. Results. The advantage of the geometric mean and harmonic mean neurons is that their software implementation does not require preliminary calculation of the first statistical moments of small samples of biometric data. The latter removes the problem of hiding the statistical moments of the “Friend” image through encryption of the tables’ connections and weight coefficients of the artificial neurons network. Conclusions. The use of artificial neurons of the geometric mean and harmonic mean is a promising direction for creating neural network decision rules of artificial intelligence in a performance protected from knowledge extraction.

Keywords

• artificial neurons
• harmonic mean neurons
• geometric mean neurons
• artificial intelligence protection