OPTIMIZATION OF NEURAL NETWORK COMPUTATION USING INTEGER ARITHMETIC

Abstract

Most of the sensor devices in the Internet of Things systems are based on energy-efficient microcontrollers, the computing resources of which are limited, as well as the amount of available memory. Increasing the security of the use of such devices with the help of neural networks is an important and urgent problem. The article describes the possibility of using artificial neural networks in small microcontrollers with limited resources. The purpose of this work is to check the possibility of calculating neural networks based on integer arithmetic to reduce the time of calculating a neural network and eliminate data normalization operations, as well as to evaluate the feasibility of using such neural networks in the field of security of the Internet of Things in comparison with traditional methods, such as black lists and white lists. The following results were obtained: when switching to integer arithmetic, compared to floating point, the accuracy of the result calculations is within the permissible error of neural network training, that is, it has not changed. Execution time decreased by 30-96%, depending on the architecture of the microcontroller. The program size is reduced by 22-48%, also depending on the microcontroller architecture. Conclusions: the possibility and expediency of using neural networks optimized for microcontrollers with limited resources was proved. This will increase the security of Internet of Things systems, especially against device authentication threats and intrusion detection. Prospects for further research are determined.