Machine learning-based detection of information security violations in swarm robotic systems

Abstract

Intensive development of swarm robotic systems actualizes the need to ensure their information security. Known approaches to information protection of the collective decision-making process in swarm robotic systems use physical parameters that strongly depend on operating environment and hardware implementation of the system. Thus, it is difficult to identify universal indicators of abnormal behavior of an agent, that are able to provide accurate rejection threshold and low false positive rate. The aim of the work is to improve the efficiency of consensus achievement in swarm robotic systems in the presence of faulty or Byzantine robots. Detection of Byzantine robots is carried out by the use of machine learning methods. To classify the robots as normal or Byzantine, we used an artificial neural network trained on a dataset generated with a previously developed analytical method. The novelty of the proposed solution lies in the choice of parameters for carrying out simulations in order to form a dataset for training the classifier. Simulation of a swarm consisting of 100 robots has been carried out. In the presence of 20 % of robots with incorrect behavior, the number of false positives is reduced by 41,07 % relative to the prototype method. The proposed approach is capable of detecting Byzantine robots regardless of their number or behavioral strategy. The method is implemented as a C++ program.