Adaptive Evolutionary Neural Networks for Resilient Anomaly Detection in Decentralized Swarm Robotic Communication Systems

This paper presents EvoCommNet, an adaptive evolutionary neural network designed specifically to find abnormalities in response to communication dynamics of robotic swarms inspired by foraging and exploration behaviours. EvoCommNet employs an evolutionary neural network to detect abnormalities in the swarm communication which might be indicative of problem with the system. Thus, through these evolutionary process, EvoCommNet optimises its structure to effectively detect anomalous behaviour with low computational requirements that will make the solution ideal for real time monitoring in resource limited environments. Designed and trained with PyTorch and tested on one unified swarm communication dataset, the vulnerability of the model to dynamic swarm communication patterns is evidenced and, therefore, the model should be useful and necessary in areas where communication continuity and proper abnormality detection are necessary for mission accomplishment.