Memory-based reservoir computing for synchronization and its applications

This paper proposes a memory-based reservoir computing (RC) architecture that incorporates an explicit linear autoregressive memory mechanism, designed to naturally align with the hereditary properties of complex dynamical systems. The model demonstrates superior capability in predicting chaotic system, achieving longer valid prediction times and significantly lower fitting errors compared to classical RC. In a drive-response configuration, heterogeneous memory-based RCs driven by a common scalar input reliably synchronize, and this coordinated behavior remains robust even under moderate-intensity observational noise. Furthermore, the proposed architecture synchronizes faster and with smaller steady-state error than conventional RC. Leveraging this robust synchronization property, we develop a chaos-based image encryption scheme that effectively obscures visual content, randomizes pixel distributions, and drastically reduces inter-pixel correlations, thereby offering strong resistance to statistical attacks while guaranteeing lossless decryption. These findings establish synchronization as an emergent behavior of trained memory-based RC systems and present a promising framework for secure communication applications.