SAECNet: Self-Attention Encryption and Compression Network Based on Bidirectional Cyclic Multiscroll Memristor Neural Network
Peizhen Li, Xiufang Feng, Herbert Ho‐Ching Iu, Shuang Zhou, Hao Zhang
Abstract
This article introduces a memristor-based neural network with bidirectional cyclic constructed from square wave pulse functions. Additionally, we design an image encryption and compression network framework incorporating a self-attention mechanism based on this memristor neural network. Specifically, this study introduces a novel model of a high-dimensional multiscroll memristor neural network (MMNN) that incorporates three neurons and a synapse made from memristive elements. Its complex dynamic behavior is analyzed in depth. Theoretical analysis and numerical simulations demonstrate that the proposed MMNN can generate an unlimited number of hyperchaotic multiscroll attractors and various other types of dynamics, which can be fine-tuned by modifying the system’s parameters or initial conditions. Additionally, a simulated equivalent circuit for MMNN proves the validity and practicality of the multiscroll attractor behavior. Finally, an image encryption and compression network framework based on end-to-end and MMNN and the self-attention mechanism is proposed, called SAECNet. Extensive performance evaluation highlights the system’s high accuracy in compression and reconstruction, strong encryption capabilities, and solid defense against potential attacks.