Remote sensing multi-image encryption based on a dual-memristor brain-inspired chaotic neural network

Remote sensing multi-image transmission faces strong spatial correlation, heterogeneous image dimensions, packet disorder, and integrity risks. To address these issues, a dual-memristor brain-inspired chaotic neural network is constructed by introducing two flux-controlled memristive synapses with nonlinear saturating memductance and weak flux coupling into a four-neuron Hopfield network, providing key-dependent and reproducible chaotic states. Based on this network, a multi-image encryption framework is developed by integrating reversible RGB-pixel interleaving, metadata-aware sub-packaging, session key derivation, trajectory-based packet masks, packet scheduling, authentication, and self-described reconstruction. The receiver regenerates identical chaotic states and spatial mappings from the master key and authenticated session information, enabling packet-wise decryption and lossless recovery without transmitting complete permutation sequences. Experiments on five 256 × 256 RGB remote sensing images achieve zero-MSE recovery, an average ciphertext entropy of 7.9972 bits, an average absolute adjacent-pixel correlation of 0.002763, and NPCR and UACI values of 99.6006% and 33.4432%, respectively. The authentication-tag overhead is 4.17%. These results demonstrate effective statistical and differential security, integrity verification, and reversible multi-image protection.