NEURAL CRYPTOGRAPHY BASED ON QUATERNION-VALUED NEURAL NETWORK

Neural cryptography is a public key exchange protocol based on mutual learning of two identically structured neural networks. After learning each other and eventual-ly reaching full synchronization, the two networks attain common network weights which can be used as a secret key for cryptographic purposes. The most common model used in neural cryptography is the tree parity machine (TPM). In this paper, neural cryptography is proposed based on quaternion-valued TPM (QVTPM). The mutual learning strategy is established and the security of QVTPM is theoretically analyzed. The main advantages of the proposed QVTPM are twofold: (i) the two communicating parties can exchange four group keys during one neural synchronization process; (ii) the security of QVTPM is higher than that of TPM and that of complex-valued TPM with the same number of input neurons and hidden units. Numerical simulation results ascertain the efficiency of QVTPM and our theoretical findings. © ICIC International.