Research on Neural Cryptographic Protocols

Research on discretized-neural-network-based cryptographic protocols is a novel topic in information security domain. At the beginning of this paper, literature on such neural cryptography is surveyed, and then an interacting neural network model, named tree parity machine (TPM), is introduced with a survey of its theoretical research. Then, several important issues regarding TPM's applications for cryptography are addressed and analyzed through engineering empirical study. Two key problems, namely, the stability of weight synchronization and the security of synchronization, are investigated in detail. Two new concept methods, the threshold in neuron activation function and hash-based synchronization check, are proposed to improve the TPM-based cryptography application scheme. ANOVA and statistical experiments results show that using threshold can achieve faster and more stable weight synchronization, while using hash function can check weight synchronization precisely. Finally, attack analysis for the improved TPM-used protocol models is discussed.