Research On Memristor Neural Network And Its Application In Encryption Technology

The human neural system is a highly complex and large-scale nonlinear dynamic system,and its dynamic behavior is closely related to human intelligence activities.Therefore,the research on the dynamic behavior of neurons or neural network models has become an important content of artificial neural networks.Memristors have become one of the most ideal devices for constructing artificial neural synapses because of their unique memory characteristics,non-volatility,compatibility with CMOS technology and other advantages.Its good nonlinearity and randomness become an important condition for constructing nonlinear systems and generating rich dynamic behaviors.In order to accurately grasp how the human brain works and accelerate the realization of artificial intelligence technology with a higher degree of intelligence,the study of neural network dynamics based on memristors has become a key field of exploration,but there are still problems such as model simplification,insufficient application and lack of hardware research.On the basis of studying the dynamic behavior of memristive neural network,this dissertation combines the neural network with encryption technology,and designs a variety of effective neural network encryption algorithms and its hardware implementation,including the following three aspects:1.Two neural network models based on memristor are proposed,which show rich dynamic behaviors.Based on the analysis of the classical memristor model,a memristor model with negative resistance state is proposed and used for synaptic connection weights to construct a Negative-resistance Memristor-based Hopfield Neural Network（NMHNN）.The dynamic behavior of NMHNN is analyzed by using phase trajectory diagram and Lyapunov exponent,and the single scroll attractor chaotic phenomenon is observed,which proves that NMHNN has the simple structure and complex dynamics.Secondly,the HP memristor is used for neuron self-feedback connection weights to construct a Memristor-based Transient Chaotic Neural Network（MTCNN）.Compared with the original Transiently Chaotic Neural Network（TCNN）,it has richer dynamic behavior,and MTCNN is 40%and 25%more efficient than TCNN in dealing with the Traveling Salesman Problem（TSP）and the Channel Assignment Problem（CAP）,respectively.2.Three memristive neural network encryption algorithms are designed,and their security performance is analyzed in detail.In view of the problems of fixed key,derivable key and small key space of traditional encryption algorithms Advanced Encryption Standard（AES）and Rivest Cipher 4（RC4）,the improved algorithms of AES and RC4encryption based on MTCNN are designed and compared with the original algorithms.The improved algorithm satisfies the strict avalanche principle,and is superior to the original algorithms in sequence randomness and histogram uniformity.Moreover,the key space increases exponentially compared with the original algorithm,reaching 10¹²³,which can effectively resist statistical analysis attacks and brute force exhaustion attack.Aiming at the low efficiency of image encryption in traditional encryption algorithms such as AES,an image encryption algorithm based on NMHNN is proposed,which uses confusion-diffusion structure to achieve image encryption and decryption.In addition,the histogram,correlation,sensitivity,randomness,information entropy,key space and encryption efficiency are analyzed.The encryption efficiency is 23 times higher than AES algorithm.3.According to the characteristics of MTCNN and NMHNN,the hardware of neural network encryption is designed,and the results are analyzed.Firstly,the single neuron circuit of the MTCNN is designed and simulated,and the results are consistent with the previous theoretical analysis.Secondly,the electrical characteristics of the LNO single crystal thin film memristor are tested,and the results are used as the self-feedback connection weight of MTCNN.Based on this,an S-box dynamic search generation algorithm is proposed,and the S-box has better performance than other related algorithms in terms of nonlinearity and differential uniformity.Finally,by studying the electrical characteristics of the Hf O₂ memristors,12 memristors are used to build a 3?4 synaptic network,and a three neurons Hopfield neural network circuit is designed and implemented.The hardware encryption of the image is successfully realized,and the average NPCR is 99.60%and UACI is 33.47%in the plaintext sensitivity test,which is better than the software algorithm.In summary,this dissertation advances the research process from the four levels of device,network,algorithm and circuit,and carries out systematic research on memristive neural network and its encryption technology,laying the foundation for the large-scale and high-performance memristive neural network encryption technology and hardware research.