Security enhancement of artificial neural network using physically transient form of heterogeneous memristors with tunable resistive switching behaviors

阻变行为可调控的物理瞬态异质结构忆阻器及安全神经网络构建

As a critical command center in organisms,the brain can execute multiple intelligent interactions through neural networks,including memory,learning and cognition.Recently,memristive-based neuromorphic devices have been widely developed as promising technologies to build artificial synapses and neurons for neural networks.However,multiple information interactions in artificial intelligence devices potentially pose threats to information security.Herein,a transient form of heterogeneous memristor with a stacked structure of Ag/MgO/SiN_(x)/W is proposed,in which both the reconfigurable resistive switching behavior and volatile threshold switching characteristics could be realized by adjusting the thickness of the SiN_(x)layer.The underlying resistive switching mechanism of the device was elucidated in terms of filamentary and interfacial effects.Representative neural functions,including short-term plasticity(STP),the transformation from STP to long-term plasticity,and integrate-and-fire neuron functions,have been successfully emulated in memristive devices.Moreover,the dissolution kinetics associated with underlying transient behaviors were explored,and the water-assisted transfer printing technique was exploited to build transient neuromorphic device arrays on the water-dissolvable poly(vinyl alcohol)substrate,which were able to formless disappear in deionized water after 10-s dissolution at room temperature.This transient form of memristive-based neuromorphic device provides an important step toward information security reinforcement for artificial neural network applications.

作为生物体的重要指挥中心,大脑能够通过神经网络执行多种智能交互,包括记忆、学习和认知.近年来,基于忆阻器实现的人工突触和人工神经元构建神经网络作为一种极具前景的技术,已经得到了广泛的研究.然而,人工智能设备中的多重信息交互对信息安全构成潜在威胁.本文提出了物理瞬态形式的异质结构Ag/MgO/SiN_(x)/W忆阻器,通过调整SiN_(x)层的厚度分别实现了可重构阻变特性和易失性阈值开关特性,并从导电细丝的形成及其界面迁移特点出发阐述了该忆阻器的阻变机制.代表性的神经功能包括短时可塑性、短时可塑性向长期可塑性的转化以及神经元的整合-放电等神经拟态功能均可基于该忆阻器进行形象模拟.此外,我们还探索了与器件瞬态行为相关的溶解动力学,并利用水辅助转移法在可溶于水的聚乙烯醇衬底上构建了瞬态形式的神经形态器件阵列,其在室温下的去离子水中浸泡10 s后就可完成物理形态的完全消失.该物理瞬态忆阻式神经形态器件对于人工神经网络应用中的信息安全加固具有重要意义.