WO2020238224A1 - 一种基于机器学习的量子通信系统的主动反馈控制方法 - Google Patents

一种基于机器学习的量子通信系统的主动反馈控制方法 Download PDF

Info

Publication number
WO2020238224A1
WO2020238224A1 PCT/CN2020/070401 CN2020070401W WO2020238224A1 WO 2020238224 A1 WO2020238224 A1 WO 2020238224A1 CN 2020070401 W CN2020070401 W CN 2020070401W WO 2020238224 A1 WO2020238224 A1 WO 2020238224A1
Authority
WO
WIPO (PCT)
Prior art keywords
phase
network
feedback control
voltage
time
Prior art date
Application number
PCT/CN2020/070401
Other languages
English (en)
French (fr)
Inventor
王琴
刘靖阳
Original Assignee
南京邮电大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 南京邮电大学 filed Critical 南京邮电大学
Priority to US17/040,996 priority Critical patent/US11817911B2/en
Publication of WO2020238224A1 publication Critical patent/WO2020238224A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L9/00—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols
    • H04L9/08—Key distribution or management, e.g. generation, sharing or updating, of cryptographic keys or passwords
    • H04L9/0816—Key establishment, i.e. cryptographic processes or cryptographic protocols whereby a shared secret becomes available to two or more parties, for subsequent use
    • H04L9/0852—Quantum cryptography
    • H04L9/0858—Details about key distillation or coding, e.g. reconciliation, error correction, privacy amplification, polarisation coding or phase coding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B10/00—Transmission systems employing electromagnetic waves other than radio-waves, e.g. infrared, visible or ultraviolet light, or employing corpuscular radiation, e.g. quantum communication
    • H04B10/70—Photonic quantum communication

Definitions

  • the invention belongs to the field of quantum information technology, and specifically relates to an active feedback control method of a quantum communication system based on machine learning.
  • Quantum cryptography is the core of quantum communication, and its security directly determines the security of quantum communication systems.
  • the security of quantum cryptography is based on the basic principles of quantum mechanics.
  • Opposite One Time OTP
  • it can provide unconditional security quantum communications for legitimate users (Alice, Bob). Since the first quantum cryptography protocol-BB84 protocol was proposed, quantum cryptography has made great progress both in theory and in experiments.
  • Existing practical quantum cryptography systems can use different encoding methods, such as phase, polarization, and time-energy encoding, among which systems based on phase encoding are the most widely used.
  • the most widely used method is the interference ring scanning-transmission method, which realizes the compensation of the system phase at intervals.
  • the quantum cryptosystem cannot transmit signals, resulting in lower overall system efficiency.
  • FPGA-based real-time phase compensation technology can be used, but this method requires relatively high hardware overhead and also increases the complexity of the entire quantum cryptographic system.
  • the purpose of the present invention is to address the above-mentioned shortcomings of the prior art and propose a method for active feedback control of a quantum communication system based on machine learning, which can be applied to a quantum key distribution (QKD) system.
  • QKD quantum key distribution
  • the pre-trained double-layer LSTM network is used to predict the zero-phase voltage value of the phase modulator at the Bob terminal at the next time based on the real-time environmental temperature, humidity, laser light intensity, and voltage changes in the past time. And by updating the network at fixed time intervals, the LSTM network can accurately predict for a long time, so that the quantum key distribution system always maintains a stable and efficient operation state.
  • the scheme of the invention greatly improves the transmission efficiency of the quantum key distribution system without increasing the complexity of the system hardware.
  • An active feedback control method for a quantum communication system based on machine learning applies a machine learning model of a long short-term memory neural network (LSTM) to a quantum communication system, and uses phase-encoded quantum key distribution (QKD) )
  • LSTM long short-term memory neural network
  • QKD phase-encoded quantum key distribution
  • the QKD system includes at least two user terminals, Alice and Bob;
  • the method includes a training phase, a prediction phase, and an update phase in sequence;
  • the length of the sequence represents the time span of the data
  • the data at each time point in the sequence is composed of corresponding features and labels; wherein, the features can be composed of environmental temperature, humidity, The intensity of the laser, the voltage at the current time point, and the voltage at the previous four time points; the label can be composed of the voltage at the next time point;
  • the entire training process requires at least 270 rounds , All data needs to be standardized by Z-score before entering the network;
  • the LSTM network can be connected to the phase voltage control system of the phase modulator on the Bob side.
  • the LSTM network needs to read the current temperature and humidity from the temperature and humidity detector in real time, and read the current laser from the optical power meter in real time Light intensity, read the displacement voltages at five time points from the shift register in real time.
  • the voltages at these five time points are the voltage value at the current time point and the voltage value at the previous four times; the network predicts the next one based on the input data Zero phase voltage at the time point, and input the voltage value into the phase modulator at Bob's end to realize the phase stability control of the system;
  • LSTM In order to enable LSTM to have the ability to accurately predict for a long time, a combination of prediction and update is adopted; after a period of network prediction, the interference loop scanning program is briefly run to obtain accurate zero-phase voltage and fed back to the LSTM network, LSTM According to the accurate label value, the error back propagation method is used to update its weight and bias value, and the updated LSTM network is converted to the prediction mode again.
  • the typical characteristics of the data at each time point in the sequence are composed of temperature, humidity, laser light intensity, and displacement voltage.
  • the displacement voltage can be composed of a current voltage and the voltage at the previous four times.
  • the composition is not limited to the typical characteristics given above, and other values can be selected according to actual needs.
  • the feedback control of the active feedback control method can be completed by the interference loop scanning program, and the feedback process is to update the LSTM network; the real-time voltage scanned by the interference loop scanning program is fed back to the LSTM network, so that the network updates its weight And the bias voltage value; the feedback control of the active feedback control method is not limited to the interference ring scanning procedure mentioned above, and can also be completed by other scanning procedures including polarization scanning.
  • the training phase and the prediction phase of the active feedback control method can be separated from each other.
  • the method uses the method of continuously updating the network before each prediction so that the LSTM network only needs to be fine-tuned according to the actual situation in the prediction phase, thereby achieving training Separation of phase and prediction phase.
  • the active feedback control method uses a double-layer LSTM network for the stable phase modulation process of the QKD system, but is not limited to using a double-layer LSTM network, and the network structure can be adjusted according to the complexity of the actual quantum communication system.
  • the beneficial effects of the present invention are: compared with the conventional interference loop scanning program and the FPGA-based real-time phase compensation scheme, the present invention scheme adopts the prediction and feedback control method based on the software-controlled long and short-term memory neural network without adding additional hardware
  • the equipment not only eliminates the system complexity caused by the use of additional equipment, but also avoids the risk of possible side channel vulnerabilities.
  • this method can greatly improve the transmission efficiency of the entire QKD system, realize real-time software-based phase compensation control, and the interval update process can also make the system run stably for a long time.
  • the invention can ensure that the QKD system can operate stably and efficiently for a long time under the condition of ensuring the same error level as the traditional method.
  • Figure 1 is a flow chart of the solution of the present invention.
  • Fig. 2 is an internal structure diagram of an LSTM network according to an embodiment of the present invention.
  • Figure 3 is a diagram of the data structure of each time point in the sequence in the present invention.
  • Fig. 4 is a diagram of an experimental device of the QKD system used in an embodiment of the present invention.
  • Fig. 5 is a comparison diagram of error codes between the present invention and the "scan-transmit" scheme.
  • Figure 6 is a graph of the long-term running test results of the present invention.
  • the phase voltage control system of the present invention mainly uses the LSTM network to predict the zero-point phase voltage.
  • the following is the working principle of the LSTM network:
  • the LSTM network consists of a series of repeated neural network modules. As shown in Figure 1, its core is the cell conveyor belt C t . The knowledge learned by LSTM network training is transmitted along this conveyor belt and runs through the entire chain. The LSTM network deletes or adds information to the cell state through the gate structure.
  • An LSTM block has three gate structures to control the state of a cell unit, which are called forget gate, input gate and output gate.
  • the first is that the forget gate determines which information from the previous LSTM block needs to be discarded.
  • the features x t of the current state and the output h t of the previous LSTM block will be processed by the corresponding weights W f and paranoia b f , and then processed by a sigmoid activation function ⁇ , and the output result f t is as follows:
  • the network realizes the input of specific information through two processes.
  • an input gate composed of the sigmoid layer plays a role in updating the information in the cell, the formula is as follows:
  • a tanh layer adds the information that helps the network to realize the memory function into the cell structure in the form of a vector, the formula is as follows:
  • the cell structure realizes the update of the memory information in the cell through the multiplication operation with the above two gate structures, the formula is as follows:
  • the LSTM network can realize the long-term memory effect of information.
  • the training phase needs to divide the training data into many sequences before and after the time.
  • the length of the sequence represents the time span of the data.
  • the data at each time point in the sequence is composed of corresponding features and labels. Among them, the characteristics are composed of ambient temperature, humidity, laser intensity, voltage at the current time point, and voltage at the previous four time points.
  • the label consists of the voltage at the next point in time.
  • the training phase first collects training data.
  • the process first uses the traditional interference ring scanning program to obtain the data of the zero-phase voltage applied by the phase modulator on the Bob end with time, and the temperature, humidity, and laser intensity changes with time.
  • the data is spliced with zero-phase voltage data.
  • Each row of the data is a specific feature value at a certain moment, and each column is a feature that changes over time.
  • Each piece of training data consists of 3,600 data points, a total of ten pieces of training data.
  • the duty cycle of the experimental system used in the solution of the present invention is 0.5, that is, another 10 seconds are needed for phase compensation after every 10 seconds of transmission. Therefore, the time span corresponding to each piece of training data is 20 hours.
  • the solution of the present invention uses the Adam optimization algorithm.
  • the indicator of the training process is the mean square error.
  • the entire training process requires at least 270 rounds, and all data needs to be standardized by Z-score before being input into the network.
  • the LSTM network After training, the LSTM network should be connected to the phase voltage control system of the phase modulator on the Bob side.
  • the LSTM network needs to read the current temperature and humidity from the temperature and humidity detector in real time, and the current laser from the optical power meter in real time.
  • Light intensity read the displacement voltages at five time points from the shift register in real time. The voltages at these five time points are the voltage value at the current time point and the voltage value at the previous four time points.
  • the network predicts the zero-phase voltage at the next time point based on the input data, and inputs the voltage value into the phase modulator at the Bob end to realize the phase stability control of the system.
  • the solution of the present invention adopts a mode of changing to an update mode after continuously predicting 25 voltage values, wherein the time span of each continuous prediction is 5 minutes.
  • the solution of the present invention adopts a working mode that combines prediction and update. After the network predicts for a period of time, the interference loop scanning program is run for a short time to obtain the accurate zero-phase voltage and fed back to the LSTM network. The LSTM uses the error back propagation method to update its weights and bias values according to the accurate label value. The network reverts to predictive mode. The time for each update phase is 50 seconds.
  • FIG. 3 is a diagram of the experimental device of the present invention.
  • the laser (repetition frequency 1MHz, center wavelength 1550nm) outputs laser light to a 1:99 beam splitter (BS), which splits the beam into two paths, and 1% of the light is sent to the optical power meter, 99% The light is sent to the Faraday Michelson interference ring (FMI) at the Alice side.
  • FMI Faraday Michelson interference ring
  • Each light pulse sent into the FMI is randomly encoded into a state in the BB84 protocol, and then transmitted to the receiving end Bob through a commercial single-mode fiber.
  • the Bob end randomly selects the X or Z base to measure the quantum state, and the measurement is done through the FMI on the Bob end.
  • a control box (CB) and a computer are placed on the Alice side and Bob side. These two devices are used to run the LSTM network and deliver voltage to the phase modulator.
  • the solution of the invention is also equipped with an optical power meter (OPM), a temperature and humidity detector (THD), and a single photon detector (SPD) for real-time recording of laser light intensity, temperature, humidity, and photon count rate.
  • OPM optical power meter
  • TDD temperature and humidity detector
  • SPD single photon detector
  • the single photon detector used in the experimental system of the present invention is an InGaAs detector working in a gate mode.
  • the scheme experiment of the present invention has passed 50 kilometers and 150 kilometers of optical fiber detection, and at the same time, the results are compared with the traditional interference ring scanning program.
  • the transmission process uses three kinds of intensity (signal state intensity 0.5, decoy state intensity 0.1, vacuum state intensity 0) to modulate the light pulse.
  • the background error of the experimental system of the scheme of the present invention is 1.23%
  • the detector efficiency is 10%
  • the dark count rate is 0.8MHz.
  • Fig. 4 is a comparison diagram of error codes between the scheme of the present invention and the traditional scheme.
  • Figure 4 (a) and (b) show the change of the system signal state qubit error rate (QBER) at 50km and 150km in 48 hours, respectively. It can be seen from the two figures that the QBER result of the scheme of the present invention is basically at the same level as the traditional interference curve scanning scheme, which proves the stability and reliability of the scheme of the present invention.
  • Figure 4(c) shows the comparison between the result of the code rate experiment and the theoretical simulation result.
  • the solid line in the figure is the theoretical simulation result, the square dots are the experimental results of the traditional scheme, and the circular dots are the experimental results of the inventive scheme.
  • Fig. 5 is a test result diagram of the long-term running of the scheme of the present invention.
  • the QBER did not show a significant increase, which proves that the solution of the present invention can still maintain the accuracy of the prediction and the stability of the QBER of the system for a long time.
  • the present invention has experimentally verified a phase prediction and feedback control scheme for a quantum key distribution system based on a long and short-term memory neural network.
  • the transmission efficiency of the QKD system can be increased to more than 83% , While ensuring that the QBER of the system remains at the same level as the traditional solution.
  • the scheme of the invention can be extended to any QKD protocol and system.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Electromagnetism (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Optics & Photonics (AREA)
  • Computer Security & Cryptography (AREA)
  • Optical Modulation, Optical Deflection, Nonlinear Optics, Optical Demodulation, Optical Logic Elements (AREA)
  • Optical Communication System (AREA)

Abstract

一种基于机器学习的量子通信系统的主动反馈控制方法，在量子密钥分发系统的传输过程中，本发明利用已预先训练完成的双层LSTM网络，根据外界环境中实时温度、湿度、激光器光强起伏，以及过去时刻的电压变化预测下一时刻接收端的相位调制器的零相位电压值，并通过固定时间间隔对网络进行更新，使该LSTM网络能够长时间准确预测，从而保证量子密钥分发系统长时间高效稳定运行。本发明通过主动预测、反馈控制的方法，极大地提高了量子密钥分发系统的传输效率。本发明不仅限于应用在量子密钥分发系统或相位编码系统之中，也同样适用于基于其他编码方式的量子密钥分发系统或量子通信网络之中。

Description

一种基于机器学习的量子通信系统的主动反馈控制方法 技术领域
本发明属于量子信息技术领域，具体涉及一种基于机器学习的量子通信系统的主动反馈控制方法。
背景技术
量子密码是量子通信的核心，它的安全与否直接决定着量子通信系统的安全性。量子密码的安全性建立在量子力学的基本原理之上，同时通过结合香农提出的“一次一密”(OTP)定理，原则上可以为合法用户(Alice、Bob)提供无条件安全性的量子通信。自从第一个量子密码协议——BB84协议提出以来，量子密码无论是在理论上还是在实验上均取得了巨大的进步。现有实用化的量子密码系统可以使用不同的编码方式，比如相位、偏振、时间-能量编码等，其中基于相位编码的系统应用最为广泛。但是由于该类系统中存在着相位漂移问题，因而需要不断对发送端和接收端的相位进行实时校准。目前使用最广泛的是干涉环扫描-传输的方法，该方法每隔一段时间实现对系统相位的补偿。然而，在干涉环扫描程序工作期间，量子密码系统无法传输信号，导致系统整体效率较低。为了提高量子密码系统的传输效率，可以使用基于FPGA的相位实时补偿技术，但这种方法需要比较高昂的硬件开销，同时也增加了整个量子密码系统的复杂度。
发明内容
本发明目的在于针对上述现有技术的不足，提出了一种基于机器学习的量子通信系统主动反馈控制方法，该方法可以应用于量子密钥 分发(QKD)系统中。在QKD系统的稳定调相阶段，利用已预先训练完成的双层LSTM网络根据实时环境温度、湿度、激光器光强及过去时刻的电压变化预测下一时刻Bob端的相位调制器的零相位电压值，并通过固定时间间隔的对网络进行更新，使LSTM网络能够长时间准确预测，从而使得量子密钥分发系统始终保持稳定的高效率运行状态。本发明方案在不提高系统硬件复杂度的前提下极大的提高了量子密钥分发系统的传输效率。
一种基于机器学习的量子通信系统的主动反馈控制方法，所述方法将长短期记忆神经网络(LSTM)的机器学习模型应用于量子通信系统之中，并且以相位编码的量子密钥分发(QKD)作为其中一个应用场景，但不仅限于QKD系统或相位编码系统；该QKD系统至少包括两个用户端Alice端和Bob端；
所述方法依次包括训练阶段、预测阶段和更新阶段；
训练阶段：
根据时间前后将训练数据分为很多序列，序列的长度代表该段数据的时间跨度，序列中每一个时间点的数据由相应的特征和标签组成；其中，所述特征可以由环境温度、湿度、激光器的强度、当前时间点的电压以及前四个时间点的电压组成；所述标签则可以由下一个时间点的电压组成；在训练网络时，使用Adam优化算法，整个训练过程至少需270轮，所有数据在输入进网络前需经过Z-score的标准化；
预测阶段：
训练完成后的LSTM网络可以接入Bob端的相位调制器的相位电 压调控系统，LSTM网络在该阶段需要从温湿度探测器实时读取当前时间的温度、湿度，从光功率计实时读取当前激光光强，从移位寄存器实时读取五个时间点的位移电压，这五个时间点的电压分别为当前时间点的电压值以及前四个时刻的电压值；网络根据输入数据预测出下一个时间点的零相位电压，并将该电压值输入Bob端的相位调制器，以此实现系统的相位稳定控制；
更新阶段：
为使LSTM具备长时间准确预测的能力，采取了预测与更新相结合的工作模式；在网络预测一段时间后，通过短暂运行干涉环扫描程序以获取准确的零相位电压并反馈回LSTM网络，LSTM根据准确的标签值使用误差逆传播方法更新其权重与偏置值，更新后的LSTM网络则重新转为预测模式。
进一步地，所述训练阶段中，序列中每一个时间点的数据，典型的特征由温度、湿度、激光光强、位移电压组成，其中，位移电压可以由一个当前电压和前四个时刻的电压组成，但不仅仅限于以上给出的典型特征，可以根据实际需求选择其他数值。
进一步地，所述主动反馈控制方法的反馈控制可以由干涉环扫描程序完成，其反馈过程以更新LSTM网络为目的；将干涉环扫描程序扫描出的实时电压反馈回LSTM网络，使网络更新其权重与偏置电压值；所述主动反馈控制方法的反馈控制不仅仅限于以上提到的干涉环扫描程序完成，也可以通过包括偏振扫描在内的其他扫描程序完成。
进一步地，所述主动反馈控制方法的训练阶段与预测阶段可以互 相分开，本方法采用在每次预测前连续更新网络的方式使得LSTM网络只需根据预测阶段的实际情况进行微调，从而实现了训练阶段与预测阶段的分离。
进一步地，所述主动反馈控制方法将双层LSTM网络用于QKD系统的稳定调相过程，但不仅仅限于使用双层LSTM网络，可以根据实际量子通信系统的复杂程度对网络结构做相应调整。
本发明的有益效果为：相比常规的干涉环扫描程序以及基于FPGA的相位实时补偿方案，本发明方案采用基于软件控制的长短期记忆神经网络的预测和反馈控制方法，不需要加入额外的硬件设备，不仅消除了使用额外设备所带来的系统复杂度，同时也避免了可能的侧信道漏洞风险。而且该方法可以极大地提高整个QKD系统的传输效率，实现实时的基于软件控制的相位补偿控制，并且间隔的更新过程也能使系统长时间稳定运行。本发明能够在保证与传统方法同等误码水平的条件下，QKD系统长时间高效率地稳定运行。
附图说明
图1是本发明方案的流程图。
图2是本发明实施例的LSTM网络的内部结构图。
图3是本发明中序列中每一个时间点的数据的结构图。
图4是本发明实施例使用的QKD系统实验装置图。
图5是本发明与“扫描-传输”方案的误码对比图。
图6是本发明的长时间运行测试结果图。
具体实施方式
下面结合说明书附图对本发明的技术方案做进一步的详细说明。
本发明方案的相位电压控制系统主要通过使用LSTM网络来预测零点相位电压。以下为该LSTM网络的工作原理：
LSTM网络由一连串重复的神经网络模块所组成。如图1所示，其核心为细胞传送带C t。LSTM网络训练所学习到的知识沿这条传送带进行传递，并贯穿整个链条。而LSTM网络是通过门结构来删除或添加信息到细胞状态。一个LSTM块有三个门结构来控制一个细胞单元的状态，分别称为忘记门、输入门和输出门。
首先是忘记门决定了由上一个LSTM块传递而来的哪些信息需要被丢弃。当前状态的特征x t和上一LSTM块的输出h t会经过相对应的权重W f和偏执b f的处理，而后通过一个sigmoid激活函数σ的处理，其输出结果f t如下：
f t＝σ(W f·[h t-1,x t]+b f)。
下一步网络通过两个过程实现了特定信息的输入。首先，一个由sigmoid层组成的输入门起到更新细胞内信息的作用，公式如下：
i t＝σ(W i·[ht -1,x t]+b i)，
接着一个tanh层将有助于网络实现记忆功能的信息以向量的形式添加进细胞结构，公式如下：
Figure PCTCN2020070401-appb-000001
细胞结构通过与上述两个门结构的乘法操作实现细胞内记忆信息的更新，公式如下：
Figure PCTCN2020070401-appb-000002
最后，将细胞状态通过tanh层的处理并与输出门做乘法操作，实现特定信息的输出，公式如下：
O t＝σ(W o·[h t-1,x t]+b o)，
h t＝O t×tanh(C t)。
通过以上方式的计算，LSTM网络能够实现信息的长期记忆效果。
下面将详细介绍基于长短期记忆神经网络的量子密钥分发相位预测与反馈控制方案的实现过程：
训练阶段：
考虑到LSTM网络的时间记忆特性，训练阶段需根据时间前后将训练数据分为很多序列，序列的长度代表该段数据的时间跨度，序列中每一个时间点的数据由相应的特征和标签组成。其中，特征由环境温度、湿度、激光器的强度、当前时间点的电压以及前四个时间点的电压组成。而标签则由下一个时间点的电压组成。
训练阶段首先收集训练数据，该过程先使用传统的干涉环扫描程序，用于获取Bob端的相位调制器所加的零相位电压随时间变化的数据，并将温度、湿度、激光器强度随时间变化的数据与零相位电压数据拼接。该数据的每一行为某时刻的具体特征值，每一列为一个特征在时间变化。每一段训练数据由3600个数据点组成，总共十段训练数据。本发明方案所使用的实验系统的占空比为0.5，即每传输10秒之后需要另外10秒时间用于补偿相位。因此，每一段训练数据对应的时间跨度为20小时。
在训练网络时，本发明方案使用Adam优化算法，训练过程的指 标为均方误差，整个训练过程至少需270轮，所有数据在输入进网络前需经过Z-score的标准化。
预测阶段：
训练完成后的LSTM网络应接入Bob端的相位调制器的相位电压调控系统，LSTM网络在该阶段需要从温湿度探测器实时读取当前时间的温度、湿度，从光功率计实时读取当前激光光强，从移位寄存器实时读取五个时间点的位移电压，这五个时间点的电压分别为当前时间点的电压值以及前四个时刻的电压值。网络根据输入数据预测出下一个时间点的零相位电压，并将该电压值输入Bob端的相位调制器，以此实现系统的相位稳定控制。
本发明方案采用连续预测25个电压值后改为更新模式的方式，其中每次连续预测的时间跨度为5分钟。
更新阶段：
为使LSTM具备长时间准确预测的能力，本发明方案采取了预测与更新相结合的工作模式。在网络预测一段时间后，通过短暂运行干涉环扫描程序以获取准确的零相位电压并反馈回LSTM网络，LSTM根据准确的标签值使用误差逆传播方法更新其权重与偏置值，更新后的LSTM网络则重新转为预测模式。其中每次更新阶段的时间为50秒。
为使本发明的目的、技术方案和优点更加清楚明白，以下结合具体实施例，并参照附图，对本发明作进一步的详细说明。
图3是本发明方案的实验装置图。Alice端，激光器(重复频率 1MHz、中心波长1550nm)输出激光给1:99的分束器(BS)，该分束器将光束分为两路，1％的光送入光功率计，99％的光送入Alice端的法拉第迈克尔逊干涉环(FMI)。每一个送入FMI的光脉冲都被随机地编码为BB84协议中的一个态，而后经过商用单模光纤传输给接收端Bob。
Bob端随机选择X基或Z基对量子态进行测量，测量通过Bob端的FMI完成。在Alice端和Bob端均放置有控制箱(CB)以及电脑，这两项设备用于运行LSTM网络以及给相位调制器输送电压。本发明方案同时配有光功率计(OPM)、温湿度探测器(THD)、单光子探测器(SPD)分别用于实时记录激光光强、温度、湿度、光子计数率。本发明方案的实验系统所使用的单光子探测器为工作于门模式的InGaAs探测器。
本发明方案实验经过了50公里和150公里光纤的检测，同时也与传统的干涉环扫描程序进行了结果对比。传输过程分别使用了三种强度(信号态强度0.5、诱骗态强度0.1、真空态强度0)对光脉冲进行调制。同时本发明方案的实验系统的本底误码为1.23％，探测器效率为10％，暗计数率为0.8MHz。
图4为本发明方案与传统方案的误码对比图。图4(a)和(b)分别显示了48小时内50km和150km下系统信号态量子比特误码率(QBER)的变化。从两幅图可以看出，本发明方案的QBER结果与传统的干涉曲线扫描方案基本处于同一水平，这证明了本发明方案的稳定性与可靠性。图4(c)展示了成码率实验的结果与理论仿真结 果的对比，图中实线为理论仿真结果，方形点为传统方案的实验结果，圆形点为本发明方案的实验结果。
图5是本发明方案长时间运行的测试结果图。图中，在系统连续运行的十天内，QBER没有表现出明显的上升，证明了本发明方案长时间运行依然可以维持预测的准确性和系统QBER的稳定。
综上，本发明实验验证了一种基于长短期记忆神经网络的量子密钥分发系统相位预测与反馈控制方案，利用预测电压加更新网络的方法，能够将QKD系统的传输效率提高至83％以上，同时还能保证系统的QBER保持在与传统方案同等大小的水平。此外本发明方案还可以扩展到任何QKD协议及系统。
以上所述仅为本发明的较佳实施方式，本发明的保护范围并不以上述实施方式为限，但凡本领域普通技术人员根据本发明所揭示内容所作的等效修饰或变化，皆应纳入权利要求书中记载的保护范围内。

Claims (5)

  1. 一种基于机器学习的量子通信系统的主动反馈控制方法，其特征在于：
    所述方法将长短期记忆神经网络(LSTM)的机器学习模型应用于量子通信系统之中，并且以相位编码的量子密钥分发(QKD)作为其中一个应用场景，但不仅限于QKD系统或相位编码系统；该QKD系统至少包括两个用户端Alice端和Bob端；
    所述方法依次包括训练阶段、预测阶段和更新阶段；
    训练阶段：
    根据时间前后将训练数据分为很多序列，序列的长度代表该段数据的时间跨度，序列中每一个时间点的数据由相应的特征和标签组成；其中，所述特征可以由环境温度、湿度、激光器的强度、当前时间点的电压以及前四个时间点的电压组成；所述标签则可以由下一个时间点的电压组成；在训练网络时，使用Adam优化算法，整个训练过程至少需270轮，所有数据在输入进网络前需经过Z-score的标准化；
    预测阶段：
    训练完成后的LSTM网络可以接入Bob端的相位调制器的相位电压调控系统，LSTM网络在该阶段需要从温湿度探测器实时读取当前时间的温度、湿度，从光功率计实时读取当前激光光强，从移位寄存器实时读取五个时间点的位移电压，这五个时间点的电压分别为当前时间点的电压值以及前四个时刻的电压值；网络根据输入数据预测出下一个时间点的零相位电压，并将该电压值输入Bob端的相位调制器，以此实现系统的相位稳定控制；
    更新阶段：
    为使LSTM具备长时间准确预测的能力，采取了预测与更新相结合的工作模式；在网络预测一段时间后，通过短暂运行干涉环扫描程序以获取准确的零相位电压并反馈回LSTM网络，LSTM根据准确的标签值使用误差逆传播方法更新其权重与偏置值，更新后的LSTM网络则重新转为预测模式。
  2. 根据权利要求1所述的一种基于机器学习的量子通信系统的主动反馈控制方法，其特征在于：所述训练阶段中，序列中每一个时间点的数据，典型的特征由温度、湿度、激光光强、位移电压组成，其中，位移电压可以由一个当前电压和前四个时刻的电压组成，但不仅仅限于以上给出的典型特征，可以根据实际需求选择其他数值。
  3. 权利要求1所述的一种基于机器学习的量子通信系统的主动反馈控制方法，其特征在于：所述主动反馈控制方法的反馈控制可以由干涉环扫描程序完成，其反馈过程以更新LSTM网络为目的；将干涉环扫描程序扫描出的实时电压反馈回LSTM网络，使网络更新其权重与偏置电压值；所述主动反馈控制方法的反馈控制不仅仅限于以上提到的干涉环扫描程序完成，也可以通过包括偏振扫描在内的其他扫描程序完成。
  4. 权利要求1所述的一种基于机器学习的量子通信系统的主动反馈控制方法，其特征在于：所述主动反馈控制方法的训练阶段与预测阶段可以互相分开，本方法采用在每次预测前连续更新网络的方式 使得LSTM网络只需根据预测阶段的实际情况进行微调，从而实现了训练阶段与预测阶段的分离。
  5. 权利要求1所述的一种基于机器学习的量子通信系统的主动反馈控制方法，其特征在于：所述主动反馈控制方法将双层LSTM网络用于QKD系统的稳定调相过程，但不仅仅限于使用双层LSTM网络，可以根据实际量子通信系统的复杂程度对网络结构做相应调整。
PCT/CN2020/070401 2019-05-31 2020-01-06 一种基于机器学习的量子通信系统的主动反馈控制方法 WO2020238224A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/040,996 US11817911B2 (en) 2019-05-31 2020-01-06 Active feedback control method for quantum communication system based on machine learning

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910469672.XA CN110365473B (zh) 2019-05-31 2019-05-31 一种基于机器学习的量子通信系统的主动反馈控制方法
CN201910469672.X 2019-05-31

Publications (1)

Publication Number Publication Date
WO2020238224A1 true WO2020238224A1 (zh) 2020-12-03

Family

ID=68214965

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/070401 WO2020238224A1 (zh) 2019-05-31 2020-01-06 一种基于机器学习的量子通信系统的主动反馈控制方法

Country Status (3)

Country Link
US (1) US11817911B2 (zh)
CN (1) CN110365473B (zh)
WO (1) WO2020238224A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113572532A (zh) * 2021-04-14 2021-10-29 华南师范大学 光纤干涉环主动相位补偿方法、装置和量子密钥分发系统
CN114511070A (zh) * 2022-04-20 2022-05-17 南京邮电大学 一种适用于双场量子密钥分发系统的零相位电压估计方法

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110365473B (zh) * 2019-05-31 2021-11-26 南京邮电大学 一种基于机器学习的量子通信系统的主动反馈控制方法
CN110798314B (zh) * 2019-11-01 2023-02-24 南京邮电大学 一种基于随机森林算法的量子密钥分发参数优化方法
CN111082938B (zh) * 2020-03-25 2020-08-28 北京中创为南京量子通信技术有限公司 一种提高量子密钥分发系统成码率的方法及装置
CN111738426A (zh) * 2020-08-07 2020-10-02 南京邮电大学 一种基于gru模型的量子密码系统稳定控制方法
CN113055167B (zh) * 2021-03-22 2022-06-03 上海循态量子科技有限公司 基于芯片cvqkd实际系统中安全漏洞的防御方法及系统
CN113411185A (zh) * 2021-06-09 2021-09-17 南京邮电大学 一种基于gru模型的量子密码系统稳定控制方法
CN113810186B (zh) * 2021-09-18 2023-11-07 上海循态量子科技有限公司 自适应量子效率高精度实时预测方法及系统
CN114912608B (zh) * 2022-05-16 2023-04-07 南京邮电大学 一种适用于双场量子密钥分发系统的全局相位跟踪预测方法
CN117458261B (zh) * 2023-12-26 2024-04-16 东莞市湃泊科技有限公司 激光器封装系统及其智能散热方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006124209A1 (en) * 2005-05-17 2006-11-23 Lucent Technonologies Inc. Phase locking in a multi-channel quantum communication system
CN105490805A (zh) * 2015-11-24 2016-04-13 上海斐讯数据通信技术有限公司 一种基于扩展卡尔曼滤波降低qkd系统误码率的系统及方法
CN107612688A (zh) * 2017-09-30 2018-01-19 上海交通大学 基于机器学习的cvqkd实时性能优化方法及系统
CN110365473A (zh) * 2019-05-31 2019-10-22 南京邮电大学 一种基于机器学习的量子通信系统的主动反馈控制方法

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101272373B (zh) * 2008-05-07 2010-09-08 北京北方烽火科技有限公司 一种自适应模拟正交调制失衡补偿方法和装置
US9742593B2 (en) * 2015-12-16 2017-08-22 Kumu Networks, Inc. Systems and methods for adaptively-tuned digital self-interference cancellation
CN106254071B (zh) * 2016-09-27 2019-05-14 四川航天机电工程研究所 一种量子密钥分发的实时相位补偿系统和方法
US10275721B2 (en) * 2017-04-19 2019-04-30 Accenture Global Solutions Limited Quantum computing machine learning module
US10644882B2 (en) * 2017-04-25 2020-05-05 Bank Of America Corporation Electronic security keys for data security based on quantum particle states
US10990677B2 (en) * 2017-06-05 2021-04-27 Microsoft Technology Licensing, Llc Adversarial quantum machine learning
CN108365953B (zh) * 2018-02-06 2020-06-23 中南大学 基于深度神经网络的自适应差分相移量子密钥分发系统及其实现方法
CN108535878B (zh) * 2018-04-17 2020-06-19 图灵人工智能研究院（南京）有限公司 稳定量子光源的方法及稳定量子光源
CN109547202B (zh) * 2018-12-18 2021-10-22 国科量子通信网络有限公司 量子密钥回收的方法、装置及系统
CN110190959B (zh) * 2019-06-28 2021-05-07 中南大学 基于连续变量量子神经网络的加解密方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006124209A1 (en) * 2005-05-17 2006-11-23 Lucent Technonologies Inc. Phase locking in a multi-channel quantum communication system
CN105490805A (zh) * 2015-11-24 2016-04-13 上海斐讯数据通信技术有限公司 一种基于扩展卡尔曼滤波降低qkd系统误码率的系统及方法
CN107612688A (zh) * 2017-09-30 2018-01-19 上海交通大学 基于机器学习的cvqkd实时性能优化方法及系统
CN110365473A (zh) * 2019-05-31 2019-10-22 南京邮电大学 一种基于机器学习的量子通信系统的主动反馈控制方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LIU, JINGYANG ET AL.: "Practical Phase-Modulation Stabilization in Quantum Key Distribution via Machine Learning", PHYSICAL REVIEW APPLIED 12, 014059 (2019), 30 July 2019 (2019-07-30), XP081382191, DOI: 20200326184816PX *
LIU, WEIQI ET AL.: "Integrating Machine Learning to Achieve an Automatic Parameter Prediction for Practical Continuous-Variable Quantum Key Distribution", PHYSICAL REVIEW A 97, 022316 (2018), 12 February 2018 (2018-02-12), XP055762825, DOI: 20200326172217A *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113572532A (zh) * 2021-04-14 2021-10-29 华南师范大学 光纤干涉环主动相位补偿方法、装置和量子密钥分发系统
CN113572532B (zh) * 2021-04-14 2022-06-28 华南师范大学 光纤干涉环主动相位补偿方法、装置和量子密钥分发系统
CN114511070A (zh) * 2022-04-20 2022-05-17 南京邮电大学 一种适用于双场量子密钥分发系统的零相位电压估计方法

Also Published As

Publication number Publication date
CN110365473A (zh) 2019-10-22
CN110365473B (zh) 2021-11-26
US11817911B2 (en) 2023-11-14
US20220182152A1 (en) 2022-06-09

Similar Documents

Publication Publication Date Title
WO2020238224A1 (zh) 一种基于机器学习的量子通信系统的主动反馈控制方法
Zhang et al. Continuous-variable QKD over 50 km commercial fiber
Argyris et al. Photonic machine learning implementation for signal recovery in optical communications
CN108290733B (zh) 使用玻色子模式进行量子纠错的技术及相关系统和方法
Argyris et al. Comparison of photonic reservoir computing systems for fiber transmission equalization
Liu et al. Practical phase-modulation stabilization in quantum key distribution via machine learning
CN102916807B (zh) 连续变量量子密钥分发系统的偏振补偿实现方法
Katumba et al. Neuromorphic computing based on silicon photonics and reservoir computing
WO2019236250A1 (en) Real-number photonic encoding
US20090074425A1 (en) Optical transmitter and control method thereof
Summers et al. Monolithic InP‐based coherent transmitter photonic integrated circuit with 2.25 Tbit/s capacity
Niu et al. End-to-end deep learning for long-haul fiber transmission using differentiable surrogate channel
Zhang et al. Timing and synchronisation for high‐loss free‐space quantum communication with Hybrid de Bruijn Codes
Tajima et al. Practical quantum cryptosystem for metro area applications
Liu et al. Noise reduction in qubit readout with a two-mode squeezed interferometer
Donaldson et al. Quantum state correction using a measurement-based feedforward mechanism
Djordjevic et al. Discrete variable (dv) qkd
Isoe et al. Capacity upgrade in short-reach optical fibre networks: simultaneous 4-PAM 20 Gbps data and polarization-modulated PPS clock signal using a single VCSEL carrier
Peng et al. A photonics-inspired compact network: toward real-time AI processing in communication systems
Zuo et al. Integrated silicon photonic reservoir computing with pso training algorithm for fiber communication channel equalization
CN113992323B (zh) 一种基于芯片的测量设备无关量子密钥分发系统及方法
Di Candia et al. Two-way covert microwave quantum communication
Vatsavai et al. Silicon photonic microring based chip-scale accelerator for delayed feedback reservoir computing
Fernandez et al. Differentiable machine learning-based modeling for directly-modulated lasers
Zhang et al. Automatic phase compensation of a continuous-variable quantum-key-distribution system via deep learning

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20812678

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20812678

Country of ref document: EP

Kind code of ref document: A1