Deep learning-based satellite-to-ground continuous-variable quantum key distribution system and method
By using a deep learning phase estimation module to perform phase correction based on intensity distribution images, the phase distortion problem caused by atmospheric turbulence in the satellite-to-ground continuous variable quantum key distribution system is solved. This achieves low-cost, high-reliability real-time correction, improving system performance and robustness.
Patent Information
- Application Number
- CN202610836907.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-10
- Publication Date
- 2026-07-24
Smart Images

Figure CN122457255A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to quantum key distribution, and more specifically to a satellite-to-ground continuous variable quantum key distribution system and method based on deep learning. Background Technology
[0002] Quantum key distribution is theoretically an absolutely secure communication method. Continuous-Variable Quantum Key Distribution (CV-QKD) can achieve high-speed transmission using standard optical communication devices, but in satellite-to-ground links, the signal needs to traverse a complex atmospheric turbulence environment. Atmospheric turbulence causes dynamic and random distortions in the wavefront of light waves, severely disrupting the coherence between the quantum signal at the satellite end and the local oscillator at the ground station, leading to a sharp drop in the secure key rate, even to zero.
[0003] To restore coherence, existing research (e.g., the studies by Kish SP et al. at ICC 2021 and Villasenor E et al. at GLOBECOM 2020) typically employs a "real local oscillator + direct phase measurement of the reference pulse" technique. The basic steps of this technique are as follows: at a ground station, the phase wavefront distribution of the classical reference pulse transmitted along with the quantum signal is directly measured using sophisticated equipment such as a Shaker-Hartmann wavefront sensor; subsequently, phase distortion information is extracted from the sensor data using phase reconstruction algorithms such as Zernike polynomial fitting; finally, this information is used to modulate the phase of the real local oscillator to match the wavefront of the quantum signal. However, this technique has the following inherent drawbacks:
[0004] 1) The system is complex and costly: Wavefront sensors are precision and expensive optical instruments, which greatly increases the hardware complexity and deployment cost of ground stations;
[0005] 2) Poor real-time performance: Phase reconstruction algorithms (such as Zernike polynomial fitting) have high computational complexity, making it difficult to meet the millisecond-level real-time correction requirements of high-speed satellite-to-ground communication (Knapek M, doctoral dissertation, 2011).
[0006] 3) Performance bottleneck under strong turbulence: Under strong turbulence conditions, the measurement signal-to-noise ratio of the wavefront sensor itself will be severely degraded, leading to the failure of phase estimation and forming the "ceiling" of the entire system performance;
[0007] 4) Potential security risks: The complex measurement and calibration process may provide new attack surfaces for eavesdropper (such as reference pulse attacks).
[0008] Meanwhile, a long-standing technological bias in this field holds that to achieve high-precision phase correction, direct phase measurement is necessary first. This bias has prevented satellite-to-ground CV-QKD systems from escaping their reliance on complex phase measurement equipment, severely hindering their development towards large-scale, low-cost, and high-reliability systems. Summary of the Invention
[0009] (a) Technical problems to be solved
[0010] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a satellite-to-ground continuous variable quantum key distribution system and method based on deep learning, which can effectively overcome the shortcomings of the existing technology in accurately and in real time correcting the phase distortion caused by atmospheric turbulence in the satellite-to-ground link.
[0011] (II) Technical Solution
[0012] To achieve the above objectives, the present invention provides the following technical solution:
[0013] The deep learning-based satellite-to-ground continuous variable quantum key distribution system includes a signal receiving and splitting module, an intensity measurement module, a deep learning phase estimation module, a phase modulation module, and a coherent detection and key generation module deployed at the ground station.
[0014] The signal receiving and splitting module receives multiplexed optical signals from the satellite, which contain quantum signals and classical reference pulses, and separates the quantum signals and classical reference pulses to the coherent detection and key generation module and the intensity measurement module, respectively.
[0015] The intensity measurement module acquires a two-dimensional intensity distribution image of the classical reference pulse;
[0016] The deep learning phase estimation module has a built-in pre-trained deep learning model. It takes a two-dimensional intensity distribution image as input and outputs a two-dimensional phase correction matrix with the same resolution as the input image.
[0017] The phase modulation module performs real-time phase correction on the locally generated light wavefront based on a two-dimensional phase correction value matrix to compensate for phase distortion caused by atmospheric turbulence.
[0018] The coherent detection and key generation module performs zero-difference detection on the corrected optical wavefront and quantum signal to complete coherent detection, and generates a secure key through post-processing.
[0019] Preferably, the intensity measurement module acquires a two-dimensional intensity distribution image of a classical reference pulse through an area array photodetector.
[0020] Preferably, the deep learning model is built on a convolutional neural network (CNN) with an encoder-decoder architecture. The model is trained using a large amount of satellite-to-ground channel simulation data to learn the complex nonlinear mapping relationship between the intensity distortion mode caused by atmospheric turbulence and the corresponding phase distortion mode. The pre-trained deep learning model is deployed to the deep learning phase estimation module. During actual satellite-to-ground communication, it performs forward inference on the real-time acquired two-dimensional intensity distribution image and quickly outputs the corresponding two-dimensional phase correction matrix.
[0021] Preferably, the deep learning model is constructed using a convolutional neural network (CNN) with an encoder-decoder architecture based on SegNet;
[0022] The encoder consists of 13 convolutional layers and 3 max pooling layers. Its main structure is based on VGG16 and it is responsible for extracting multi-level features from the two-dimensional intensity distribution image and compressing them into a highly abstract feature vector.
[0023] The decoder consists of 13 convolutional layers and 3 upsampling layers. Its main structure is a mirror image of the encoder. It is responsible for decoding the feature vectors and reconstructing them into a two-dimensional phase correction matrix with the same resolution as the two-dimensional intensity distribution image.
[0024] Preferably, the deep learning model is trained in an offline, large-scale simulation environment. A large amount of satellite-to-ground channel simulation data is generated by simulating the phase screen method, which includes tens of thousands of data pairs of "turbulence intensity map-phase distortion map". The goal of model training is to minimize the mean square error between the predicted phase and the true phase.
[0025] Preferably, the phase modulation module performs real-time phase correction on the light wavefront generated by the local oscillator through a spatial light modulator based on a two-dimensional phase correction value matrix to compensate for phase distortion caused by atmospheric turbulence.
[0026] Preferably, the coherent detection and key generation module performs coherent detection on the corrected light wavefront and quantum signal using a zero-difference detector.
[0027] A deep learning-based satellite-to-ground continuous-variable quantum key distribution method includes the following steps:
[0028] S1. Receive the multiplexed optical signal from the satellite, which contains quantum signals and classical reference pulses, and separate the quantum signals and classical reference pulses;
[0029] S2. Obtain the two-dimensional intensity distribution image of the classical reference pulse;
[0030] S3. Input the two-dimensional intensity distribution image into the pre-trained deep learning model and output a two-dimensional phase correction matrix with the same resolution as the input image.
[0031] S4. Based on the two-dimensional phase correction value matrix, perform real-time phase correction on the locally generated light wavefront to compensate for phase distortion caused by atmospheric turbulence.
[0032] S5. Perform zero-difference detection on the corrected optical wavefront and quantum signal to complete coherent detection, and generate a secure key through post-processing.
[0033] Preferably, the deep learning model is built on a convolutional neural network (CNN) with an encoder-decoder architecture. The model is trained using a large amount of satellite-to-ground channel simulation data to learn the complex nonlinear mapping relationship between the intensity distortion mode caused by atmospheric turbulence and the corresponding phase distortion mode. The pre-trained deep learning model is deployed to the deep learning phase estimation module. During actual satellite-to-ground communication, it performs forward inference on the real-time acquired two-dimensional intensity distribution image and quickly outputs the corresponding two-dimensional phase correction matrix.
[0034] Preferably, the deep learning model is constructed using a convolutional neural network (CNN) with an encoder-decoder architecture based on SegNet;
[0035] The encoder consists of 13 convolutional layers and 3 max pooling layers. Its main structure is based on VGG16 and it is responsible for extracting multi-level features from the two-dimensional intensity distribution image and compressing them into a highly abstract feature vector.
[0036] The decoder consists of 13 convolutional layers and 3 upsampling layers. Its main structure is a mirror image of the encoder. It is responsible for decoding the feature vectors and reconstructing them into a two-dimensional phase correction matrix with the same resolution as the two-dimensional intensity distribution image.
[0037] (III) Beneficial Effects
[0038] Compared with existing technologies, the deep learning-based satellite-to-ground continuous variable quantum key distribution system and method provided by this invention have the following advantages:
[0039] 1) A leap in system complexity and cost: The expensive and sophisticated wavefront sensor has been completely eliminated, and the core equipment of the ground station has been simplified to a conventional array detector and a deep learning inference unit, resulting in a significant reduction in hardware cost and maintenance difficulty.
[0040] 2) Significant improvement in performance and robustness: Deep learning models have a higher tolerance for noise and distortion. Experiments have shown that under strong turbulent conditions where the performance of traditional wavefront sensors deteriorates, this invention can still maintain effective phase estimation capabilities and improve coherence efficiency from near zero to 0.53, breaking through the performance bottleneck of existing technologies.
[0041] 3) Achieve true real-time high-performance processing: The forward inference time of the pre-trained deep learning model is in the millisecond range, which can fully meet the millisecond-level real-time correction requirements of high-speed satellite-to-ground communication, and fundamentally solve the computational delay problem caused by complex phase reconstruction algorithms.
[0042] 4) Opening up new technical paths: This invention verifies the feasibility of "intensity is phase" in specific application scenarios, and provides a brand-new technical paradigm for signal processing in the entire field of free space quantum communication. Attached Figure Description
[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0044] Figure 1 This is a schematic diagram illustrating the working principle of the space-to-ground CV-QKD system in this invention;
[0045] Figure 2 This is a schematic diagram illustrating the working principle of the deep learning model in this invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0047] The core of this invention lies in breaking the technical bias of "no phase correction without phase measurement" and proposing a new paradigm of indirect phase correction based on "intensity as phase." In atmospheric turbulence channel models, there is a strong correlation between the spatial distribution distortion of light intensity caused by atmospheric turbulence and the resulting spatial distribution distortion of phase. This correlation can be learned and predicted with high precision using deep learning models. During the model training phase, this hidden mapping relationship is learned end-to-end using massive amounts of "turbulence intensity map-phase distortion map" data pairs (generated from satellite-ground channel simulation). During the model deployment phase, the model uses the real-time measured intensity distortion patterns as "fingerprints" to quickly infer the corresponding phase distortion patterns, thereby achieving precise pre-distortion modulation of the RLO and restoring its coherence with the quantum signal.
[0048] The following describes the specific functional modules of the deep learning-based satellite-to-ground continuous variable quantum key distribution system provided by this invention, using concrete examples (such as...). Figure 1 (As shown) and its technical effects. The system functional modules include: a signal receiving and splitting module, an intensity measurement module, a deep learning phase estimation module, a phase modulation module, and a coherent detection and key generation module deployed at the ground station;
[0049] The signal receiving and splitting module receives multiplexed optical signals from the satellite, which contain quantum signals and classical reference pulses, and separates the quantum signals and classical reference pulses to the coherent detection and key generation module and the intensity measurement module, respectively.
[0050] The intensity measurement module acquires a two-dimensional intensity distribution image of the classical reference pulse;
[0051] The deep learning phase estimation module has a built-in pre-trained deep learning model. It takes a two-dimensional intensity distribution image as input and outputs a two-dimensional phase correction matrix with the same resolution as the input image.
[0052] The phase modulation module performs real-time phase correction on the locally generated light wavefront based on a two-dimensional phase correction value matrix to compensate for phase distortion caused by atmospheric turbulence.
[0053] The coherent detection and key generation module performs zero-difference detection on the corrected optical wavefront and quantum signal to complete coherent detection, and generates a secure key through post-processing.
[0054] I. Strength Measurement Module
[0055] The intensity measurement module acquires a two-dimensional intensity distribution image of the classical reference pulse through an area array photodetector.
[0056] II. Deep Learning Phase Estimation Module
[0057] The deep learning model is built on a convolutional neural network (CNN) with an encoder-decoder architecture. The model is trained using a large amount of satellite-to-ground channel simulation data to learn the complex nonlinear mapping relationship between the intensity distortion mode caused by atmospheric turbulence and the corresponding phase distortion mode. The pre-trained deep learning model is deployed to the deep learning phase estimation module. In actual satellite-to-ground communication, it performs forward inference on the real-time acquired two-dimensional intensity distribution image and quickly outputs the corresponding two-dimensional phase correction matrix.
[0058] In the technical solution of this application, the deep learning model is constructed using a convolutional neural network (CNN) with an encoder-decoder architecture based on SegNet, such as... Figure 2 As shown;
[0059] The encoder consists of 13 convolutional layers and 3 max pooling layers. Its main structure is based on VGG16 and it is responsible for extracting multi-level features from the two-dimensional intensity distribution image and compressing them into a highly abstract feature vector.
[0060] The decoder consists of 13 convolutional layers and 3 upsampling layers. Its main structure is a mirror image of the encoder. It is responsible for decoding the feature vectors and reconstructing them into a two-dimensional phase correction matrix with the same resolution as the two-dimensional intensity distribution image.
[0061] The deep learning model is trained in an offline, large-scale simulation environment. A large amount of satellite-to-ground channel simulation data is generated by simulating the phase screen method, which includes tens of thousands of data pairs of "turbulence intensity map-phase distortion map". The goal of model training is to minimize the mean square error between the predicted phase and the true phase.
[0062] III. Phase Modulation Module
[0063] The phase modulation module performs real-time phase correction on the light wavefront generated by the local oscillator through a spatial light modulator based on a two-dimensional phase correction value matrix to compensate for phase distortion caused by atmospheric turbulence.
[0064] IV. Coherent Detection and Key Generation Module
[0065] The coherent detection and key generation module performs coherent detection on the corrected light wavefront and quantum signal using a zero-difference detector.
[0066] Based on the aforementioned deep learning-based satellite-to-ground continuous variable quantum key distribution system, this invention also discloses a deep learning-based satellite-to-ground continuous variable quantum key distribution method, comprising the following steps:
[0067] S1. Receive the multiplexed optical signal from the satellite, which contains quantum signals and classical reference pulses, and separate the quantum signals and classical reference pulses;
[0068] S2. Obtain the two-dimensional intensity distribution image of the classical reference pulse;
[0069] S3. Input the two-dimensional intensity distribution image into the pre-trained deep learning model and output a two-dimensional phase correction matrix with the same resolution as the input image.
[0070] S4. Based on the two-dimensional phase correction value matrix, perform real-time phase correction on the locally generated light wavefront to compensate for phase distortion caused by atmospheric turbulence.
[0071] S5. Perform zero-difference detection on the corrected optical wavefront and quantum signal to complete coherent detection, and generate a secure key through post-processing.
[0072] In the technical solution of this application, the deep learning model is built based on the convolutional neural network (CNN) with an encoder-decoder architecture. The model is trained with a large amount of satellite-to-ground channel simulation data to learn the complex nonlinear mapping relationship between the intensity distortion mode caused by atmospheric turbulence and the corresponding phase distortion mode. The pre-trained deep learning model is deployed to the deep learning phase estimation module. In actual satellite-to-ground communication, it performs forward inference on the real-time acquired two-dimensional intensity distribution image and quickly outputs the corresponding two-dimensional phase correction value matrix.
[0073] The deep learning model is constructed using a convolutional neural network (CNN) with an encoder-decoder architecture based on SegNet.
[0074] The encoder consists of 13 convolutional layers and 3 max pooling layers. Its main structure is based on VGG16 and it is responsible for extracting multi-level features from the two-dimensional intensity distribution image and compressing them into a highly abstract feature vector.
[0075] The decoder consists of 13 convolutional layers and 3 upsampling layers. Its main structure is a mirror image of the encoder. It is responsible for decoding the feature vectors and reconstructing them into a two-dimensional phase correction matrix with the same resolution as the two-dimensional intensity distribution image.
[0076] To better illustrate the technical solution of this application, the following detailed explanation is provided in conjunction with three experiments.
[0077] Experiment 1: Performance Verification under Typical Satellite-Ground Channel
[0078] • Experimental objective: To verify the basic performance of the present invention under typical moderate to strong turbulent star-to-ground channels.
[0079] • Experimental scenario and parameter settings:
[0080] 1) Satellite-to-ground link:
[0081] Satellite orbital altitude: 300km;
[0082] Ground station elevation: 2km;
[0083] Turbulence intensity: Moderate to strong turbulence conditions;
[0084] Transmission wavelength: 1550nm;
[0085] 2) Deep learning model configuration:
[0086] Network architecture: SegNet (encoder-decoder architecture, based on VGG16 backbone).
[0087] Input / output image resolution: 64×64 pixels;
[0088] Training set: 57,600 pairs of "turbulence intensity map-phase true map" data, generated by simulation using the phase screen method;
[0089] Test set: 6,400 independent simulation data sets, covering different turbulence intensities and noise levels;
[0090] 3) Compare to the baseline:
[0091] Traditional method: Phase reconstruction method based on Zernike polynomial fitting, using an ideal wavefront sensor to measure the phase;
[0092] Uncorrected: No phase correction is performed, serving as the lower limit of performance.
[0093] • Experimental steps:
[0094] 1) Model training:
[0095] Using the Adam optimizer, learning rate = 0.001, batch size = 32, train for 100 epochs;
[0096] Loss function: the mean square error between the predicted phase and the true phase;
[0097] 2) Performance evaluation:
[0098] The method of this invention and two baseline methods were run on the test set, and the coherence efficiency, mutual information between Alice and Bob, and security key rate were recorded.
[0099] •Experimental Results:
[0100] Table 1. Performance experimental results under typical moderate to strong turbulence star-to-ground channels.
[0101]
[0102] •Experimental conclusions:
[0103] This invention improves the coherence efficiency from near zero to 0.53 under moderate to strong turbulence conditions, which is significantly better than existing solutions.
[0104] The system ultimately achieved a secure key rate of 0.112 bits / pulse, proving that the present invention can not only achieve phase correction, but also has the ability to generate actual keys.
[0105] This verifies the technical feasibility of "achieving high-precision phase estimation based solely on intensity information".
[0106] Experiment 2: Robustness Test under Different Turbulence Intensities
[0107] • Experimental objective: To verify the adaptability and robustness of the present invention under different atmospheric turbulence intensities, covering typical satellite-to-ground communication scenarios from weak to extremely strong turbulence.
[0108] • Experiment setup:
[0109] Turbulence intensity range: ;
[0110] Use a uniformly trained SegNet model without retraining or fine-tuning.
[0111] Comparison method: The phase reconstruction method based on Zernike polynomial fitting uses an ideal wavefront sensor to measure the phase.
[0112] •Experimental Results:
[0113] Under weak turbulent conditions, the performance of this invention is close to that of existing solutions;
[0114] In strong turbulence ( When the existing solution is used, the performance drops sharply, while the present invention still maintains a high coherence efficiency (>0.4).
[0115] This demonstrates that the present invention possesses strong generalization ability and turbulence robustness, and is particularly suitable for dynamically changing satellite-to-ground channels.
[0116] Experiment 3: System Real-Time Verification and Hardware Deployment Testing
[0117] • Experimental objective: To verify the real-time processing capability of this invention on a real hardware platform.
[0118] • Experiment setup:
[0119] Hardware platform: NVIDIA Jetson Xavier edge computing module;
[0120] Input source: Simulated real-time generated intensity image stream (10fps);
[0121] Task: Continuously perform phase prediction and calculate end-to-end processing latency.
[0122] •Experimental Results:
[0123] Average processing time per frame: 15.2ms;
[0124] Maximum sustained frame rate: 65fps;
[0125] System power consumption: <20W.
[0126] •Experimental conclusions:
[0127] This invention can meet the millisecond-level real-time correction requirements of high-speed satellite-to-ground communication;
[0128] It can be deployed on low-power edge devices and is suitable for embedded integration in space or ground stations.
[0129] It provides a hardware implementation path for the lightweight and practical application of the space-to-ground CV-QKD system.
[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A deep learning-based satellite-to-ground continuous-variable quantum key distribution system, characterized in that: It includes a signal receiving and splitting module, an intensity measurement module, a deep learning phase estimation module, a phase modulation module, and a coherent detection and key generation module deployed at ground stations; The signal receiving and splitting module receives multiplexed optical signals from the satellite, which contain quantum signals and classical reference pulses, and separates the quantum signals and classical reference pulses to the coherent detection and key generation module and the intensity measurement module, respectively. The intensity measurement module acquires a two-dimensional intensity distribution image of the classical reference pulse; The deep learning phase estimation module has a built-in pre-trained deep learning model. It takes a two-dimensional intensity distribution image as input and outputs a two-dimensional phase correction matrix with the same resolution as the input image. The phase modulation module performs real-time phase correction on the locally generated light wavefront based on a two-dimensional phase correction value matrix to compensate for phase distortion caused by atmospheric turbulence. The coherent detection and key generation module performs zero-difference detection on the corrected optical wavefront and quantum signal to complete coherent detection, and generates a secure key through post-processing.
2. The deep learning-based satellite-to-ground continuous-variable quantum key distribution system according to claim 1, characterized in that: The intensity measurement module acquires a two-dimensional intensity distribution image of the classical reference pulse through an area array photodetector.
3. The deep learning-based satellite-to-ground continuous-variable quantum key distribution system according to claim 1, characterized in that: The deep learning model is built on a convolutional neural network (CNN) with an encoder-decoder architecture. It is trained using a large amount of satellite-to-ground channel simulation data to learn the complex nonlinear mapping relationship between the intensity distortion mode caused by atmospheric turbulence and the corresponding phase distortion mode. The pre-trained deep learning model is deployed to the deep learning phase estimation module. During actual satellite-to-ground communication, it performs forward inference on the real-time acquired two-dimensional intensity distribution image and quickly outputs the corresponding two-dimensional phase correction matrix.
4. The deep learning-based satellite-to-ground continuous-variable quantum key distribution system according to claim 3, characterized in that: The deep learning model is constructed using a convolutional neural network (CNN) with an encoder-decoder architecture based on SegNet. The encoder consists of 13 convolutional layers and 3 max pooling layers. Its main structure is based on VGG16 and it is responsible for extracting multi-level features from the two-dimensional intensity distribution image and compressing them into a highly abstract feature vector. The decoder consists of 13 convolutional layers and 3 upsampling layers. Its main structure is a mirror image of the encoder. It is responsible for decoding the feature vectors and reconstructing them into a two-dimensional phase correction matrix with the same resolution as the two-dimensional intensity distribution image.
5. The deep learning-based satellite-to-ground continuous-variable quantum key distribution system according to claim 4, characterized in that: The deep learning model is trained in an offline, large-scale simulation environment. A large amount of satellite-to-ground channel simulation data is generated by simulating the phase screen method, which includes tens of thousands of data pairs of "turbulence intensity map-phase distortion map". The goal of model training is to minimize the mean square error between the predicted phase and the true phase.
6. The deep learning-based satellite-to-ground continuous-variable quantum key distribution system according to claim 1, characterized in that: The phase modulation module performs real-time phase correction on the light wavefront generated by the local oscillator through a spatial light modulator based on a two-dimensional phase correction value matrix, in order to compensate for phase distortion caused by atmospheric turbulence.
7. The deep learning-based satellite-to-ground continuous-variable quantum key distribution system according to claim 1, characterized in that: The coherent detection and key generation module performs coherent detection on the corrected light wavefront and quantum signal using a zero-difference detector.
8. A deep learning-based satellite-to-ground continuous variable quantum key distribution method, adapted to the deep learning-based satellite-to-ground continuous variable quantum key distribution system as described in claim 1, characterized in that: Includes the following steps: S1. Receive the multiplexed optical signal from the satellite, which contains quantum signals and classical reference pulses, and separate the quantum signals and classical reference pulses; S2. Obtain the two-dimensional intensity distribution image of the classical reference pulse; S3. Input the two-dimensional intensity distribution image into the pre-trained deep learning model and output a two-dimensional phase correction matrix with the same resolution as the input image. S4. Based on the two-dimensional phase correction value matrix, perform real-time phase correction on the locally generated light wavefront to compensate for phase distortion caused by atmospheric turbulence. S5. Perform zero-difference detection on the corrected optical wavefront and quantum signal to complete coherent detection, and generate a secure key through post-processing.
9. The deep learning-based satellite-to-ground continuous variable quantum key distribution method according to claim 8, characterized in that: The deep learning model is built on a convolutional neural network (CNN) with an encoder-decoder architecture. It is trained using a large amount of satellite-to-ground channel simulation data to learn the complex nonlinear mapping relationship between the intensity distortion mode caused by atmospheric turbulence and the corresponding phase distortion mode. The pre-trained deep learning model is deployed to the deep learning phase estimation module. During actual satellite-to-ground communication, it performs forward inference on the real-time acquired two-dimensional intensity distribution image and quickly outputs the corresponding two-dimensional phase correction matrix.
10. The deep learning-based satellite-to-ground continuous variable quantum key distribution method according to claim 9, characterized in that: The deep learning model is constructed using a convolutional neural network (CNN) with an encoder-decoder architecture based on SegNet. The encoder consists of 13 convolutional layers and 3 max pooling layers. Its main structure is based on VGG16 and it is responsible for extracting multi-level features from the two-dimensional intensity distribution image and compressing them into a highly abstract feature vector. The decoder consists of 13 convolutional layers and 3 upsampling layers. Its main structure is a mirror image of the encoder. It is responsible for decoding the feature vectors and reconstructing them into a two-dimensional phase correction matrix with the same resolution as the two-dimensional intensity distribution image.
