Large-scale MIMO key generation method based on channel mapping and compression quantization

Abstract

The invention relates to a large-scale MIMO (Multiple Input Multiple Output) key generation method based on channel mapping and compression quantization, which specifically comprises the following steps of: aiming at the problem that the channel mapping accuracy of the existing model is insufficient, providing a CoSTMNet model based on a hybrid codec, which is used for modeling a channel mapping function which cannot be mathematically solved and improving the channel mapping accuracy; in order to solve the problem of insufficient key consistency of the existing FDD key scheme, the invention provides a double-bit quantization algorithm PCGQ based on partition compression and gray code coding, and designs a full-process physical layer key generation scheme in an FDD mode in combination with a CoSTMNet channel mapping model. Compared with a comparison object, the algorithm scheme provided by the invention has better performance in the aspects of normalized mean square error, Pearson correlation coefficient, key consistency and other indexes.

Description

Technical Field

[0001] This invention relates to the field of physical layer security, and more particularly to physical layer keys and large-scale MIMO systems, specifically a method for generating large-scale MIMO keys based on channel mapping and compression quantization. Background Technology

[0002] With the rapid development of 5G and future 6G communication technologies, wireless systems face increasingly severe security challenges. How to effectively ensure the security of wireless systems is a pressing issue. Traditional secure communication solutions mainly rely on classical encryption algorithms, such as symmetric and asymmetric encryption algorithms. However, these methods depend on public key infrastructure and trusted third parties for key distribution and management, resulting in high maintenance costs and susceptibility to single points of failure. Furthermore, breakthroughs in quantum computing also pose a substantial threat to the security of traditional encryption algorithms. Against this backdrop, physical layer key generation technology, as an emerging physical layer security solution, can utilize the inherent randomness of wireless channels to directly generate shared keys for legitimate communicators, providing a more secure and efficient encryption method.

[0003] Most existing physical layer key research is based on ideal wireless communication environment assumptions, failing to consider the impact of non-ideal factors such as signal blocking, interference, and fading in complex real-world scenarios. In real-world wireless network environments, signals are easily affected by buildings, terrain, and various moving obstacles. In such complex environments, some research has introduced additional relay nodes or random parameters to enhance the performance of key schemes, but this leads to higher power consumption and computational complexity, thus limiting its application scenarios. Furthermore, with the continuous increase in communication frequencies in 5G and future 6G systems, signal attenuation will become more significant, making them more susceptible to blocking or interference from various obstacles. In addition, Massive MIMO, as a key technology in 5G and 6G, has received relatively little research on physical layer key generation. Therefore, how to design an efficient and secure physical layer key generation scheme for the complex wireless network environments of today's 5G and future 6G remains an open problem that urgently needs to be solved.

[0004] Unlike TDD (Time Division Duplex) Massive MIMO systems, obtaining reciprocal wireless channel characteristics for key generation in Frequency Division Duplex (FDD) Massive MIMO systems presents a significant challenge. In TDD systems, uplink and downlink use the same frequency band for data transmission, resulting in high natural reciprocity of the CSI (Channel Signal Indicator) between the communicating parties. However, in FDD Massive MIMO systems, uplink and downlink use different frequency bands and experience different channel fading, leading to a loss of natural reciprocity in their CSI information. Furthermore, other reciprocal channel parameters in TDD Massive MIMO, such as Received Signal Strength (RSS), envelope, and channel phase, may differ significantly in FDD Massive MIMO systems. Given that FDD Massive MIMO is widely used in various wireless scenarios in today's 5G communication architecture, researching information-theoretically secure physical layer key technologies for FDD Massive MIMO systems holds significant theoretical and practical value.

[0005] Currently, research on physical layer key schemes in FDD frequency division duplex mode can be mainly divided into two categories: key schemes based on prior construction and key schemes based on deep learning.

[0006] The core of key schemes based on prior construction is to extract frequency-independent channel features or construct reciprocal channel parameters to generate keys. Current research has proposed quantization algorithms and error negotiation schemes based on packet shifting in FDD mode, as well as FDD key schemes designed using the eigenvalues ​​of the uplink and downlink channel covariance matrices. While these methods have some feasibility, they often require large bandwidth or special antenna configurations, and are difficult to accurately separate channel paths in complex multipath environments. Another type of scheme based on loopback utilizes additional reverse channel training and feedback to construct reciprocal channel parameters, but this type of scheme significantly increases the complexity of channel detection and system computational overhead.

[0007] The core of deep learning-based key generation schemes lies in learning and extracting wireless channel characteristics through neural networks, without relying too heavily on system models or prior knowledge. Existing research has shown that downlink channel information can be inferred to some extent by analyzing the uplink channel in FDD systems. Subsequent research has proposed domain adversarial networks and convolutional autoencoder models to predict downlink channel state information, and further constructed corresponding key generation schemes.

[0008] However, existing research is mostly limited to traditional FDD SISO or MIMO systems, and research on key schemes for FDD Massive MIMO in 5G and future 6G communications is still relatively lacking. Furthermore, current research is largely based on idealized wireless channel environment assumptions, making it difficult to guarantee the consistency and security of key generation in complex environments and non-ideal scenarios such as signal congestion. Summary of the Invention

[0009] The purpose of this invention is to propose a large-scale MIMO key generation method based on channel mapping and compression quantization, thereby improving the accuracy of model channel mapping and further reducing the initial key inconsistency rate of existing FDD key schemes.

[0010] The objective of this invention is achieved as follows:

[0011] For FDD Massive MIMO systems in complex environments, this proposal addresses the issue of insufficient channel mapping accuracy in existing models by proposing a CoSTMNet model based on a hybrid codec. This model models channel mapping functions that cannot be mathematically solved, thereby improving channel mapping accuracy. Furthermore, to address the key consistency issues in existing FDD keying schemes, this proposal presents a PCGQ two-bit quantization algorithm based on partitioned compression and gray code encoding. Combined with CoSTMNet, a complete physical layer key generation scheme for FDD mode is designed.

[0012] The specific method is as follows:

[0013] A method for generating large-scale MIMO keys based on channel mapping and compressed quantization includes the following steps:

[0014] Step 1: Obtain the channel reciprocity characteristics of the communicating parties through the CoSTMNet channel mapping model. Using CoSTMNet, Alice can estimate the center frequency f it detected. A uplink channel H A The mapping is to the center frequency f at the Bob end during the coherence time. B downlink channel H B In other words, in an FDD Massive MIMO system, Alice and Bob do not need additional information exchange. Instead, they can obtain the channel reciprocity characteristics through CoSTMNet and use this as the basis for subsequent key generation between Alice and Bob. CoSTMNet adopts an encoder-decoder architecture. The encoder part converts the input H... A (f AThe encoder maps the signal to a latent representation space, thereby extracting frequency-independent features. Simultaneously, noise removal and redundant information compression are also performed in the encoder. The decoder then performs frequency alignment and information reconstruction, reconstructing the latent representation output by the encoder into the target channel H. B (f B To minimize the error between the reconstructed results and the real channel, the CSI data of Alice and Bob needs to be preprocessed before CoSTMNet model training to improve the stability of model training and accelerate convergence. First, the amplitude and phase in the CSI are separated, and then amplitude and phase are normalized. After data preprocessing, a CSI pair dataset is constructed. In each CSI pair, there is a dataset containing Alice's center frequency f. A H A And the center frequency of Bob's end during the coherence time is f B H B The CoSTMNet model is trained using supervised learning. A (f A H is used as the model input. B (f B The true value label is then used. The model is trained using the MSE (mean squared error) loss function, taking into account both amplitude and phase mapping errors.

[0015] Step 2: Quantize the reciprocal channel data of both communicating parties using the PCGQ (Peripherally Compressed and Gray Coded) algorithm to obtain the initial key bit string for both parties. Using the CoSTMNet channel mapping model from the previous step, Alice and Bob can obtain the reciprocal channel information (CSI). The dimension of CSI is the number of receiver antennas × the number of transmitter antennas × the number of subcarriers. In an FDD Massive MIMO channel model with a large number of antennas, the number of elements in CSI will be much larger than the key length. Therefore, this invention designs the PCGQ channel compression quantization algorithm to convert CSI into the initial key bit string. Setting the final generated physical layer key length to 1024 bits, the initial key bit string length is set to L = 2048. The PCGQ algorithm process includes CSI amplitude and phase fusion, global statistics calculation and CSI partitioning, partitioned compression quantization, and gray code mapping.

[0016] Step 3: Complete the physical layer key generation for both communicating parties using the designed FDD mode full-process physical layer key generation scheme. The key generation scheme includes steps such as channel sounding and estimation, preprocessing and channel mapping, compression and quantization, information negotiation and privacy amplification.

[0017] Step 3.1: Channel sounding estimation. The channel sounding and estimation at Alice and Bob's ends uses the least squares method. After channel sounding estimation, Alice and Bob can respectively obtain the uplink channel state information H. A With downlink channel state information H B In FDD Massive MIMO systems, H A The center frequency is f A H B The center frequency is f B And f A ≠f B Due to the difference in their frequencies, Alice measured H... A H measured by Bob B Reciprocity is no longer present; the construction of reciprocal information between the two parties will be carried out in subsequent processes.

[0018] Step 3.2, Preprocessing and Channel Mapping: In this step, the reciprocity channel information for Alice and Bob in FDD mode is constructed. First, the H values ​​estimated by the probes of Alice and Bob are processed separately. A With H B Preprocessing is performed, including amplitude-phase separation and amplitude-phase normalization. Further, the Alice end performs channel mapping using the CoSTMNet model designed in Section 4.3. The result of the CoSTMNet channel mapping at the Alice end is denoted as H. B Through channel mapping, the Alice terminal can use a center frequency of f. A H A To indirectly estimate the center frequency as f B H B This allows for the construction of reciprocal channel information between Alice and Bob, which is then used as the basis for subsequent key generation processes.

[0019] Step 3.3: In this step, the PCGQ two-bit channel quantization algorithm from Step 2 is used to convert the reciprocity channel information of Alice and Bob into their initial key bit strings. Specifically, Alice converts the H obtained after CoSTMNet channel mapping... B Bob performs PCGQ channel compression quantization, while Bob directly processes the preprocessed H... B Perform PCGQ compression quantization. After compression quantization, both parties obtain their respective initial keys, denoted as q. A With q B Because the reciprocity of Alice and Bob's channel information is not perfect, therefore, in the initial key q... A With q BThere is a certain key inconsistency rate (KDR) between the two parties. In subsequent information negotiation processes, the inconsistency in the initial keys of both parties will be eliminated.

[0020] Step 3.4, Information Negotiation: The main function of information negotiation is to correct inconsistencies in the initial key bit strings of Alice and Bob, reducing the key inconsistency rate (KDR) of the initial key strings to 0, thus obtaining identical key bit streams. As seen in the previous quantization process, despite the isolation protection band within the quantization interval, inconsistencies in the feature vectors of Alice and Bob inevitably remain after quantization. Therefore, it is necessary to correct these inconsistencies using an error correction algorithm, while ensuring that only a very small amount of key information is leaked on the public channel. Commonly used error correction algorithms in key information negotiation include the Cascade block-based interactive error correction protocol and algorithms based on error-correcting code schemes, such as BCH codes, Polar codes, and Turbo codes. This study uses the Cascade error correction protocol algorithm for information negotiation of the initial key between Alice and Bob. The main idea of ​​the Cascade error correction protocol is to divide the initial key into blocks and exchange the hash values ​​of the blocks in each round of interaction. If the hash values ​​are inconsistent, binary recursion is used to further divide the blocks to quickly locate and correct errors until all inconsistencies are finally eliminated.

[0021] Step 3.5, Privacy Amplification: The main function of privacy amplification is to further strengthen the consensus key sequence obtained through information negotiation using cryptographic security functions, and generate the final physical layer key for Alice and Bob. During the aforementioned information negotiation step, the eavesdropper Eve may obtain some error correction information. Privacy amplification transforms the consensus sequence into a 1024-bit final key using a secure hash function. Due to the avalanche effect and collision resistance of cryptographic hashes, the final generated physical layer key can be considered an unpredictable uniform random sequence, thus eliminating the risk of information leakage during key negotiation. Commonly used secure hash functions in physical layer key generation include SHAKE-128 / 256, SHA-256 / 512, or iterative construction based on HMAC. This study uses the SHAKE-256 secure hash function for privacy amplification. SHAKE-256 can output hash results of arbitrary lengths according to actual needs. Here, the output length of SHAKE-256 is set to 1024, meaning the final generated physical layer key is 1024 bits long.

[0022] The positive effects of this invention are:

[0023] This invention addresses RIS-assisted FDD Massive MIMO systems, designing a full-process physical layer key generation scheme based on channel mapping and partitioned compression quantization. First, a CoSTMNet model based on a hybrid codec is proposed to model the channel mapping function and improve its accuracy. The CoSTMNet model exhibits superior performance in terms of normalized mean square error and Pearson correlation coefficient, indicating higher channel mapping accuracy. Subsequently, this invention proposes a two-bit quantization algorithm, PCGQ, based on partitioned compression and gray code encoding, to efficiently convert the reciprocal channel information of Alice and Bob into a low-KDR initial key. Combining the CoSTMNet channel mapping model and the PCGQ quantization algorithm, this invention designs a full-process physical layer key generation scheme in FDD mode. The designed scheme demonstrates superior initial key consistency performance and passes the NIST randomness test, indicating that the generated key meets cryptographic randomness security requirements. Attached Figure Description

[0024] Figure 1 It is the CoSTMNet channel mapping model structure.

[0025] Figure 2 This is a schematic diagram of the DSConv and Swing Transformer structures in the CoSTMNet model.

[0026] Figure 3 It is a full-process physical layer key generation scheme in FDD mode. Detailed Implementation

[0027] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0028] Step 1: Obtain the channel reciprocity characteristics of the communicating parties through the CoSTMNet channel mapping model. The overall structure of CoSTMNet is as follows: Figure 1 As shown. Using CoSTMNet, the Alice terminal can estimate the center frequency f it detected. A uplink channel H A The mapping is to the center frequency f at the Bob end during the coherence time. B downlink channel H B In other words, in an FDD Massive MIMO system, Alice and Bob do not need additional information exchange. Instead, they can obtain the channel reciprocity characteristics through CoSTMNet and use this as the basis for subsequent key generation between Alice and Bob. CoSTMNet adopts an encoder-decoder architecture. The encoder part converts the input H... A (f AThe encoder maps the signal to a latent representation space, thereby extracting frequency-independent features. Simultaneously, noise removal and redundant information compression are also performed in the encoder. The decoder then performs frequency alignment and information reconstruction, reconstructing the latent representation output by the encoder into the target channel H. B (f B And minimize the error between the reconstruction result and the real channel.

[0029] A detailed analysis of the CoSTMNet model structure is provided. The CoSTMNet encoder section consists of an amplitude-phase fusion module and a serial N... c Each convolutional block and N s A number of Swing Transformer blocks

[34] The module is composed of two 1×1 convolutions. Specifically, the amplitude-phase fusion module consists of two 1×1 convolutions. Each convolutional block is composed of a depthwise separable convolution (DSConv) and CBAM channel and spatial attention. The depthwise separable convolution (DSConv) consists of depthwise convolution and pointwise convolution. The depthwise convolution performs convolution operations on each independent channel, and the pointwise convolution performs feature fusion between channels using a 1×1 kernel. Figure 2 (a) illustrates the computation process of depthwise separable convolution (DSConv). DSConv has significant advantages in computational efficiency; when computational resources are limited in the physical layer domain, using DSConv can significantly reduce computational overhead. CBAM, on the other hand, is a concise and efficient attention mechanism. By introducing dual attention—channel and spatial—CBAM can guide the model to adaptively focus on important features across different dimensions.

[35] This improves the feature extraction and representation capabilities of the CoSTMNet encoder unit. Overall, the main function of the convolutional blocks in the encoder is to focus on and extract local frequency-independent features from the channel information. Using a CNN structure as the pre-encoder efficiently extracts local patterns and structural features from the channel information, providing concise and effective feature input for the subsequent Swin Transformer.

[0030] After capturing local features through convolutional structures, the encoder further captures long-range global feature dependencies in channel information through the Swin Transformer structure. Since CoSTMNet aims to map channel information across different frequency bands, cross-band and cross-antenna array global feature extraction is also extremely important. The Swin Transformer, through a moving window-based self-attention mechanism, can effectively model global features in channel information. Similar to ViT, the Swin Transformer first divides the input data into fixed-size patches, using these patches as the basic data units in its processing. In the code implementation, each patch is set to a size of 2×2. After patch segmentation, it is flattened into a one-dimensional vector through linear embedding, and then enters the core self-attention module of the Swin Transformer. Figure 2 (b) It can be seen that the self-attention mechanism of Swin Transformer includes Window-based Self-Attention (WSA) and Shifted Window Self-Attention (SWSA). In WSA, the input data is divided into multiple non-overlapping small windows, and self-attention is calculated within each small window, thus effectively reducing the high computational cost of global self-attention. Since WSA only calculates attention within a window, Shifted Window Self-Attention (SWSA) is further designed in Swin Transformer to enhance the flow of global information across windows. SWSA creates intersections between windows through window movement operations, avoiding the isolation of information within local windows, thus enabling self-attention to learn global feature information across windows. Overall, through WSA and SWSA, Swin Transformer balances local efficiency with global modeling capabilities.

[0031] CoSTMNet's encoder employs a hybrid design combining convolutional and Swing Transformers, enabling effective processing of channel information from multiple dimensions. This hybrid encoder can both precisely capture the local features of channel information and possess the ability to model global features across frequency bands. Therefore, when processing channel information with complex spatiotemporal structures, the CoSTMNet encoder can fully leverage the advantages of its hybrid structure, thereby providing the decoder with multi-level effective features for frequency alignment and information reconstruction.

[0032] The CoSTMNet decoder employs a hybrid architecture design symmetrical to the encoder. Specifically, as shown in Figure 4, the decoder consists of stacked symmetrical Swing Transformers and convolutional structures. Furthermore, skip connections are introduced between corresponding network layers of the encoder and decoder, allowing the decoder to utilize multi-level hierarchical feature maps from the encoder to improve the accuracy of channel information reconstruction. In the CoSTMNet decoder, the feature information output from the encoder is first aligned to the target frequency and restored to the global channel topology via a Swing Transformer, and then further refined by subsequent convolutional structures to refine local channel details. The Swing Transformer block and convolutional block in the decoder use PatchExpanding and UpConvolution, respectively, to restore the channel structure. It is also important to note that the channel information is in complex form, containing both amplitude and phase information; therefore, an amplitude-phase separation module corresponding to the encoder's front end is designed at the decoder's end. Finally, after passing through the amplitude-phase separation module containing two 1×1 convolutions, the CoSTMNet decoder can complete the cross-frequency band channel information H... B (f B The reverse reconstruction process of ).

[0033] Before training the CoSTMNet model, the CSI data for Alice and Bob needs to be preprocessed to improve the stability of model training and accelerate convergence. First, the amplitude and phase in the CSI data are separated, and then amplitude and phase are normalized. After data preprocessing, a CSI pair dataset is constructed. In each CSI pair, the data includes Alice's center frequency f. A H A And the center frequency of Bob's end during the coherence time is f B H B .

[0034] The CoSTMNet model is trained using supervised learning. A (f A H is used as the model input. B (f B The true value label is then represented. During model training, the mean squared error (MSE) loss function is used, and considering the mapping error between amplitude and phase, the loss function can be expressed as:

[0035]

[0036] Among them, A B With A B These represent the H values ​​output by the CoSTMNet mapping, respectively. B Compared with the true value HB The magnitude matrix, P B With P B Then they represent H respectively B With H B The phase matrix, where V represents the batch size during training, and N... H Then it means H B The number of elements after flattening. The ADAM optimizer is used as the optimizer during model training. CoSTMNet model training can be performed online, and after training, the model is deployed offline to the Alice base station for inference.

[0037] Step 2: Quantize the reciprocal channel data of both communicating parties using the PCGQ (Peripherally Compressed and Gray Coded) algorithm to obtain the initial key bit string for both parties. Using the CoSTMNet channel mapping model from the previous step, Alice and Bob can obtain reciprocal channel information (CSI). The dimension of CSI is the number of receiver antennas × the number of transmitter antennas × the number of subcarriers. In an FDD Massive MIMO channel model with a large number of antennas, the number of elements in CSI will be much larger than the key length. Therefore, this invention designs the PCGQ channel compression quantization algorithm to convert CSI into the initial key bit string. Setting the final generated physical layer key length to 1024 bits, the initial key bit string length is set to L = 2048. The specific process of the PCGQ algorithm will be described in detail below. First, CSI amplitude and phase fusion is performed. Each element in Alice and Bob's CSI contains both amplitude and phase information. Quantization algorithms cannot directly process complex numerical data; therefore, element-wise amplitude and phase fusion is first performed on Alice and Bob's CSI. In the previous section, amplitude and phase were normalized; here, we only need to consider their fusion. Specifically, the amplitude-phase fused CSI can be expressed as H. M = α·A + β·P. Where A and P represent amplitude and phase information, respectively. α and β represent the fusion weights of amplitude and phase, satisfying α + β = 1. Here, for the sake of balance, α = β = 0.5. Following this is the calculation of global statistics and CSI partitioning. In this step, firstly, based on the H after amplitude and phase fusion... M Calculate its global mean and variance statistics. During the calculation, H... M Flattened into a vector, denoted as h M and with N M h M The number of elements in the vector, h M [n] represents h M The nth element in h, then M Global mean μ and global variance σ 2 It can be represented as:

[0038]

[0039] Due to h M The number of elements in h is much greater than the set final physical layer key length of 1024, further affecting h. M Perform partitioning operations. Specifically, partition h M Divide the key into 1024 partitions, each with the same length as the final key. Except for the last partition, the length of each partition can be represented as... Ultimately, the resulting set contains 1024 partitions:

[0040] S={h M [c·(i-1)+1:min(c·i,N M )]∣i=1,2,…,1024}

[0041] The next step is partitioned compression quantization, in which each partition s in the set S is compressed. i Perform double-bit compressed quantization. First, based on the global mean μ and variance σ... 2 Construct a Gaussian distribution N(μ,σ) 2 The inverse cumulative distribution function ICDF of ) is denoted as F -1 (·). Four quantization thresholds are set as {τ1=0.25, τ2=0.5, τ3=0.75, τ4=1} to divide the space into four equally probable quantization intervals. Then, each interval is subjected to element-wise double-bit quantization and assigned a corresponding quantization code (00, 01, 10, 11), which can be represented as follows:

[0042]

[0043] Among them, s i [j] represents the i-th partition s of set S. i The j-th element in the expression. ε represents the quantization factor, used to set the isolation band between different quantization intervals to reduce the increase in key inconsistency rate (KDR) caused by edge points. Similarly, edge points outside the four quantization intervals will generate an unreliable tag with a value of "-1" at their position after quantization.

[0044] After element-wise double-bit quantization within a partition, partition-level quantization compression and decision-making are performed. For each partition, unreliable bits with a value of "-1" are ignored, and the frequency of the four quantization codes is calculated. The quantization code with the highest frequency is used as the overall compression quantization result for that partition. For example, if a partition s i If the quantization code for "01" appears most frequently, then the result of the compression quantization of this partition is Q(s). i ) = 01.

[0045] Finally, there is the grey code mapping and initial key generation. In this step, grey code mapping is performed for partitioned compressed quantization to reduce bit differences between adjacent quantization codes. Because channel information reciprocity is not perfect, Alice and Bob have a small number of inconsistent quantization results for corresponding partitions. Grey code encoding can reduce the difference between adjacent quantization codes to only one bit, thus effectively reducing the KDR of the initial key. Specifically, grey code mapping performs the following transformations on the quantization code of each partition:

[0046] (00→00, 01→01, 10→11, 11→10)

[0047] For two-bit quantization encoding, gray code mapping transforms (00,01,10,11) into (00,01,11,10). When Alice and Bob's partitions have adjacent codes due to reciprocity discrepancies, gray code mapping only introduces one inconsistent bit. Finally, after gray code mapping for the quantization encoding of each partition, the two-bit gray codes of the 1024 partitions are concatenated sequentially to obtain an initial key bit string of length L = 2048.

[0048] Step 3: Generate physical layer keys for both communicating parties using the designed FDD mode full-process physical layer key generation scheme, such as... Figure 3 As shown, the FDD key generation scheme includes steps such as channel sounding estimation, preprocessing and channel mapping, compression quantization, information negotiation and privacy amplification;

[0049] Step 3.1: Channel sounding estimation. The channel sounding and estimation at Alice and Bob's ends uses the least squares method. After channel sounding estimation, Alice and Bob can respectively obtain the uplink channel state information H. A With downlink channel state information H B In FDD Massive MIMO systems, H A The center frequency is f A H B The center frequency is f B And f A ≠f B Due to the difference in their frequencies, Alice measured H... A H measured by Bob B Reciprocity is no longer present; the construction of reciprocal information between the two parties will be carried out in subsequent processes.

[0050] Step 3.2, Preprocessing and Channel Mapping: In this step, the reciprocity channel information for Alice and Bob in FDD mode is constructed. First, the H values ​​estimated by the probes of Alice and Bob are processed separately. A With H BPreprocessing is performed, including amplitude-phase separation and amplitude-phase normalization. Further, the Alice end performs channel mapping using the CoSTMNet model designed in Section 4.3. The result of the CoSTMNet channel mapping at the Alice end is denoted as H. B Through channel mapping, the Alice terminal can use a center frequency of f. A H A To indirectly estimate the center frequency as f B H B This allows for the construction of reciprocal channel information between Alice and Bob, which is then used as the basis for subsequent key generation processes.

[0051] Step 3.3: In this step, the PCGQ two-bit channel quantization algorithm from Step 2 is used to convert the reciprocity channel information of Alice and Bob into their initial key bit strings. Specifically, Alice converts the H obtained after CoSTMNet channel mapping... B Bob performs PCGQ channel compression quantization, while Bob directly processes the preprocessed H... B Perform PCGQ compression quantization. After compression quantization, both parties obtain their respective initial keys, denoted as q. A With q B Because the reciprocity of Alice and Bob's channel information is not perfect, therefore, in the initial key q... A With q B There is a certain key inconsistency rate (KDR) between the two parties. In subsequent information negotiation processes, the inconsistency in the initial keys of both parties will be eliminated.

[0052] Step 3.4, Information Negotiation: The main function of information negotiation is to correct inconsistencies in the initial key bit strings of Alice and Bob, reducing the key inconsistency rate (KDR) of the initial key strings to 0, thus obtaining identical key bit streams. As seen in the previous quantization process, despite the isolation protection band within the quantization interval, inconsistencies in the eigenvectors of Alice and Bob inevitably remain after quantization. Therefore, a certain error correction algorithm is needed to correct these inconsistencies, while ensuring that only a very small amount of key information is leaked on the public channel. Commonly used error correction algorithms in key information negotiation include the Cascade block-based interactive error correction protocol and algorithms based on error-correcting code schemes, such as BCH codes, Polar codes, and Turbo codes. This study uses the Cascade error correction protocol algorithm for information negotiation of the initial key between Alice and Bob. The main idea of ​​the Cascade error correction protocol is to divide the initial key into blocks and exchange the hash values ​​of the blocks in each round of interaction. If the hash values ​​are inconsistent, binary recursion is used to further divide the blocks to quickly locate and correct errors until all inconsistencies are finally eliminated. In the Cascade algorithm protocol, Alice and Bob exchange error correction information via a public channel. This public information could potentially be intercepted by Eve, a passive eavesdropper. However, because the error correction information mainly consists of one-way irreversible hash values ​​of blocks, and Eve cannot know the position and order of the sub-blocks, Eve cannot obtain valid key information from it. Furthermore, through subsequent privacy amplification processes, the passive leakage of information in this step can be reduced to a negligible level.

[0053] Step 3.5, Privacy Amplification: The main function of privacy amplification is to further strengthen the consensus key sequence obtained through information negotiation using cryptographic security functions, and generate the final physical layer key for Alice and Bob. During the aforementioned information negotiation step, the eavesdropper Eve may obtain some error correction information. Privacy amplification transforms the consensus sequence into a 1024-bit final key using a secure hash function. Due to the avalanche effect and collision resistance of cryptographic hashes, the final generated physical layer key can be considered an unpredictable uniform random sequence, thus eliminating the risk of information leakage during key negotiation. Commonly used secure hash functions in physical layer key generation include SHAKE-128 / 256, SHA-256 / 512, or iterative construction based on HMAC. This study uses the SHAKE-256 secure hash function for privacy amplification. SHAKE-256 can output hash results of arbitrary lengths according to actual needs. Here, the output length of SHAKE-256 is set to 1024, meaning the final generated physical layer key is 1024 bits long. Specifically, privacy amplification can be expressed as follows:

[0054] Key = SHAKE - 256(k||salt)| 0～1023bit

[0055] Where k is the consistent bit string obtained by Alice and Bob after information negotiation. salt is a one-time random number negotiated by Alice and Bob to prevent eavesdropper Eve from conducting offline searches or constructing rainbow table attacks. Finally, the first 1024 bits of the SHAKE-256 hash output are the physical layer key Key generated by Alice and Bob.

Claims

1. A method for generating large-scale MIMO keys based on channel mapping and compressed quantization, characterized in that, The method includes: Step 1: Obtain the channel reciprocity characteristics of the two communicating parties through the CoSTMNet channel mapping model. Through CoSTMNet, Alice can map the uplink channel with the center frequency of its detected and estimated center frequency to the downlink channel with the center frequency of Bob's center frequency within the coherent time. Step 2: Quantize the reciprocal channel data of the two communicating parties using the PCGQ algorithm based on partitioned compression and gray code encoding to obtain the initial key bit string of the two communicating parties. The PCGQ algorithm process includes steps such as CSI amplitude and phase fusion, global statistics calculation and CSI partitioning, partitioned compression quantization and gray code mapping. Step 3: Complete the physical layer key generation for both communicating parties using the designed FDD mode full-process physical layer key generation scheme. The key generation scheme includes steps such as channel sounding and estimation, preprocessing and channel mapping, compression and quantization, information negotiation and privacy amplification.

2. The method according to claim 1, characterized in that, In step 1, Alice can use CoSTMNet to estimate the center frequency f. A uplink channel H A The mapping is to the center frequency f at the Bob end during the coherence time. B downlink channel H B In other words, in an FDD Massive MIMO system, Alice and Bob do not need additional information exchange. Instead, they can obtain the channel reciprocity characteristics through CoSTMNet and use this as the basis for subsequent key generation between Alice and Bob. CoSTMNet adopts an encoder-decoder architecture, where the encoder part converts the input H... A (f A The encoder maps the signal to a latent representation space, thereby extracting frequency-independent features. Simultaneously, noise removal and redundant information compression are performed in the encoder. The decoder then performs frequency alignment and information reconstruction, reconstructing the latent representation output by the encoder into the target channel H. B (f B To minimize the error between the reconstructed results and the actual channel, preprocessing of Alice and Bob's CSI data is necessary before CoSTMNet model training. This improves model training stability and accelerates convergence. First, the amplitude and phase in the CSI data are separated, and then amplitude and phase are normalized. After preprocessing, a CSI pair dataset is constructed. In each CSI pair, Alice's center frequency is f... A H A And the center frequency of Bob's end during the coherence time is f B H B The CoSTMNet model is trained using supervised learning. A (f A H is used as the model input. B (f B The true value label is 0. The mean squared error loss function (MSE) is used during model training, and the mapping error between amplitude and phase is taken into account.

3. The method according to claim 1, characterized in that, In step 2, the reciprocal channel data of the two communicating parties is quantized using the PCGQ algorithm based on partitioned compression and gray code encoding to obtain the initial key bit string of the two communicating parties. Through the CoSTMNet channel mapping model in the previous step, Alice and Bob can obtain the reciprocal channel information (CSI). The dimension of CSI is the number of receiver antennas × the number of transmitter antennas × the number of subcarriers. In the FDD Massive MIMO channel model with a large number of antennas, the number of elements in CSI will be much larger than the key length. Therefore, this invention designs the PCGQ channel compression quantization algorithm to convert CSI into the initial key bit string. The length of the final generated physical layer key is set to 1024 bits, and the length of the initial key bit string is set to L = 2048. The PCGQ algorithm process includes CSI amplitude and phase fusion, global statistics calculation and CSI partitioning, partitioned compression quantization, and gray code mapping.

4. The method according to claim 1, characterized in that, In step 3, the first step is channel sounding estimation. The channel sounding and estimation at Alice and Bob's ends uses the least squares method. After channel sounding estimation, Alice and Bob can respectively obtain the uplink channel state information H. A With downlink channel state information H B In FDD Massive MIMO systems, H A The center frequency is f A H B The center frequency is f B And f A ≠f B Due to the difference in their frequencies, Alice measured H... A H measured by Bob B Since reciprocity is no longer present, the construction of reciprocity information between Alice and Bob is handled in subsequent processes, including preprocessing and channel mapping. In this step, the reciprocal channel information for Alice and Bob in FDD mode is constructed. First, the H values ​​estimated by Alice and Bob's probes are processed separately. A With H B Preprocessing is performed, including amplitude-phase separation and amplitude-phase normalization. Further, the Alice end performs channel mapping using the CoSTMNet model designed in Section 4.

3. The result of the CoSTMNet channel mapping at the Alice end is denoted as H. B Through channel mapping, the Alice terminal can use a center frequency of f A H A To indirectly estimate the center frequency as f B H B This constructs the reciprocity channel information between Alice and Bob, which is then used as the basis for subsequent key generation. For compression quantization, the PCGQ two-bit channel quantization algorithm from step 2 is used to convert the reciprocity channel information between Alice and Bob into their initial key bit strings. Specifically, Alice... (The sentence is incomplete and requires more context to translate accurately.) B Bob performs PCGQ channel compression quantization, while Bob directly processes the preprocessed H... B After performing PCGQ compression quantization, both parties obtain their respective initial keys, denoted as q. A With q B Because the reciprocity of Alice and Bob's channel information is not perfect, therefore, in the initial key q A With q B There is a certain key inconsistency rate (KDR) between them. In the subsequent information negotiation process, the inconsistency in the initial keys of both parties will be eliminated to further facilitate information negotiation. In the information negotiation process, the main function is to correct the inconsistent bits in the initial key bit strings of Alice and Bob, that is, to reduce the key inconsistency rate (KDR) of the initial key strings to 0, so as to obtain completely identical key bit streams. As can be seen from the previous quantization process, although there is an isolation guard band in the quantization interval, the feature vectors of Alice and Bob will still inevitably have inconsistent bits after quantization. At this time, it is necessary to correct the inconsistent bits of both parties through a certain error correction algorithm, and it is necessary to ensure that only a very small amount of key information is leaked on the public channel. In key information negotiation, commonly used error correction algorithms include the Cascade block interactive error correction protocol and algorithms based on error correction code schemes, such as BCH codes, Polar codes, Turbo codes, etc. This study uses the Cascade error correction protocol algorithm to negotiate the initial key information of Alice and Bob. The main idea of ​​the Cascade error correction protocol is to divide the initial key into blocks and exchange the hash values ​​of the blocks in each round of interaction. If the hash values ​​are inconsistent, an error correction algorithm is adopted. Binary recursion is used to further divide the data into blocks to quickly locate and correct errors until all inconsistent bits are eliminated. Finally, privacy amplification is performed. In this process, the main function is to further strengthen the consensus key sequence obtained through information negotiation using cryptographic security functions, and generate the final physical layer key for Alice and Bob. During the aforementioned information negotiation steps, the eavesdropper Eve may have obtained some error correction information. Privacy amplification transforms the consensus sequence into a 1024-bit final key using a secure hash function. Due to the avalanche effect and collision resistance of cryptographic hashes, the final physical layer key can be considered an unpredictable uniform random sequence, thus eliminating the risk of information leakage during key negotiation. Commonly used secure hash functions in physical layer key generation include SHAKE-128 / 256, SHA-256 / 512, or iterative construction based on HMAC. This study uses the SHAKE-256 secure hash function for privacy amplification. SHAKE-256 can output hash results of arbitrary lengths according to actual needs; here, the output length of SHAKE-256 is set to 1024, meaning the final physical layer key is 1024 bits long.

Images