A privacy-preserving quadrature amplitude modulation method for wireless distributed learning

Abstract

The present invention discloses a privacy-preserving orthogonal amplitude modulation method for wireless distributed learning. The method is aimed at processing floating-point data in communication between a central server and a client: the client uses random sparsification processing of shared random numbers to reduce the amount of data transmitted; then, modulated data is obtained after random encoding and orthogonal amplitude modulation using probability parameters in random sparsification; the modulated data is sent to the central server via wireless communication, and then binary demodulated data is obtained after orthogonal amplitude demodulation. The demodulated data is decoded using probability parameters in random sparsification using shared random numbers to obtain restored data of the client's updated data. The present invention can effectively improve communication efficiency while meeting the requirements of privacy protection and accuracy, and effectively solves the two major challenges of privacy protection and communication efficiency faced by distributed learning in wireless communication scenarios, and the accuracy shows good robustness to the signal-to-noise ratio.

Description

Technical Field

[0001] The present invention relates to the fields of differential privacy and wireless communications, and in particular to a privacy-preserving quadrature amplitude modulation method for wireless communications. Background Art

[0002] Currently, massive amounts of data transmission impose high communication overhead on centralized data analysis paradigms. Therefore, federated data analysis has been proposed as a new distributed computing paradigm that allows data to be used and analyzed while retaining the original data on user devices. Federated data analysis faces two key challenges: data privacy and communication efficiency. On the one hand, users' local data may contain sensitive information, necessitating the design of privacy protection mechanisms. On the other hand, user devices typically have limited communication capabilities, necessitating the use of mechanisms such as compression to improve communication efficiency.

[0003] Differential privacy (DP) has become the gold standard for privacy protection due to its rigorous theoretical basis and easy implementation. The most commonly used method is to add Gaussian or Laplace noise to the data. However, Gaussian and Laplace noise are continuous, and modern computers / mobile devices can usually only generate and store variables of limited precision, which weakens the privacy protection capabilities of Gaussian and Laplace mechanisms. On the other hand, in order to improve the communication efficiency of users, compression methods such as quantization and sparsification are often adopted. In this case, discrete noise is often better. Therefore, in recent years, various privacy protection mechanisms based on discrete noise have been proposed, such as Bernoulli binomial noise, discrete Gaussian mechanism and Poisson binomial mechanism. On this basis, some work studies the basic trade-off between communication efficiency, privacy protection and accuracy in the classical (ε, δ)-DP framework, such as achieving differential privacy while ensuring communication efficiency through quantization, sampling and perturbation entries, but the mean square error of these mechanisms is often different from d 2 In addition, the design of existing discrete differential privacy protection mechanisms often ignores the integration with actual wireless communication scenarios. Summary of the Invention

[0004] To address the problems in the background technology, the present invention provides a privacy-preserving quadrature amplitude modulation method for wireless communications. This method can effectively improve communication efficiency while meeting both privacy protection and accuracy requirements. In wireless communication scenarios, it effectively addresses the two major challenges of privacy protection and communication efficiency faced by distributed learning, and its accuracy demonstrates good robustness to signal-to-noise ratio.

[0005] The technical solution adopted in the present invention is:

[0006] The method is to perform the following processing on floating-point data communicated between a central server and multiple clients:

[0007] S1. For the floating-point type update data (wireless communication data) that the client intends to send to the central server, the client uses random sparse processing of shared random numbers to reduce the data transmission volume;

[0008] S2, obtaining modulated data by performing random encoding and orthogonal amplitude modulation on the randomly thinned updated data using probability parameters in the random thinning;

[0009] S3. The modulated data is sent from the client to the central server via wireless communication. The central server performs orthogonal amplitude demodulation on the received modulated data to obtain binary demodulated data. The demodulated data is decoded using the probability parameters in random sparsification using a shared random number to obtain the recovered data of the client's updated data.

[0010] The specific implementation of step S1 is as follows:

[0011] The client's updated data is represented as a d-dimensional vector x = [x1, x2, ..., x d ],1≤i≤d,x i Represents the i-th element in the vector, the element type is a floating point number, x i satisfy Where c represents x i The upper bound parameter of the absolute value, Letters represent arbitrary symbols;

[0012] The client updates each element of the data with The probability is randomly sparsely distributed by randomly discarding random numbers to reduce the amount of data transmission. A and B are the first probability parameters and the second probability parameters, and B>A is required.

[0013] Specifically, step S2 first randomly codes the randomly thinned update data to become binary coded data containing only two symbols, 0 and 1, and then orthogonally modulates the coded data to become modulated data.

[0014] The random encoding utilizes a first probability parameter in random sparsification.

[0015] The random code in step S2 is mapped and processed in the following manner to become binary type encoded data:

[0016]

[0017] Where code(·) represents the random encoding operation, and p represents the element x in the updated data. i The mapping probability of A represents the first probability parameter in random sparsification, and the first probability parameter A> the upper bound parameter c.

[0018] The decoding of the demodulated data utilizes the second probability parameter in the random sparsification.

[0019] In step S3, after the demodulated data is obtained, the following processing is performed:

[0020] S31. Map the binary demodulated data as follows:

[0021]

[0022] Among them, map ( / ) represents the mapping operation, s j Represents the jth symbol in the demodulated data, and the mapping operation is s j When it is equal to 1, it is mapped to 1, s j When it is equal to 0, it is mapped to -1;

[0023] S32, multiplying the mapped demodulated data by the second probability parameter B in the random thinning to obtain decoded data;

[0024] S33. The central server uses the shared random number in the random thinning process of step S1 to fill the positions of the elements randomly discarded during the random thinning in the decoded data with 0 to obtain the client's recovery data.

[0025] The two probability parameters in the random sparsification satisfy the following relationship:

[0026]

[0027] Among them, μ is a parameter to measure the privacy protection ability.

[0028] The present invention improves communication efficiency by randomly sparsifying the updated data on the client side, and then directly encodes the randomly sparsified updated data into binary coded data through random coding. The coded data is then orthogonally amplitude modulated and sent from the client to the central server via wireless communication. The central server demodulates the received modulated data through orthogonal amplitude modulation to obtain binary demodulated data, and decodes the demodulated data to obtain the recovered data of the client's updated data, which is an unbiased estimator of the client's updated data. Based on the joint analysis and processing of the privacy protection gain provided by random sparsification and random coding, as well as the analysis of the mean square error, the reasonable selection of parameters can ensure the privacy protection and accuracy equivalent to the Gaussian mechanism, and the accuracy is independent of the sparsity. By adjusting the sparsity, the communication efficiency can be effectively improved. In addition, the accuracy of this method shows good robustness to noise.

[0029] The innovation of the present invention lies in the random sparsification processing of the data on the client, and the combination of random sparsification processing parameters and random discarding positions in the subsequent random encoding and orthogonal amplitude modulation, as well as the reception and decoding by the central server, so that the final data forms an unbiased estimated distribution, realizing efficient and accurate privacy protection in wireless communication and improving processing efficiency.

[0030] The present invention has the following beneficial effects:

[0031] The orthogonal amplitude modulation method based on random sparsification and random coding proposed in the present invention can provide privacy protection equivalent to that of a Gaussian mechanism and communication efficiency exceeding that of a Gaussian mechanism. The mean square error of this method is independent of sparsity, so communication efficiency can be improved by adjusting sparsity while ensuring privacy and accuracy.

[0032] At the same time, the central server recovery data of this orthogonal amplitude modulation method is an unbiased estimator of the client update data. The accuracy of this method shows good robustness to noise under Gaussian channel. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The specific structural diagram of the present invention is shown by taking 4QAM, a quadrature amplitude modulation, as an example.

[0034] Figure 2 、 3 This is a comparison chart of the mean square error (MSE) of the privacy-preserving orthogonal amplitude modulation method for wireless distributed learning of the present invention and the orthogonal amplitude modulation method combined with the traditional Gaussian mechanism and 32-bit quantization and sparsification, corresponding to different sparsities and different signal-to-noise ratios under the same privacy protection requirements. In the figure, SNR represents the signal-to-noise ratio.

[0035] Figure 4 、 5 This is a comparison chart of the mean square error (MSE) of the privacy-preserving orthogonal amplitude modulation method for wireless distributed learning of the present invention and the orthogonal amplitude modulation method combined with the traditional Gaussian mechanism and 32-bit quantization and sparsification, corresponding to different privacy protections and different signal-to-noise ratios at the same sparsity. DETAILED DESCRIPTION

[0036] The present invention will be further described below with reference to the accompanying drawings and specific implementations.

[0037] The embodiments of the present invention and their implementation process are as follows:

[0038] This embodiment is aimed at the scenario of distributed learning using wireless communication as a means of information transmission, and includes a central server and multiple clients, where the client serves as a communication sending end and the central server serves as a communication receiving end. In the distributed learning task, each client sends local update data to the central server using wireless communication, and the central server receives the update data from each client via wireless communication, where the type of each element in the update data is a floating point number.

[0039] S1. For the floating-point type update data (wireless communication data) that the client intends to send to the central server, the client uses random sparse processing of shared random numbers to reduce the data transmission volume;

[0040] The client's updated data is represented as a d-dimensional vector x = [x1, x2, ..., x d ],1≤i≤d,x i Represents the i-th element in the vector, indicating the position of the element in the vector. The element type is a floating point number. i satisfy Where c represents x i The upper bound parameter of the absolute value, Letters represent arbitrary symbols;

[0041] The client updates each element of the data with The probability is randomly discarded by random numbers to achieve random sparseness to reduce the amount of data transmission. A and B are the first probability parameter and the second probability parameter that allow the size to be flexibly set according to actual needs, and B>A is required. The value of represents the sparsity of random sparsification, The smaller it is, the greater the probability of discarding each element, and the sparser the updated data after random sparsification.

[0042] During random sparsification, the client shares the random number with the central server, meaning that both the client and the central server share the random number using the same random number seed. Because the random number is shared, the central server knows the position of the elements randomly discarded by the client during random sparsification in the updated data during subsequent demodulation.

[0043] S2, obtaining modulated data by performing random encoding and orthogonal amplitude modulation on the randomly thinned updated data using probability parameters in the random thinning;

[0044] Specifically, step S2 first randomly codes the randomly thinned update data to become binary coded data containing only two symbols, 0 and 1, and then orthogonally modulates the coded data to become modulated data.

[0045] Random encoding uses the first probability parameter in random sparsification, and is specifically mapped and processed in the following way to become binary type encoded data:

[0046]

[0047] Where code(·) represents the random encoding operation, and p represents the element x in the updated data. i The mapping probability of S1 is 1, A represents the first probability parameter in random sparsification, and the first probability parameter A> the upper bound parameter c. A is equal to A during random sparsification of S1, and is a parameter that allows the size to be set according to actual needs to control the level of randomness.

[0048] The above operation is to randomly sparsely update each element x in the data i By probability Map it to symbol 1, with probability Map it to symbol 0.

[0049] S3. The modulated data is sent from the client to the central server via wireless communication. The central server performs orthogonal amplitude demodulation on the received modulated data to obtain binary demodulated data. The demodulated data is decoded using the probability parameters in random sparsification using shared random numbers to obtain the recovered data of the client's updated data. The recovered data is an unbiased estimator of the client's updated data.

[0050] The decoding of the demodulated data utilizes the second probability parameter in the random sparsification. Specifically, after obtaining the demodulated data, the following processing is performed:

[0051] S31. Map the binary demodulated data as follows:

[0052]

[0053] Among them, map(·) represents the mapping operation, s j Represents the jth symbol in the demodulated data, and the mapping operation is s j When it is equal to 1, it is mapped to 1, s j When it is equal to 0, it is mapped to -1;

[0054] S32, multiplying the mapped demodulated data by the second probability parameter B in the random thinning to obtain decoded data;

[0055] S33. The central server uses the shared random number in the random thinning process of step S1 to fill the positions of the elements randomly discarded during the random thinning in the decoded data with 0 to obtain the client's recovery data.

[0056] Represent the recovered data as a d-dimensional vector z, and we get E[z i]=x i , that is, the recovered data is an unbiased estimator of the client's updated data, where z i Represents the i-th element in the received vector.

[0057] The present invention can reduce the amount of data required to be transmitted for distributed learning in wireless scenarios while achieving privacy protection through the coordinated relationship between random sparsification, random coding and modulation between steps S1 and S2.

[0058] The existing technology usually uses the Gaussian mechanism as the standard for privacy protection. The present invention can provide privacy protection equivalent to the Gaussian mechanism, but its communication efficiency exceeds that of the Gaussian mechanism.

[0059] In terms of privacy protection, assuming that the Gaussian mechanism satisfies μ-GDP, GDP is a standard for measuring privacy protection, and μ is a parameter for measuring privacy protection capabilities. When the selection of A and B satisfies It can provide privacy protection equivalent to that of the Gaussian mechanism.

[0060] In terms of communication efficiency, the Gaussian mechanism that satisfies μ-GDP is implemented by adding Gaussian noise is used to provide privacy protection, where n represents Gaussian noise. represents Gaussian distribution, σ represents the standard deviation of Gaussian distribution, The updated data after adding Gaussian noise is of floating-point type. Therefore, in a wireless scenario, each element of the updated data needs to be further quantized into 32 binary symbols by 32 bits before being modulated. Compared with the S2 random encoding process of the present invention, each element of the updated data is directly encoded into 1 binary symbol. The Gaussian mechanism has a large amount of transmitted data and low communication efficiency.

[0061] In an ideal channel, the mean square error is close to the Gaussian mechanism and remains unchanged under different sparsities; in a Gaussian channel, the mean square error shows good robustness to noise. The specific analysis is as follows:

[0062] Under ideal channel, the present invention Gaussian mechanism mse'=σ 2 , where mse represents the i-th element z of the central server recovery data of the present invention i And the client updates the i-th element x of the data i The mean square error between them, mse' represents the i-th element z of the data recovered by the central server under the Gaussian mechanism i And the client updates the i-th element x of the data i The mean square error between them is small and the accuracy is high. When the present invention and the Gaussian mechanism provide the same privacy protection, the following formula AB=c is satisfied: 2 +σ 2,So Usually small, so the mean square error of the two is close. In addition, we can see that mse(z i ) is only related to the product AB of A and B, and the sparsity of random sparsification Therefore, increasing the sparsity to reduce the amount of transmitted data will not affect the mean square error and accuracy.

[0063] Under Gaussian white noise channel conditions, the bit error rate of MQAM (M-ary quadrature amplitude modulation) is expressed as P b , for the updated data i-th element x i , the corresponding element z in the restored data i =1 is z i = -1 has a probability of Mean square error generally Therefore, AB dominates the mean square error. As the noise changes, the bit error rate P b The impact on the mean squared error is small, so the accuracy shows good robustness to noise.

[0064] Through computer simulation verification, it can be seen that: Figure 2 、 3 As shown in the figure, compared with the orthogonal amplitude modulation method combined with 32-bit quantization and sparsification of the Gaussian mechanism, the privacy-preserving orthogonal amplitude modulation method for wireless distributed learning proposed in this invention has the same privacy protection requirements, and the MSE does not change with the sparsity. Therefore, the sparsity can be adjusted to achieve higher communication efficiency. Figure 2 、 3 Alternatively, as shown in Figures 4 and 5, the proposed method demonstrates superior robustness to noise under Gaussian channel conditions, compared to the traditional Gaussian mechanism combined with 32-bit quantization and sparsification, using the MSE metric. Therefore, the privacy-preserving quadrature amplitude modulation method for wireless distributed learning proposed in this paper provides a highly effective modulation scheme that balances privacy protection and communication efficiency in wireless communication scenarios.

[0065] Specific examples:

[0066] The client updates the data x = [x1, x2, x3, x4] = [0.2, 0.4, 0.25, 0.1], A = 0.5, B = 1, and uses 4QAM. First, after step S1, the client updates each element of the data with Probabilistic random dropping is used to achieve random sparsification. Assume that the updated data after random sparsification is x'=[x1,x3]=[0.2,0.25], where x' represents the updated data after random sparsification, i.e., the two elements x2 and x4 are discarded after random sparsification. Next, in step S2, each element of the updated data after random sparsification is randomly encoded. Specifically, the element x1 in the updated data after random sparsification is randomly encoded with probability Coded as 1, with probability Encoded as 0, the same operation is performed on element x3, with probability Coded as 1, with probability The encoding is 0. Assuming the randomly encoded data is [1, 0], this encoded data undergoes quadrature amplitude modulation to become modulated data. The modulated data is sent from the client to the central server via wireless communication. The central server performs quadrature amplitude demodulation on the received modulated data to obtain the binary demodulated data [1, 0]. During the decoding process, the demodulated data is first mapped to [1, -1] and then multiplied by B = 1 to obtain the recovered data [1, -1]. When the number of clients is large, the central server can obtain an unbiased estimate of the mean of the updated data for each client in a statistically significant manner.

[0067] The above specific embodiments are used to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

[0068] The above description is only a preferred embodiment of the present invention. Therefore, any equivalent changes or modifications made according to the structure, characteristics and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims

1. A privacy-preserving quadrature amplitude modulation method for wireless distributed learning, characterized by: The method is to perform the following processing on floating-point data communicated between a central server and multiple clients: S1. For the floating-point type update data that the client intends to send to the central server, the client uses random sparse processing of shared random numbers to reduce the amount of data transmission; S2, obtaining modulated data by performing random encoding and orthogonal amplitude modulation on the randomly thinned updated data using probability parameters in the random thinning; S3. The modulated data is sent from the client to the central server via wireless communication. The central server performs orthogonal amplitude demodulation on the received modulated data to obtain binary demodulated data. The demodulated data is decoded using the probability parameters in random sparsification using a shared random number to obtain the recovered data of the client's updated data.

2. The privacy-preserving quadrature amplitude modulation method for wireless distributed learning according to claim 1, characterized in that: The specific implementation of step S1 is as follows: The client's updated data is represented as a d-dimensional vector x = [x1, x2, ..., x d ],1≤i≤d,x i Represents the i-th element in the vector, the element type is a floating point number, x i satisfy Where c represents x i Upper bound parameter of absolute value; The client updates each element of the data with The probability is randomly sparsely distributed by randomly discarding random numbers to reduce the amount of data transmission. A and B are the first probability parameters and the second probability parameters, and B>A is required.

3. The privacy-preserving quadrature amplitude modulation method for wireless distributed learning according to claim 1, characterized in that: Specifically, step S2 first randomly codes the randomly thinned update data to become binary coded data containing only two symbols, 0 and 1, and then orthogonally modulates the coded data to become modulated data.

4. A privacy-preserving quadrature amplitude modulation method for wireless distributed learning according to claim 1 or 3, characterized in that: The random encoding utilizes a first probability parameter in random sparsification.

5. A privacy-preserving quadrature amplitude modulation method for wireless distributed learning according to claim 1 or 3, characterized in that: The random code in step S2 is mapped and processed in the following manner to become binary type encoded data: Where code(·) represents the random encoding operation, and p represents the element x in the updated data. i The mapping probability of A represents the first probability parameter in random sparsification, and the first probability parameter A> the upper bound parameter c.

6. The privacy-preserving quadrature amplitude modulation method for wireless distributed learning according to claim 1, characterized in that: The decoding of the demodulated data utilizes the second probability parameter in the random sparsification.

7. The privacy-preserving quadrature amplitude modulation method for wireless distributed learning according to claim 1, characterized in that: In step S3, after the demodulated data is obtained, the following processing is performed: S31. Map the binary demodulated data as follows: Among them, map(·) represents the mapping operation, s j Represents the jth symbol in the demodulated data, and the mapping operation is s j When it is equal to 1, it is mapped to 1, s j When it is equal to 0, it is mapped to -1; S32, multiplying the mapped demodulated data by the second probability parameter B in the random thinning to obtain decoded data; S33. The central server uses the shared random number in the random thinning process of step S1 to fill the positions of the elements randomly discarded during the random thinning in the decoded data with 0 to obtain the client's recovery data.

8. The privacy-preserving quadrature amplitude modulation method for wireless distributed learning according to claim 2, characterized in that: The two probability parameters in the random sparsification satisfy the following relationship: Among them, μ is a parameter to measure the privacy protection ability.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium or electronic device having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images