Quantum-resistant robust parameter aggregation federated learning method and device for genomics

Abstract

An embodiment of the present invention relates to a quantum-resistant robust parameter aggregation federated learning method and device based on genomics. The method includes the following steps: the client quantizes and encrypts local model parameters within the TEE; the client exports the ciphertext of the encrypted model parameters from the TEE and uploads it to a central server; the central server robustly aggregates the encrypted model parameters in the encryption domain to obtain optimized model parameters and sends them to the client; the client decrypts and dequantizes the optimized model parameters and then updates the local model. The technical solution of the embodiment of the present invention, by combining the TEE for local computing and the TFHE method for parameter transmission and aggregation, can keep the original data and model updates encrypted throughout the entire process, eliminating the risk of data leakage through inference attacks.

Description

Technical Field

[0001] Embodiments of the present invention relate to the field of secure computing technology, and in particular to a quantum-resistant robust parameter aggregation federated learning method and device for genomics. Background Art

[0002] Federated learning (FL) is a distributed machine learning paradigm that allows multiple institutions to collaboratively train a shared model without exchanging raw data. This approach is particularly important in fields where data privacy must be strictly protected, such as genomics, where single-cell RNA sequencing (scRNA-seq) data contains sensitive genetic information. In traditional FL, each participating institution trains a local model on its private dataset and shares only the model parameters (such as weights or gradients) with a central server. The central server aggregates these parameters to update the global model, which is then redistributed to the institutions for further training. This iterative process maintains data locality and reduces the risk of direct data exposure.

[0003] Despite the aforementioned advantages of federated learning, traditional federated learning still suffers from the following problems when applied to single-cell RNA sequencing (scRNA-seq) data analysis:

[0004] Transmitting model parameters in plain text poses a significant privacy threat. Attackers may use these parameters to reconstruct part of the original data or infer sensitive information of individuals through inference attacks (such as membership inference attacks). This vulnerability undermines the privacy guarantees required for genomic data processing.

[0005] Existing encryption methods typically allow computation on encrypted data, but are limited to approximate arithmetic. These schemes cannot support the precise comparison operations required by advanced aggregation algorithms, limiting their ability to ensure robust and accurate model updates in a secure manner.

[0006] Traditional federated learning systems typically assume benign participants. However, in real-world scenarios, malicious nodes (e.g., Byzantine adversaries) may submit corrupted updates to disrupt the training process or degrade model performance. Standard aggregation methods lack mechanisms to detect or mitigate such attacks, and system reliability degrades when a large number of nodes are compromised. Summary of the Invention

[0007] Based on the above-mentioned situation of the prior art, the purpose of the embodiments of the present invention is to provide a quantum-resistant robust parameter aggregation federated learning method and device for genomics. Through the TFHE algorithm and the enhanced robust aggregation algorithm, the problems of privacy, computational accuracy and anti-adversarial interference existing in traditional federated learning are solved.

[0008] To achieve the above objectives, according to a first aspect of the present invention, a quantum-resistant robust parameter aggregation federated learning method for genomics is provided, comprising the steps of:

[0009] The client quantizes and encrypts local model parameters within the TEE.

[0010] The client exports the ciphertext of the encrypted model parameters from the TEE and uploads it to the central server;

[0011] The central server performs robust aggregation processing on the encrypted model parameters in the encryption domain to obtain optimized model parameters and sends them to the client;

[0012] The client decrypts and dequantizes the optimized model parameters and then updates the local model;

[0013] Among them, the local model parameters are obtained by the client by training the local model in the TEE based on the local private data set; the robust aggregation processing optimizes and updates the global model parameters based on the similarity between the encrypted model parameters of each client.

[0014] Furthermore, the method further comprises the steps of:

[0015] The client obtains the change value of each updated optimized model parameter relative to the model parameter in the previous round, and uploads the change value and the ciphertext of the encrypted model parameter in this round to the central server;

[0016] The central server determines the importance of each parameter based on the change value, and performs truncation processing on the encryption model parameters of each client according to the importance.

[0017] Furthermore, the local private dataset includes raw single-cell RNA sequencing data; the method further comprises the steps of:

[0018] Preprocess the raw single-cell RNA sequencing data of the local private dataset.

[0019] Furthermore, the preprocessing includes removing zero-expression genes in the data, converting the original single-cell RNA sequencing data into CPM, selecting highly expressed genes based on the converted data, and performing PCA dimensionality reduction on the selected genes.

[0020] Furthermore, the robust aggregation process includes:

[0021] Calculate the Euclidean distance between each client's encryption model parameter vector;

[0022] Based on the calculated Euclidean distance, the parameters of the encryption model participating in the global model are adjusted;

[0023] The global model parameters are optimized and updated based on the adjustment results.

[0024] Furthermore, based on the calculated Euclidean distance, the encryption model parameters participating in the global model are adjusted, including:

[0025] For each client, calculate the sum of the Euclidean distances between the client and the encryption model parameters of other clients as the distance set of the client;

[0026] Among all the client distance sets, the client model parameters corresponding to the largest distance set are eliminated.

[0027] Furthermore, the global model parameters are optimized and updated based on the adjustment results, including:

[0028] After removing the client model parameters corresponding to the largest distance set, the average value of the encrypted model parameter vectors of the remaining clients is calculated;

[0029] The average value is used as the optimization update parameter of the global model.

[0030] Furthermore, the central server performs robust aggregation processing on the ciphertext by executing homomorphic operations in the compiled TFHE computation circuit.

[0031] Furthermore, the TFHE calculation circuit is pre-generated at any client according to the local model parameter vector of the client.

[0032] According to another aspect of the present invention, a quantum-resistant robust parameter aggregation federated learning device for genomics is provided, comprising:

[0033] The client parameter encryption module is used to quantize and encrypt local model parameters within the TEE.

[0034] The client parameter upload module is used to export the ciphertext of the encryption model parameters from the TEE and upload it to the central server;

[0035] A central server aggregation module is used to perform robust aggregation processing on the encryption model parameters in the encryption domain to obtain optimized model parameters and send them to the client;

[0036] A client parameter updating module is used to update the local model after decrypting and dequantizing the optimized model parameters;

[0037] Among them, the local model parameters are obtained by the client by training the local model in the TEE based on the local private data set; the robust aggregation processing optimizes and updates the global model parameters based on the similarity between the encrypted model parameters of each client.

[0038] In summary, an embodiment of the present invention provides a quantum-resistant robust parameter aggregation federated learning method and device based on genomics. The method includes the following steps: the client quantizes and TFHE encrypts local model parameters within the TEE; the client exports the ciphertext of the encrypted model parameters from the TEE and uploads it to the central server; the central server robustly aggregates the encrypted model parameters in the encryption domain to obtain optimized model parameters and sends them to the client; the client decrypts and dequantizes the optimized model parameters and then updates the local model. The technical solution of this embodiment combines a TEE for local computation with TFHE for parameter transmission and aggregation. This ensures that both raw data and model updates remain encrypted throughout the entire process, eliminating the risk of data leakage through inference attacks and providing a zero-exposure guarantee for sensitive single-cell RNA sequencing (scRNA-seq) data. The enhanced Krum algorithm effectively mitigates the impact of malicious nodes by selecting updates consistent with the majority, maintaining model integrity even in the presence of adversarial perturbations. This robustness is critical for real-world deployments where trustworthiness of all participants cannot be assumed. Dynamic truncation reduces the computational overhead of TFHE operations while maintaining sufficient accuracy for model training and aggregation. TEE-based local training further optimizes performance by avoiding encryption during the computationally intensive training phase, balancing security and efficiency. The technical solution provided by this embodiment supports accurate cell type classification and other single-cell RNA sequencing (scRNA-seq) tasks across distributed datasets, enabling collaboration among clients without sacrificing data privacy or model quality, and is compatible with a variety of machine learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flowchart of a genomics-oriented quantum-resistant robust parameter aggregation federated learning method provided by an embodiment of the present invention;

[0040] Figure 2 is a flowchart of a genomics-oriented quantum-resistant robust parameter aggregation federated learning method provided by another embodiment of the present invention;

[0041] Figure 3 This is a flowchart of a genomics-oriented quantum-resistant robust parameter aggregation federated learning method provided by another embodiment of the present invention. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present invention.

[0043] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in one or more embodiments of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprising" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0044] For single-cell RNA sequencing (scRNA-seq) data, federated learning facilitates cross-institutional collaboration by leveraging diverse datasets to enhance cell type classification and other analyses while complying with privacy regulations. Typically, model parameters are transmitted in plaintext to facilitate efficient computation and aggregation. Existing implementations often rely on basic aggregation methods (such as averaging) and may incorporate encryption schemes for enhanced security. However, while suitable for distributed training, this basic approach exhibits vulnerabilities in highly sensitive data scenarios and is vulnerable to data poisoning attacks, limiting its application. It also poses potential privacy risks, encryption limitations, and vulnerability to malicious attacks.

[0045] The embodiment of the present invention provides a genomics-oriented quantum robust parameter aggregation federated learning method. Figure 1 The flowchart of the federated learning method according to an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes the following steps:

[0046] S11. The client quantizes and encrypts the local model parameters within the TEE. The local model parameters are obtained by training the local model within the TEE based on the local private dataset. In this embodiment of the present invention, the local private dataset includes raw single-cell RNA sequencing data, and the step of preprocessing the raw single-cell RNA sequencing data of the local private dataset is also included.

[0047] S12. The client exports the ciphertext of the encrypted model parameters from the TEE and uploads it to the central server.

[0048] S13. The central server performs robust aggregation processing on the encryption model parameters in the encryption domain to obtain optimized model parameters and sends them to the client. This robust aggregation process optimizes and updates the global model parameters based on the similarity between the encryption model parameters of each client.

[0049] S14. After the client decrypts and dequantizes the optimized model parameters, it updates the local model.

[0050] An embodiment of the present invention further provides a genomics-oriented quantum robust parameter aggregation federated learning device, the device comprising:

[0051] The client parameter encryption module is used to quantize and encrypt local model parameters within the TEE.

[0052] The client parameter upload module is used to export the ciphertext of the encryption model parameters from the TEE and upload it to the central server;

[0053] A central server aggregation module is used to perform robust aggregation processing on the encryption model parameters in the encryption domain to obtain optimized model parameters and send them to the client;

[0054] A client parameter updating module is used to update the local model after decrypting and dequantizing the optimized model parameters;

[0055] Among them, the local model parameters are obtained by the client by training the local model in the TEE based on the local private data set; the robust aggregation processing optimizes and updates the global model parameters based on the similarity between the encrypted model parameters of each client.

[0056] The above-mentioned embodiments of the present invention involve a method and system for quantum-resistant robust parameter aggregation federated learning based on genomics, in which multiple clients and a central server interact to complete the federated learning process. The technical solutions of the above-mentioned embodiments are described below from the perspectives of the client and the central server respectively.

[0057] The embodiment of the present invention also provides a genomics-oriented quantum robust parameter aggregation federated learning method, which is applied to the client. Figure 2 The flowchart of the federated learning method according to an embodiment of the present invention is shown in FIG. Figure 2 As shown, the method includes the following steps:

[0058] S102: Based on the local private dataset, a local model is trained within the TEE to obtain local model parameters. In this embodiment of the present invention, local model training is performed within a Trusted Execution Environment (TEE), a hardware-isolated enclave that ensures the confidentiality and integrity of the local model training process. A TEE-enabled device (e.g., one supporting Intel SGX, AMD SEV, or equivalent technology) can be deployed on each client participating in joint learning. The TEE encrypts memory and isolates the computing process from the host operating system, preventing unauthorized access even if the device is compromised. Key management for subsequent encryption steps can also be securely handled within the TEE. Within the TEE, raw single-cell RNA sequencing (scRNA-seq) data from the local private dataset is preprocessed. This includes removing zero-expressed genes from the data, converting the raw scRNA-seq data to CPM (counts per million), selecting highly expressed genes based on the converted data, and performing PCA dimensionality reduction on the selected genes. Single-cell data often contain a large number of zero values, which may be caused by technical noise or true biological non-expression. Removing zero-expressed genes usually refers to filtering out genes with zero expression in most cells (for example, retaining genes expressed in at least 10% of cells), thereby reducing noise and data dimensionality. By normalizing the raw data to CPM, differences in sequencing depth can be eliminated, making expression levels comparable between different cells. Highly expressed or highly variable genes can be screened based on the mean or variance of gene expression. PCA dimensionality reduction is performed on the screened genes to extract the main variation directions, which can reduce redundant information and accelerate subsequent analysis. After data preprocessing, plaintext data is used to train a local model, such as a neural network or support vector machine for cell type classification.

[0059] S104. Quantize and perform TFHE on the local model parameters within the TEE. TFHE (Tiny Fully Homomorphic Encryption) is an improved homomorphic encryption method. After model training is complete, the local model parameters are extracted from the model within the TEE. The local model parameters are in plaintext floating-point format. To make subsequent TFHE and homomorphic computation more efficient and feasible while maintaining sufficient accuracy, the extracted floating-point local model parameters are quantized within the TEE. Quantization maps floating-point numbers to a discrete integer range with a fixed bit width (e.g., 8 or 16 bits). In this embodiment of the present invention, quantization includes pre-quantization and post-quantization. Pre-quantization determines the range of model parameters. Relevant parameters of each client participating in federated learning can be pre-collected, including the maximum and minimum values ​​of the parameters. The range of the model parameters is determined based on the maximum and minimum values. Post-quantization maps the floating-point model parameters to quantized integers. The quantized model parameters are then represented in fixed-point or integer form.

[0060] According to some optional embodiments, the method further comprises the steps of:

[0061] S1041. Obtain the change value of each updated optimized model parameter relative to the model parameter of the previous round, and upload the change value and the ciphertext of the encrypted model parameter of this round to the central server, so that the central server determines the importance of each parameter based on the change value, and truncates the encrypted model parameters of each client according to the importance. In this embodiment of the present invention, a dynamic truncation method is used for processing. For each client's local model parameter, its importance score is calculated. Scoring can be performed in a variety of ways, such as calculating the gradient norm of the parameter, the absolute value of the parameter, etc. In one embodiment of the present invention, the average value of the absolute value of the parameter change is used as the importance scoring indicator. Specifically, for each round of local training of the client, a number of local model parameters will be generated for a local model. For each model parameter , where i represents the client number and j represents the model parameter index. Calculate the absolute value of the model parameter change between the current round and the previous round. , and then calculate the average value of the absolute value of all changes in the model parameters over the past rounds , serving as an indicator for the importance score of the model parameters in this round, the average value is encrypted or uploaded to the central server in plain text, so that the central server can determine the importance of each parameter based on the change value. The larger the average value, the more important the corresponding model parameter.

[0062] The quantized model parameters are encrypted within the TEE using the TFHE algorithm. This embodiment of the present invention preferably uses a TFHE library that supports exact integer arithmetic and comparisons, such as the TFHE-based library provided by Zama (e.g., Concrete). The client encrypts each quantized parameter (or parameter block) using a private key to generate the corresponding TFHE ciphertext. TFHE supports exact arithmetic and comparison operations on encrypted data, overcoming the limitations of approximate encryption schemes. After training, the local model parameters are quantized to a discrete format within the TEE and then encrypted into TFHE ciphertext. This embodiment of the present invention uses the TFHE algorithm to encrypt the quantized model parameters within the TEE on the client side, addressing the problem that existing encryption methods (such as semi-homomorphic encryption and fully homomorphic encryption not based on TFHE) cannot perform complex aggregation operations in ciphertext. For example, semi-homomorphic encryption cannot perform addition and multiplication simultaneously, and fully homomorphic encryption not based on TFHE cannot perform complex computational operations (such as CKKS).

[0063] S106: The ciphertext of the encrypted model parameters and the absolute values ​​of the model parameter changes are exported from the TEE and uploaded to the central server, allowing the central server to perform robust aggregation processing on the encrypted model parameters in the encryption domain. The ciphertext of the encrypted model parameters is securely exported from the TEE and uploaded to the central server via a secure channel. At this point, even if the ciphertext is intercepted, the original local model parameter information cannot be directly obtained.

[0064] S108. Receive the optimized model parameters from the central server. After completing the encrypted aggregation, the central server sends the optimized model parameters, i.e., the encrypted global model update ciphertext (or the selected client update ciphertext), back to the relevant client via a secure channel. After receiving the ciphertext within the TEE, the client performs the following steps to update its local model:

[0065] S110. Decrypt and dequantize the optimized model parameters to obtain updated optimized model parameters. Within the TEE, the received ciphertext is decrypted using the client's locally held TFHE private key. The decryption operation is performed in the secure environment of the TEE to ensure that the private key is not leaked. The decryption result is the quantized integer parameter. For the quantized integer parameter obtained by decryption, the dequantization algorithm is executed within the TEE to convert the quantized integer parameter back to an approximate original floating-point representation. Dequantization is the inverse process of quantization.

[0066] S112: Update the local model using the updated optimized model parameters. Use the floating-point parameters obtained after dequantization to update the client local model (for example, perform operations such as replacing or weighted averaging the model parameters).

[0067] The embodiment of the present invention also provides a quantum robust parameter aggregation federated learning method for genomics, which is applied to a central server. Figure 3 The flowchart of the federated learning method according to an embodiment of the present invention is shown in FIG. Figure 3 As shown, the method includes the following steps:

[0068] S202: Receive the encrypted model parameter ciphertext uploaded by the client. The encrypted model parameter is obtained by the client locally quantizing and performing TFHE processing on the local model parameter. The central server receives the encrypted model parameter ciphertext uploaded by all participating federated learning clients and executes the enhanced Krum robust aggregation algorithm in the encrypted domain. The central server does not need to hold the encrypted private key; all computations in the central server are performed homomorphically on the ciphertext.

[0069] S204: Robust aggregation is performed on the encryption model parameters in the encryption domain to obtain optimized model parameters. In this embodiment of the present invention, robust aggregation employs an enhanced Krum aggregation algorithm to optimize and update global model parameters based on the similarity between the encryption model parameters of each client. The enhanced Krum aggregation algorithm comprises a series of computational steps, including calculating the Euclidean distances between model update vectors, sorting these distances, selecting nearest neighbors, summing the distances, and finding the client with the smallest sum. These steps involve basic arithmetic operations (subtraction, multiplication, and addition) and comparison operations. This embodiment of the present invention defines this computational logic as a computation graph or program. Compiler tools (such as Concrete) within the ZamaTFHE ecosystem are used to compile this computation graph into a TFHE computation circuit suitable for efficient execution on TFHE ciphertext. This compilation process is part of the algorithm definition phase and can be completed offline before system deployment or by a coordinator or other trusted role at the start of a federated learning session. The compiled circuit is then provided to the central server. The compiled circuit defines a series of homomorphic gate operations (such as homomorphic addition gates, homomorphic multiplication gates, etc.) and noise reduction (Bootstrapping) operations that the server needs to execute in sequence. The central server loads the ciphertext of the client's encrypted model parameters and the pre-compiled TFHE calculation circuit. The central server uses the loaded ciphertext of the encrypted model parameters as input to start the execution of the TFHE calculation circuit. The entire calculation process is performed entirely in the encrypted domain, and the central server processes the ciphertext by executing the homomorphic operations in the compiled TFHE calculation circuit. The TFHE calculation circuit can be pre-generated on any client: the client's local model vector is input into the TFHE toolkit, and the required logical operations are simultaneously input into the TFHE toolkit. The TFHE calculation circuit is generated through the TFHE toolkit, which can calculate Euclidean distance, compare sizes, and other functions in a secret state. The robust aggregation processing performed by the central server through the TFHE calculation circuit includes the following steps:

[0070] S2041. Calculate the Euclidean distance between each pair of encryption model parameter vectors of each client. The encryption model parameters uploaded by each client constitute an encryption model parameter vector. In this step, calculate the Euclidean distance between each pair of encryption model parameter vectors uploaded by each client.

[0071] S2042. Based on the calculated Euclidean distances, adjust the encryption model parameters of the participating global model. For each client, calculate the sum of the Euclidean distances between the client and the encryption model parameters of all other clients, and use this sum as the client's distance set. From the distance sets of all clients, remove the client model parameter corresponding to the largest distance set.

[0072] S2043: Select the encrypted model parameters with the smallest sum to update the parameters of the global model. After eliminating the client model parameters corresponding to the largest distance set, calculate the average of the encrypted model parameter vectors of the remaining clients and use this average as the optimized update parameter of the global model.

[0073] According to certain optional embodiments, before performing robust aggregation processing, the importance of each model parameter is determined based on the model parameter change value uploaded by the client, and the encrypted model parameters of each client are truncated according to the importance to improve the consistency and efficiency of data processing. Specifically, the absolute values ​​of the model parameter change values ​​uploaded by each client are integrated to obtain a global parameter importance score, which can be integrated using a simple average method or a weighted average method. In this embodiment, the simple average method is used, that is, for each model parameter index j, the average value of the importance score of the model parameter of all clients is calculated:

[0074] ;

[0075] Where N represents the number of clients. represents the importance score of the jth parameter of the i-th client. The importance score can be obtained by the central server by sorting the absolute values ​​of the model parameter changes uploaded by the clients.

[0076] All parameters are sorted according to the global parameter importance score. The sorting can be in ascending or descending order. In this embodiment, ascending order is adopted, that is, the parameters with lower importance scores are placed at the front. According to the sorting results, different truncation bit numbers are assigned to different parameters. Specifically, a set of optional truncation bit number sets can be pre-set. ,in, ; Then, according to the sorting order of the parameters, the elements in the quantization bit number set are assigned to each parameter in turn. For example, if there are M parameters and the truncation bit number set has n elements, then the first The number of quantization bits is assigned to the kth parameter (k starts counting from 0). The central server receives the truncated parameters sent by each client and truncates the data according to the different truncation bits during robust aggregation processing to obtain the truncated model parameters.

[0077] S206: Send the optimized update parameters to the client, so that the client decrypts and dequantizes the optimized update parameters and then updates the local model.

[0078] The above steps use a method based on model vector similarity to identify and eliminate model parameters uploaded by malicious nodes, thereby ensuring the reliability of the federated learning system.

[0079] The client-side processing and central server aggregation processes in each of the above embodiments constitute a training iteration of federated learning. By repeatedly executing these iterative processes, the global model is gradually optimized until a predetermined convergence condition or a maximum number of iterations is reached. Throughout this process, the original data and model parameters exist only in plaintext within the client's TEE, while the central server always processes encrypted data, thus greatly ensuring data security.

[0080] An embodiment of the present invention further provides a genomics-oriented quantum-resistant robust parameter aggregation federated learning device, which is applied to a client and includes:

[0081] The model training module is used to train the local model in the TEE based on the local private dataset and obtain the local model parameters;

[0082] Parameter processing module, used to quantize and perform TFHE processing on local model parameters within TEE;

[0083] The parameter upload module is used to export the ciphertext of the encryption model parameters from the TEE and upload it to the central server, so that the central server can perform robust aggregation processing on the encryption model parameters in the encryption domain.

[0084] An embodiment of the present invention further provides a genomics-oriented quantum-resistant robust parameter aggregation federated learning device, which is applied to a central server and includes:

[0085] The parameter receiving module is used to receive the ciphertext of the encrypted model parameters uploaded by the client. The encrypted model parameters are obtained by the client by quantizing and TFHE processing the local model parameters locally;

[0086] The parameter aggregation module is used to perform robust aggregation processing on the encrypted model parameters in the encrypted domain to obtain optimized model parameters.

[0087] In the quantum-resistant robust parameter aggregation federated learning device for genomics provided by the above embodiments of the present invention, the specific process of each module implementing its function is the same as the steps of the quantum-resistant robust parameter aggregation federated learning method for genomics provided by the above embodiments of the present invention, and its repeated description will be omitted here.

[0088] In addition, an embodiment of the present invention may also be a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, enable the processor to execute the steps in the quantum-resistant robust parameter aggregation federated learning method for genomics of various embodiments of the present invention.

[0089] Computer-readable storage media can be any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or components, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0090] It should be understood that the processor in the embodiments of the present invention may be a central processing unit (CPU), but may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0091] In summary, an embodiment of the present invention relates to a quantum-resistant robust parameter aggregation federated learning method and device based on genomics, which includes the following steps: the client quantizes and TFHE encrypts local model parameters within the TEE; the client exports the ciphertext of the encrypted model parameters from the TEE and uploads it to a central server; the central server robustly aggregates the encrypted model parameters in the encryption domain to obtain optimized model parameters and sends them to the client; the client decrypts and dequantizes the optimized model parameters and then updates the local model. The technical solution of the embodiments of the present invention, by combining TEE for local computation and TFHE for parameter transmission and aggregation, can keep the original data and model updates encrypted throughout the entire process, eliminating the risk of data leakage through inference attacks and providing a zero-exposure guarantee for sensitive single-cell RNA sequencing (scRNA-seq) data. The enhanced Krum algorithm can effectively mitigate the impact of malicious nodes by selecting updates consistent with the majority of updates, maintaining the integrity of the model even in the presence of adversarial interference. This robustness is critical for real-world deployments where the trustworthiness of all participants cannot be assumed. The use of dynamic truncation reduces the computational overhead of TFHE operations while retaining sufficient accuracy for model training and aggregation. TEE-based local training further optimizes performance by avoiding encryption during the computationally intensive training phase, balancing security and efficiency. The technical solution provided by the embodiments of the present invention supports accurate cell type classification and other single-cell RNA sequencing (scRNA-seq) tasks across distributed datasets, enabling clients to collaborate without sacrificing data privacy or model quality, and is compatible with various machine learning models.

[0092] It should be understood that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention (including the claims) is limited to these examples. Within the spirit of the present invention, the technical features of the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of one or more embodiments of the present invention described above, which are not provided in detail for the sake of clarity. The above specific embodiments of the present invention are merely intended to illustrate or explain the principles of the present invention and do not constitute a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present invention should be included within the scope of protection of the present invention. In addition, the appended claims of the present invention are intended to cover all variations and modifications that fall within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A quantum-resistant robust parameter aggregation federated learning method for genomics, characterized by: Including steps: The client quantizes and encrypts local model parameters within the TEE. The client exports the ciphertext of the encrypted model parameters from the TEE and uploads it to the central server; The central server performs robust aggregation processing on the encrypted model parameters in the encryption domain to obtain optimized model parameters and sends them to the client; The client decrypts and dequantizes the optimized model parameters and then updates the local model; The local model parameters are obtained by the client training the local model in the TEE based on the local private dataset; the robust aggregation process optimizes and updates the global model parameters based on the similarity between the encrypted model parameters of each client; The method further comprises the steps of: The client obtains the change value of each updated optimized model parameter relative to the model parameter in the previous round, and uploads the change value and the ciphertext of the encrypted model parameter in this round to the central server; The central server determines the importance of each parameter based on the change value, and performs truncation processing on the encryption model parameters of each client according to the importance, including: Integrate the absolute values ​​of the model parameter changes uploaded by each client to obtain a global parameter importance score; All parameters are sorted according to the global parameter importance score, and different truncation digits are assigned to different parameters based on the sorting results.

2. The method according to claim 1, characterized in that The local private dataset includes original single-cell RNA sequencing data; the method further includes the steps of: Preprocess the raw single-cell RNA sequencing data of the local private dataset.

3. The method according to claim 2, characterized in that The preprocessing includes removing zero-expression genes in the data, converting the original single-cell RNA sequencing data into CPM, selecting highly expressed genes based on the converted data, and performing PCA dimensionality reduction on the selected genes.

4. The method according to any one of claims 1 to 3, characterized in that The robust aggregation process includes: Calculate the Euclidean distance between each client's encryption model parameter vector; Based on the calculated Euclidean distance, the parameters of the encryption model participating in the global model are adjusted; The global model parameters are optimized and updated based on the adjustment results.

5. The method according to claim 4, characterized in that Based on the calculated Euclidean distance, the parameters of the encryption model participating in the global model are adjusted, including: For each client, calculate the sum of the Euclidean distances between the client and the encryption model parameters of other clients as the distance set of the client; Among all the client distance sets, the client model parameters corresponding to the largest distance set are eliminated.

6. The method according to claim 5, characterized in that Optimizing and updating the global model parameters based on the adjustment results includes: After removing the client model parameters corresponding to the largest distance set, the average value of the encrypted model parameter vectors of the remaining clients is calculated; The average value is used as the optimization update parameter of the global model.

7. The method according to claim 6, characterized in that The central server performs robust aggregation on the ciphertext by executing homomorphic operations in the compiled TFHE computation circuit.

8. The method according to claim 7, characterized in that The TFHE calculation circuit is pre-generated at any client according to the local model parameter vector of the client.

9. A quantum-resistant robust parameter aggregation federated learning device for genomics, characterized by: include: The client parameter encryption module is used to quantize and encrypt local model parameters within the TEE. The client parameter upload module is used to export the ciphertext of the encryption model parameters from the TEE and upload it to the central server; A central server aggregation module is used to perform robust aggregation processing on the encryption model parameters in the encryption domain to obtain optimized model parameters and send them to the client; A client parameter updating module is used to update the local model after decrypting and dequantizing the optimized model parameters; The local model parameters are obtained by the client training the local model in the TEE based on the local private dataset; the robust aggregation process optimizes and updates the global model parameters based on the similarity between the encrypted model parameters of each client; The client parameter uploading module is also used to obtain the change value of each updated optimization model parameter relative to the model parameter of the previous round, and upload the change value and the ciphertext of the encrypted model parameter of this round to the central server; The central server aggregation module is further configured to determine the importance of each parameter based on the change value, and perform truncation processing on the encryption model parameters of each client according to the importance, including: Integrate the absolute values ​​of the model parameter changes uploaded by each client to obtain a global parameter importance score; All parameters are sorted according to the global parameter importance score, and different truncation digits are assigned to different parameters based on the sorting results.

Images