Ovarian cancer multi-index combined diagnosis security federated learning method based on flexible access control

Abstract

The invention discloses an ovarian cancer multi-index combined diagnosis security federal learning method based on flexible access control, and belongs to the technical field of medical artificial intelligence and privacy calculation crossing. The method comprises three parts of a security initialization stage, federated principal component analysis and a federated genetic algorithm. In the security initialization stage, secret keys are securely distributed for a client and a server in federated learning through attribute-based encryption, homomorphic encryption and digital signature technologies, and a fine-grained access control strategy is set. In the federal principal component analysis stage, a client side locally calculates a covariance matrix, encrypts and signs the covariance matrix and then uploads the covariance matrix to a server for safe aggregation, and the server carries out characteristic decomposition and generates a disturbance projection matrix. In a federated genetic algorithm stage, a client side locally calculates fitness, a server safely aggregates and generates a new population, and a globally optimal solution is obtained through multiple rounds of iteration. According to the method, data privacy and model security are effectively protected through multiple encryption and signature mechanisms.

Description

Technical Field

[0001] This invention relates to medical artificial intelligence and privacy computing, specifically to a secure federated learning method for joint diagnosis of ovarian cancer based on flexible access control and multiple indicators. Background Technology

[0002] Early research on federated learning in the medical field primarily focused on infrastructure development and the implementation of specific application scenarios. Chen et al. proposed the FedHealth framework in 2020, combining federated learning with transfer learning for health monitoring of wearable devices. This enabled cross-institutional data collaboration without sharing raw data, but faced significant communication overhead when processing data from heterogeneous devices. Wu et al., developing the PerFit architecture around the same time for IoT medical devices, used federated distillation technology to address device heterogeneity and statistical bias, improving the accuracy of activity recognition. However, its personalization process is complex and difficult to scale to large-scale medical systems. These early works laid the foundation for the application of federated learning in the medical field, but generally faced challenges such as data heterogeneity, communication efficiency, and model generalization.

[0003] With technological advancements, research focus has gradually shifted to the practical applications of federated learning in medicine. Dayan et al.'s 2021 work applied federated learning to COVID-19 detection, achieving high diagnostic accuracy through multi-center collaborative model training. Ogier du Terrail et al. applied federated learning (FL) to breast cancer histological response prediction in 2023, achieving performance comparable to centralized learning. Guan et al.'s 2024 review systematically summarized the progress of federated learning in medical image analysis, covering tasks such as classification, segmentation, and reconstruction, and emphasizing the importance of data heterogeneity and privacy protection. Zeng et al. proposed the HFed-MIL framework in 2025, applying federated learning to ovarian cancer diagnosis. This framework utilizes knowledge distillation based on patch gradient attention to handle the heterogeneity of multi-center distributed whole-slice image data, improving model performance while protecting privacy.

[0004] Early privacy protection techniques in federated learning primarily relied on differential privacy and homomorphic encryption to prevent data leakage. In 2017, Bonawitz et al. proposed a privacy-preserving aggregation scheme based on secure multi-party computation, achieving privacy protection by adding noise and cryptographic gradients. In 2019, Truex et al. further explored hybrid privacy methods, combining differential privacy and homomorphic encryption to balance privacy protection and model performance. While these methods offer some privacy protection, they still face challenges such as the impact of noise injection on model performance and the significant computational overhead of homomorphic ciphertext operations. In 2022, Li et al. systematically investigated the application of blockchain in federated learning, utilizing smart contracts to achieve decentralized aggregation and incentive mechanisms, improving system transparency and tamper resistance. However, the high computational and communication overhead of blockchain itself limits its feasibility in large-scale medical deployments. In 2024, Shen et al. proposed a novel scheme, mMFHE, based on multi-key fully homomorphic encryption, to address the privacy leaks that may occur during model updates in federated learning and the insufficient anti-collusion capabilities of existing encryption schemes. This scheme effectively resists multi-user collusion attacks; however, the complex key management, ciphertext inflation, and intensive homomorphic operations required in multi-user scenarios also lead to high computational complexity and communication overhead. In the same year, Nuria et al.'s work provided an in-depth analysis of threat models and defense mechanisms in federated learning, emphasizing the risks of privacy leaks and collusion attacks.

[0005] These studies have improved the security of federated learning to some extent. However, for sensitive application scenarios such as healthcare, existing solutions generally fail to achieve fine-grained access control and global data privacy protection, making it difficult to meet the stringent requirements for data security and the need for dynamic management of data permissions. Summary of the Invention

[0006] This invention aims to provide a secure federated learning method for the joint diagnosis of ovarian cancer based on flexible access control, in order to solve the problems of insufficient privacy protection, lack of access control, and model performance degradation caused by cross-center data heterogeneity in the existing technology.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A secure federated learning method for multi-indicator joint diagnosis of ovarian cancer based on flexible access control, characterized by the following steps:

[0009] Safety initialization phase:

[0010] CA performs global settings

[0011] CA chooses a bilinear group of prime order. (order is a prime number) ),in yes generators, bilinear mappings It satisfies non-degeneracy and computability;

[0012] Choose a random hash function and ;

[0013] Generate common parameters .

[0014] CA sets up authorization mechanism

[0015] CA randomly selects ;

[0016] calculate ;

[0017] CA Secret Storage and publicly ;

[0018] For each attribute set CA selection and calculate ;

[0019] For attribute sets Each attribute in Define a set Represents attributes All possible values;

[0020] for Each value in CA selection and calculate ;

[0021] CA Secret Storage and publicly .

[0022] CA access policy settings are shown in the following example:

[0023] "Role: Central Server";

[0024] "Role: Client AND Agency: Authorized";

[0025] "(Role: Server OR Role: Client) and Status: Active".

[0026] Generate user key

[0027] For each client (Parallel execution):

[0028] Generate user identity private key:

[0029] Randomly select secret value And calculate;

[0030] Will and Send to CA;

[0031] CA verification Randomly select a user-exclusive secret value ;

[0032] CA compute key component: ;

[0033] CA Secret Storage and will Send to .

[0034] Generate user attribute private key

[0035] based on attribute set ,for Each attribute Specific values :

[0036] CA calculation ,in It is an attribute corresponding The number of attributes;

[0037] CA will Send to .

[0038] The complete ABE private key: ;

[0039] Similarly, the CA generates the ABE private key for the CS. .

[0040] CA uses ABE to encrypt symmetric keys.

[0041] CA generates symmetric keys: .

[0042] encryption hour:

[0043] CA according to access policy Define access structure ;

[0044] Random selection ;

[0045] Computational ciphertext component: .

[0046] right arrive : .

[0047] Output ciphertext .

[0048] Similarly, the CA uses the same encryption method. and .

[0049] Secure distribution and authorized decryption:

[0050] CA will and Send to CS, and Send to each CW.

[0051] CS Decryption process:

[0052] Phase 1 - CADecrypt (Executed by CA): CS sends a decryption request to CA. After the CA verifies the CS's identity, it processes the encrypted data. Calculate the blinding factor: and return For CS.

[0053] Phase 2 - UserDecrypt (executed by CS): CS determines the access policy... Define access structure .calculate = = .

[0054] CS computation to recover the symmetric key .

[0055] The same applies to CW decryption.

[0056] Generate other keys

[0057] CA generates Paillier homomorphic encryption key pair Use ABE encryption It is then distributed to the authorized entity.

[0058] CA distributes EdDSA signature key pairs To CS, To each CW.

[0059] CA publishes Paillier homomorphic encryption public key and all EdDSA verification keys .

[0060] Federal principal component analysis:

[0061] CA initialization parameters.

[0062] CA sets the total sample size. and target explanation rate threshold ;

[0063] CA allocates sample size to each CW ,satisfy ;

[0064] CA use Encryption parameters And then, through authorized decryption, it is distributed to the corresponding CW.

[0065] Client-side local computation (parallel execution)

[0066] For each :

[0067] use Decryption yields the parameters;

[0068] Calculate the local covariance matrix ;

[0069] Using Paillier public key encryption ;

[0070] Using EdDSA private key right sign: ;

[0071] use encryption And upload it to CS.

[0072] CS security aggregation and global computation

[0073] CS use Decryption yields all homomorphic ciphertext uploaded by CW. Verify each signature ;

[0074] CS performs homomorphic aggregation: ;

[0075] CS use right sign: ;

[0076] CS use encryption And send it to the CA.

[0077] CA decryption and return results

[0078] CA use Decrypt the data and verify the CS signature;

[0079] CA use Decryption Obtain the global covariance matrix and return For CS.

[0080] CS eigendecomposition and projection matrix generation.

[0081] CS pair Perform eigenvalue decomposition ,in For the eigenvalue matrix, It is an orthogonal eigenvector matrix;

[0082] Select the smallest This makes the cumulative explained variance ;

[0083] Constructing the projection matrix ;

[0084] Generate random orthogonal matrices ;

[0085] Calculate the perturbation projection matrix: ;

[0086] CS use right sign: ;

[0087] CS use encryption And broadcast it to all CWs.

[0088] Client-side local projection

[0089] For each :

[0090] use Decrypt the data and verify the CS signature;

[0091] Calculate local projection features .

[0092] Federated Genetic Algorithm:

[0093] CA Initializes Genetic Algorithm Parameters

[0094] CA initialization maximum algebra Cross weights Random vectors Cross rate Variation rate Variation intensity ;

[0095] CA sets the initial population (in (Population size)

[0096] CA use The parameters are encrypted and signed, and then distributed to all CWs.

[0097] Genetic algorithm iteration (generation number) arrive ).

[0098] Client-side local fitness calculation (executed in parallel).

[0099] For each :

[0100] For each individual Calculate local fitness: The objective function is unsupervised.

[0101] Select the local optimal solution ;

[0102] use sign: ;

[0103] use encryption And upload it to CS.

[0104] CS aggregation and deduplication

[0105] CS use Decrypt all data and verify the signature;

[0106] CS removes duplicates from the solution set: ;

[0107] If the current generation If yes, proceed to step A; otherwise, proceed to step B.

[0108] Step A: Generate a new population.

[0109] CS use sign ;

[0110] CS use encryption Broadcast to all CWs;

[0111] For each (Parallel execution):

[0112] use Decrypt and verify the signature;

[0113] calculate The fitness of each solution is selected. Solution to form candidate set ;

[0114] .

[0115] Repeat the following operations until the set is complete. The quantity equals :

[0116] Determine the probability of choosing a roulette wheel ;

[0117] According to probability distribution From the set Select ;

[0118] by The probability of and Perform cross operation ;

[0119] If at least one crossover or mutation operation is performed. examine or Is it in the set? If a duplicate value already exists, then... Add the new individual to the set .

[0120] Set up a new population .

[0121] Step B: Selection of the final globally optimal solution.

[0122] CS use sign ;

[0123] CS use After encryption, it is randomly sent to a CW, such as ;

[0124] use After decrypting and verifying the signature, evaluate fitness of all solutions ;

[0125] Choose the globally optimal solution: ,use The encrypted data is then returned to the client (CS).

[0126] CS Decrypt and verify After signing, The uploaded encrypted data is broadcast directly to all CWs;

[0127] All CW use Decrypt and verify The signature is used to obtain the final globally optimal solution. .

[0128] The beneficial effects of this invention are:

[0129] To address data privacy and security issues in distributed learning environments, this invention introduces a federated genetic algorithm based on federated principal component analysis and optimization. This effectively mitigates the negative impact of non-independent and identically distributed data while protecting the data privacy of all parties, significantly improving the convergence speed and generalization performance of the global model. Furthermore, by constructing a full-link security framework integrating various cryptographic primitives such as attribute-based encryption, homomorphic encryption, and digital signatures, it achieves attribute-based access control, authorized decryption, authentication, and transmission protection for model parameters. This provides comprehensive and auditable security guarantees for cross-institutional collaborative training, encompassing authentication, channel security, and encrypted computation. Attached Figure Description

[0130] Figure 1 System model diagram of the present invention;

[0131] Figure 2 Detailed flowchart of the present invention;

[0132] Figure 3 Comparison chart of experimental performance of the present invention. Detailed Implementation

[0133] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings:

[0134] The system model of this framework mainly consists of three types of entities: the Central Authorization Center (CA), the Central Server (CS), and the Client Site (CW). Their interaction relationships are as follows: Figure 1 As shown.

[0135] The characters are described as follows:

[0136] The CA (Certified Authority) is the trusted institution in this framework. In medical scenarios, it is usually the medical management agency. Its main responsibilities include: performing system initialization and generating public parameters; providing identity authentication for entities within the system; creating access rules and allocating key pairs to entities; formulating federated learning tasks; and providing security foundation support for the entire cryptographic system.

[0137] A Client-Side (CS) database is often considered a semi-trusted entity, and in healthcare scenarios, it frequently corresponds to a data analytics center. Its main functions are: receiving encrypted local model parameters from various client sites; providing data storage, computation, and analysis services; performing aggregation operations on model parameters and publishing the aggregated global parameters.

[0138] A Chained Workstation (CW) is typically a group of terminal devices or workstations, representing healthcare service institutions such as hospitals or research institutes in a medical setting. Its core functions are: to perform computation and analysis on local data; to collaborate with Computer Controllers (CS) to complete the federated learning training process; and to update the local model.

[0139] The system operation process is as follows:

[0140] The specific process of this plan is as follows: Figure 2 As shown.

[0141] The entire method comprises three core algorithms: a secure initialization phase, federated principal component analysis (PCA), and a federated genetic algorithm. The secure initialization phase is led by a Certificate Authority (CA), which constructs a security framework with fine-grained access control capabilities based on an attribute-based authorization scheme. This is achieved by generating a system public key and master key, generating attribute private keys based on the attribute sets of each participant, and defining access policies, thus establishing a complete access control system. The EdDSA digital signature, AES-GCM symmetric encryption, and Paillier homomorphic encryption used in this invention directly adopt existing mature solutions. Federated principal component analysis (PCA) achieves multi-center collaborative PCA while strictly adhering to privacy principles. This method performs PCA on local data, using orthogonal transformations to synthesize highly correlated high-dimensional test indicators into a few independent new variables. PCA simplifies the data structure while retaining most of the information from the original indicators, facilitating the construction of disease diagnostic models in a low-dimensional space. Federated PCA is the core algorithm for achieving distributed feature extraction under privacy protection within this framework. The federated genetic algorithm optimizes the model through a distributed evolutionary strategy. Each client calculates its local fitness in each generation and uploads the signed optimal solution. The central server coordinates the deduplication and broadcasting of the solution set. Finally, after a specified number of generations, the selected clients evaluate and output the globally optimal model with signed verification, thus completing the entire privacy-preserving federated feature learning and optimization process.

[0142] We systematically reviewed and analyzed the computational overhead of each cryptographic primitive in the proposed method, and the results are shown in Tables 1 and 2. From the perspective of computational complexity, the main overhead in the security initialization phase comes from the bilinear pairing operation in attribute-based encryption, and its complexity is linearly related to the number of clients; the core overhead of federated principal component analysis lies in the modular exponentiation operation in the Paillier homomorphic encryption system, and its computational load is proportional to the square of the feature dimension; the performance bottleneck of the federated genetic algorithm is reflected in the EdDSA signature verification operation, and the total overhead is related to the product of the number of generations of the genetic algorithm and the number of clients.

[0143] It is worth emphasizing that despite the aforementioned computational burden, high-overhead operations such as bilinear pairing are limited to the system startup phase, and their costs can be effectively amortized over long-term operation. Furthermore, the inherent parallel computing characteristics of the federated learning framework allow most cryptographic operations to be executed in parallel across participating nodes, significantly reducing the computational pressure at single points. This enables our solution to achieve a good balance between security strength and execution efficiency while strictly meeting medical data security requirements, ensuring the feasibility of the system in real-world application environments.

[0144]

[0145] Table 1: Statistics of Encryption Operation Computational Overhead

[0146] Table 2: Symbol Meaning Table

[0147] To verify the performance of the proposed federated learning framework, this study constructed a three-center case-control simulated dataset, referencing the case-control composition and distribution parameters of 28 test indicators reported by Zhang et al. This dataset includes demographic characteristics and laboratory indicators related to ovarian cancer risk prediction, and its distribution is designed to simulate real-world clinical conditions. The application of federated learning is primarily used to construct a distributed training environment to evaluate the algorithm's effectiveness in avoiding data centralization. It is important to emphasize that all experimental procedures did not involve any real patient privacy information and did not require ethical approval.

[0148] This simulated dataset generated 1804 simulated individuals, with sample sizes at the three sites: Site A (754 cases), Site B (450 cases), and Site C (600 cases). The overall sample included four population groups: ovarian cancer patients (n=443), patients with other gynecological malignancies (n=330), patients with benign gynecological diseases (n=811), and healthy individuals undergoing physical examinations (n=220). Based on literature statistical characteristics, 28 routine laboratory test indicators were simulated and generated, and derived ratios were calculated based on these indicators. Basic demographic variables were also included.

[0149] Our simulation experiments were conducted on a laptop equipped with an Intel i9-13980HX processor, an NVIDIA GeForce RTX 4060 Laptop graphics card, 32GB of RAM, and Windows 11 Version 24H2. This device was used to simulate scenarios involving different stakeholders, including clients, a central server, and a trusted center. For automated testing, programmatic development was based on Python 3.11, utilizing libraries such as NumPy 1.26 and Pandas 2.0.

[0150] On an independent test set of 602 subjects, the system evaluated the model performance of seven modeling strategies by comparing the area under the receiver operating characteristic curve (AUC), including: the traditional CA125 single-index model, the centralized training model, the local independent model at three sites, the classic federated average algorithm (FedAvg), and our proposed FAC-SFL-OC framework.

[0151] Figure 3 The results showed that the traditional CA125 single-indicator model had an overall AUC of 0.852, a sensitivity of 81.8%, a specificity of 73.8%, an accuracy of 75.8%, and an F1 score of 0.624. This indicates that the discrimination ability of a single tumor marker in multicenter mixed populations is limited, and the false positive rate is relatively high.

[0152] After centrally modeling multiple clinical test indicators at the central hub, the model's AUC improved to 0.926, with sensitivity and specificity of 83.1% and 85.0%, respectively, accuracy reaching 84.6%, and an F1 score of 0.725. This strategy represents the "theoretical upper limit" of the model's performance, but in real-world medical scenarios, its reliance on cross-institutional aggregation of raw data presents privacy and compliance constraints.

[0153] The performance of the local models at the three sites showed significant differences. The model at site A had the highest AUC (0.934) and a specificity of 95.3% on the local test data, but its sensitivity was low (66.1%). The model at site B had an AUC of 0.892, with a sensitivity of 70.3% and a specificity of 92.9%. The model at site C performed relatively poorly, with an AUC of 0.841, a sensitivity of 63.2%, and a specificity of 85.4%. Although some local models achieved high AUCs or specificities on the local data, their overall sensitivity was generally insufficient, reflecting that when the sample size is limited and the disease spectrum varies, single-center models are prone to overfitting the local data distribution, resulting in limited cross-center generalization ability.

[0154] FedAvg achieved an AUC of 0.903, a sensitivity of 79.1%, a specificity and accuracy of 78.9%, and an F1 score of 0.648 on the overall test set. Compared to a purely local model, FedAvg mitigated performance fluctuations to some extent, but its overall performance was still significantly lower than that of centralized training schemes, indicating a performance penalty when processing such complex medical data.

[0155] Without exposing any raw data, the proposed FAC-SFL-OC framework achieves diagnostic efficacy close to that of centralized training. The model achieves an AUC of 0.915 on the overall test set, significantly outperforming the traditional CA125 single-index model and FedAvg, and approaching the performance of centralized models. Furthermore, it achieves a more balanced trade-off between sensitivity (80.4%), specificity (84.6%), and accuracy (85.7%), with an improved F1 score of 0.706. Figure 3 As can be seen from the receiver operating characteristic curves, the curves of FAC-SFL-OC are generally above the traditional CA125 single-index model and FedAvg, and basically overlap with the centralized training curves. This indicates that under strict privacy constraints, this method can still maintain a discrimination ability close to the ideal performance upper limit.

[0156] We recorded the computational and communication overhead data for the encryption and decryption process in 20 rounds of simulation experiments and presented the average values. As shown in Table 3, in the communication overhead test of attribute-based encryption and decryption of symmetric keys, as the attribute-based encryption key length increased from 256 bits to 2048 bits, the size of the attribute-based encrypted ciphertext correspondingly increased from 64 bytes to 512 bytes, the encryption time increased from 0.17 milliseconds to 29.99 milliseconds, and the decryption time increased from 0.13 milliseconds to 16.59 milliseconds. This shows that even at a high security level, encryption and decryption operations can still be completed within milliseconds, demonstrating that the scheme maintains acceptable performance loss while achieving fine-grained access control.

[0157]

[0158] Table 3: Performance overhead of attribute-based encryption and decryption of symmetric keys.

[0159] As shown in Table 4, in the communication overhead test of the symmetric encryption / decryption model parameters, encrypting 3716 bytes of plaintext using a 256-bit key yields 3748 bytes of ciphertext. The encryption time is only 0.0357 milliseconds, and the decryption time is 0.154 milliseconds. This extremely low latency highlights the high efficiency of the scheme in protecting communication channels. Authentication encryption ensures the confidentiality and integrity of data transmission, thereby improving the overall system's response speed and reliability.

[0160]

[0161] Table 4: Performance Cost of Symmetric Encryption / Decryption Parameters

[0162] As shown in Table 5, in the communication overhead analysis of EdDSA signature verification, for 3748 bytes of ciphertext, the signing time using a 512-bit key is 0.203 milliseconds, and the verification time is 0.078 milliseconds. This indicates that the scheme introduces almost imperceptible overhead in achieving strong authentication and data integrity verification, further consolidating the system's authenticability and effectively preventing unauthorized access and tampering risks.

[0163]

[0164] Table 5: Performance overhead of EdDSA signature symmetric ciphertext

[0165] Overall, the experimental results verify that the various cryptographic components used in this scheme perform well in terms of performance overhead, fully demonstrating their superior ability to balance security and operational efficiency in a federated learning environment. Our scheme maintains low communication and computational overhead even with multiple security mechanisms in place, providing a secure and efficient privacy protection framework for ovarian cancer federated learning applications.

Claims

1. A secure federated learning method for multi-indicator joint diagnosis of ovarian cancer based on flexible access control, characterized in that, This includes the establishment of a system model, which comprises three roles: Central Authority (CA), Central Server (CS), and Client Workstation (CW). The Central Authority (CA) is responsible for performing system initialization and generating public parameters; providing identity authentication for entities within the system; creating access rules and allocating key pairs to entities; formulating federated learning tasks; and providing security foundation support for the entire cryptographic system. The central server (CS) receives encrypted local model parameters from each client site; provides data storage, computation, and analysis services; performs aggregation operations on the model parameters and publishes the aggregated global parameters. The client workstation (CW) is responsible for calculating and analyzing local data. Collaborate with Computer Science to complete the federated learning training process; Complete the local model update.

2. The secure federated learning method for multi-indicator joint diagnosis of ovarian cancer based on flexible access control as described in claim 1, characterized in that, Including system operation steps: Step A: The security initialization phase, primarily executed by the Central Authorization Center (CA), aims to build a secure foundational environment, providing encryption guarantees and authentication mechanisms for subsequent federated learning tasks. Its steps are as follows: Step A-1: ​​CA performs global settings, sets system common parameters, and establishes a mathematical foundation for the security model; Step A-2: The CA sets up an authorization mechanism and generates a series of security parameters to strengthen the management of decryption permissions; Step A-3: The CA generates and sends the corresponding identity key and attribute key to the Central Server (CS) and Client Workstation (CW); Step A-4: The CA generates a symmetric key and encrypts it using an attribute base, then sends it to the CS and CW; Step A-5: Under the authorization of the CA, CS and CW establish a secure channel by decrypting the symmetric key ciphertext; Step A-6: The CA generates other encryption keys required by the security system; Step B: The Federated Principal Component Analysis (FPC) phase, executed collaboratively by CA, CS, and CW, is used to achieve global principal component extraction and feature dimensionality reduction across sites while protecting data privacy. The steps are as follows: Step B-1: CA initializes the parameters used for federated principal component analysis and sends them to CW via symmetric encryption; Step B-2: CW uses a symmetric key to decrypt and obtain the initial parameters, calculates the covariance matrix locally, and then homomorphically encrypts the result before signing and sending it to CS; Step B-3: CS receives all homomorphic ciphertext uploaded by CW and verifies the signature, performs homomorphic aggregation operation, signs the aggregated result and sends it to CA to request decryption; Step B-4: The CA verifies the CS signature, decrypts the aggregation result of the CS using the homomorphic private key, and returns the plaintext result to the CS; Step B-5: CS performs feature decomposition on the aggregation result of the plaintext state, constructs the projection matrix, finally obtains the perturbation projection matrix, and after symmetric encryption of the matrix, signs and sends it to all CWs; Step B-6: CW verifies the CS signature, decrypts the perturbation projection matrix using the symmetric key, and uses it to calculate the local projection features; Step C: Federated Genetic Algorithm, executed jointly by CA, CS, and CW, is used to find the globally optimal model parameters in an encrypted environment. Its steps are as follows: Step C-1: CA initializes the parameters used for the federated genetic algorithm and sends them to CW via symmetric encryption; Step C-2: CW decrypts to obtain the initial parameters, starts local fitness calculation, obtains the local optimal solution, signs it, symmetrically encrypts it, and uploads it to CS; Step C-3: CS uses the symmetric key to decrypt the optimal solution uploaded by CW, verifies the signature, and removes duplicates from the solution set. If the current solution is not the last generation, then proceed to step C-4. If the current generation is already the last generation, proceed directly to step C-6; Step C-4: CS signs the deduplicated solution set and sends it to all CWs using symmetric encryption; Step C-5: CW decrypts and verifies the signature, calculates the fitness of each solution in the solution set, selects the best of several to form a new solution set, the number of which is initially determined by CA, and then performs roulette wheel selection, crossover operation with probability, and mutation operation with probability on the solution set. After removing duplicates, the results are added to the new solution set until the number of solutions is sufficient. The new solution set is then used as the new generation population to repeat the above operations. Step C-6: CS randomly assigns a CW, signs the deduplicated solution set, and sends it to the assigned CW using symmetric encryption; Step C-7: The designated CW decrypts and verifies the signature, calculates the fitness of each solution in the solution set, selects the optimal solution, signs it, and sends it to the CS using symmetric encryption; Step C-8: After CS decrypts and verifies the signature, it sends the result uploaded by CW to other CWs as is. Step C-9: Decrypt all CWs and verify the signatures to obtain the final globally optimal solution.

3. The secure federated learning method for multi-indicator joint diagnosis of ovarian cancer based on flexible access control as described in claim 1, characterized in that, Including related algorithms: Algorithm A's safe initialization phase: Input: Set of client workstations, security parameters; Output: Security parameters used to protect training; Algorithm B: Federated Principal Component Analysis Input: A collection of client workstations, local standardized data; Output: Local projection features; Algorithm C: Federated Genetic Algorithm Input: A set of client workstations, and the projection characteristics of each client; Output: Global optimal solution.

Images