Method and system for carrying out model privacy protection in federated learning system
Through the improved distributed ElGamal encryption system IDEC algorithm and secret sharing mechanism, the error problems caused by privacy data leakage and disconnected users in the federated learning system are solved, and efficient and secure model training and aggregation are achieved, reducing computing and communication overhead , promoting cross-domain collaboration.
Patent Information
- Application Number
- CN202411086837.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-19
AI Technical Summary
There are risks of privacy data breaches and errors caused by disconnected users during model training in large-scale federated learning systems, and the computing cost and communication overhead of the prior art are relatively high.
The improved distributed ElGamal encryption system IDEC algorithm is used to encrypt the model, and the secret sharing mechanism of the key server ensures the secure distribution of the decryption key, ensuring the security of the participants' private data during transmission and storage.
Significantly improves the security of private data, reduces computing costs and communication overhead, enhances the security of model aggregation and system robustness, and supports cross-domain collaboration and knowledge sharing.
Smart Images

Figure CN118981792A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer application technology, and in particular, relates to a large-scale federated learning system scenario in which there are malicious participants or semi-honest aggregation servers, at which point the privacy data of the participants is at risk of being leaked and out of control. It also targets the situation in the federated learning system where unexpected offline users cause errors in model aggregation and even privacy information leakage. Background Art
[0002] In recent years, machine learning has become a common means of data processing and analysis. Many researchers have used machine learning methods to achieve many tasks that previously required a lot of manpower and material resources. They use public or private data to train machine learning models so that they can replace manual identification, classification, and creation. For example, the application of fingerprint recognition and face recognition technology in access control systems, automatic image classification in the field of computer vision, and generative artificial intelligence creation tools such as the now very popular "ChatGPT Creation" are all examples of machine learning applications. In addition, machine learning has continued to derive a number of important branches, including deep learning, reinforcement learning, supervised learning, and semi-supervised learning. Each branch has demonstrated strong capabilities and potential in its own field.
[0003] However, these centralized machine learning methods all face a common problem, namely, data privacy. Large-scale machine learning requires extremely large amounts of data, but this data contains other people's private information. This centralized machine learning method requires all data to be centralized on the server for model training, which will lead to personal privacy information leakage and loss of control. In order to address this problem, distributed machine learning came into being, the most famous of which is the federated learning framework proposed by Google in 2016. It uses a unique distributed method to allow each participant to use their own data for model training locally, and then only upload the trained model updates (not the original data) to the central server for aggregation. This strategy significantly reduces the risk of personal privacy information leakage.
[0004] However, research in the field of federated learning in recent years has shown that even if an individual's private data is retained on the local client for training, there is still a risk of privacy leakage. This is because a curious server or malicious client can obtain an individual's private information by performing mathematical analysis on the model uploaded by the participants. Later, some researchers proposed that only the gradient information of the training model should be uploaded instead of the complete model, which further reduces the risk of privacy information leakage. However, the latest research shows that even if such measures are taken, uploading only the gradient information may still leak an individual's private data in certain situations.
[0005] In addition, in the process of federated learning, common systems require participants to stay online at all times in order to complete the training process normally. Once a participant accidentally drops offline, the training process will be interrupted or the aggregation model will be wrong. In the current federated learning solution, homomorphic encryption technology is usually used to solve the privacy protection problem, but the computing cost consumed by fully homomorphic encryption and the communication overhead between devices are huge.
[0006] In summary, the problems existing in the prior art are: (1) In large-scale federated learning systems, there is a risk of leakage of participants’ private data; (2) During the model training process, some participants accidentally dropped out of the network, resulting in training errors; (3) The participating devices face huge computing costs and communication overheads. Summary of the invention
[0007] In view of the problems existing in the prior art, the present invention provides a privacy protection method applied to a large-scale federated learning system, aiming to ensure the security of the privacy data of all participants.
[0008] The present invention is implemented as follows: a method for protecting model privacy in a federated learning system, the method comprising the following steps: a. Participants use private data to train models locally; b. The participants use the improved distributed ElGamal encryption system IDEC algorithm to encrypt the trained model; c. The participants upload the encrypted model to the aggregation server; d. The aggregation server aggregates the received encryption model; e. The participants download the aggregated model from the aggregation server and obtain the decryption key by interacting with the key server to achieve secret sharing; f. The participants use the decryption key to decrypt the aggregated model and update the local model for the next round of training.
[0009] Further, the model encryption process using the improved distributed ElGamal encryption system IDEC algorithm includes the following steps: a. The key server generates the encryption key for the participants; b. The participants use the generated key to encrypt the locally trained model using the distributed ElGamal encryption system IDEC algorithm; c. The encrypted model exists in ciphertext form to ensure security during transmission and storage.
[0010] Further, the aggregation server aggregates the received encryption models, including the following steps: a. The aggregation server receives the encrypted models from each participant; b. The aggregation server performs aggregation calculation on the received encryption model to obtain an aggregated encryption model; c. The aggregation server returns the aggregated encrypted model to the participants to ensure that the privacy data of any participant is not leaked during the entire aggregation process.
[0011] Further, decrypting the aggregated model and updating the local model includes the following steps: a. The participants download the aggregated encrypted model from the aggregation server; b. The participants interact with the key server and obtain the decryption key through secret sharing; c. The participants use the decryption key to decrypt the aggregated encrypted model to obtain the decrypted model; d. The participants update the local model based on the decrypted model for the next round of model training and encrypted upload.
[0012] Furthermore, the IDEC algorithm of the distributed ElGamal encryption system is specifically: 1) Key generation: Given a modulus that is prime The multiplicative cyclic group of and one of its generators , each participant From the multiplicative cyclic group Randomly select a private key , and then the public key Sent to the key server, which then calculates the joint public key of all participants ; 2) Encryption: In the multiplication cycle group The ciphertext is determined in and joint public key After that, all participants share the same random number , is from The joint public key is then used The encryption process is as follows: , where C is the encrypted ciphertext, Represents a part of the ciphertext, Represents ciphertext data, It is an encrypted process value. It is part of the encryption result and one of the process values of the encryption operation; 3) Decryption: After receiving the secret sharing share from the key server, each participant uses the secret sharing to reconstruct the joint private key , each participant Get Message By the following equation: , ciphertext It can be quickly obtained by table lookup.
[0013] Another object of the present invention is to provide a federated learning system that implements the method of protecting model privacy in a federated learning system, comprising: a participant, responsible for using private data locally to train the model, and uploading the trained model after encryption to an aggregation server; an aggregation server, responsible for receiving encrypted models uploaded by each participant, and performing aggregation processing on these models; a key server, responsible for generating and distributing keys to the participants so that the participants can decrypt the downloaded aggregated models; wherein the model encryption adopts the distributed ElGamal encryption system IDEC algorithm to ensure the security and privacy protection of the model information in the entire learning system.
[0014] Furthermore, the model encryption and decryption method in the federated learning system adopts the improved distributed ElGamal encryption system IDEC algorithm. After the participating party encrypts the local model, it uploads it to the aggregation server for aggregation. After the aggregation is completed, the participating party downloads the aggregated model from the aggregation server, and obtains the decryption key through secret sharing interaction with the key server, so as to decrypt the aggregated model and then update the local model for the next round of training.
[0015] Another object of the present invention is to provide a key management method for a federated learning system, in which a key server is responsible for generating and distributing keys to participants, and a secret sharing mechanism is used to ensure the secure distribution of decryption keys. Participants update local models by decrypting the obtained aggregated model. This method ensures the security of the privacy information of the participants, while supporting the model information to exist in a confidential state in the entire learning system.
[0016] Another object of the present invention is to provide a privacy protection method for a federated learning system that uses an improved distributed ElGamal encryption system IDEC algorithm combined with a secret sharing mechanism. The method uses a distributed ElGamal encryption system IDEC algorithm to encrypt the model at the participating party, and uses a secret sharing mechanism to interact with a key server to obtain a decryption key, thereby improving the privacy protection performance of the federated learning system.
[0017] In summary, the advantages and positive effects of the present invention are:
[0018] First, the present invention greatly improves the privacy security of large-scale federated learning systems, so that the data of both participants and servers are effectively protected. At the same time, it solves the adverse effects of some participants being offline on the entire federated learning system, and reduces the communication overhead between participants and servers while maintaining low computing costs.
[0019] The difficulty and significance of solving the technical problems existing in existing technologies lies in protecting personal privacy data from being stolen, while improving the robustness of federated learning and reducing the computing cost and communication overhead during model training.
[0020] Second, the embodiments of the present invention have achieved significant technological progress by introducing an improved federated learning system, which is specifically reflected in the following aspects: 1. Enhanced data privacy protection: By using the improved distributed ElGamal encryption system IDEC algorithm to encrypt the participants' models and combining the secret sharing mechanism to ensure the secure distribution of keys, the present invention significantly improves the privacy protection level of data during transmission and processing. Compared with the existing technology, this method effectively prevents the leakage of privacy information during model training and ensures the security of the participants' data.
[0021] 2. Improved security of model aggregation: On the aggregation server, by securely aggregating the encrypted models uploaded by each participant, the embodiment of the present invention ensures that during the entire model training and aggregation process, no single entity can access the original data or decrypted model information, significantly improving the security of the entire system and its ability to resist attacks.
[0022] 3. Optimized system performance and efficiency: The present invention not only strengthens privacy protection but also improves the operating efficiency of the system through the improved distributed ElGamal encryption system IDEC algorithm and optimized key management method. The distributed ElGamal encryption system IDEC algorithm reduces the computational complexity in the encryption and decryption process while ensuring high security. Compared with the traditional federated learning system, the present invention can complete the training and updating of the model more quickly, effectively shortening the learning cycle.
[0023] 4. Promotes cross-domain collaboration: The embodiments of the present invention provide a secure and efficient model training and aggregation mechanism, which lowers the threshold for collaboration between different participants, allowing participants from different fields and with different data privacy requirements to participate in federated learning projects while ensuring data security. This progress has greatly expanded the application scenarios of federated learning and promoted the sharing and utilization of cross-domain knowledge.
[0024] In summary, the embodiments of the present invention have achieved significant technical progress in data privacy protection by introducing improved encryption technology and key management methods in the federated learning system, while also improving the security, efficiency and collaboration capabilities of the system, and promoting the development and application of federated learning technology.
[0025] Third, the existing technical problems solved by the present invention and the technical effects obtained are summarized as follows: Technical issues solved: 1. Data privacy leakage problem: In federated learning systems, data privacy protection is a major challenge. Traditional methods cannot effectively prevent privacy leakage during model training and sharing, especially when multiple parties are involved.
[0026] 2. Model security issues: Without proper encryption mechanisms, the model may be tampered with or stolen during transmission or storage, threatening the security and reliability of the entire learning system.
[0027] 3. Model aggregation efficiency issue: In a federated learning system, how to efficiently aggregate encrypted models from different participants while ensuring privacy protection during the aggregation process is a key technical challenge.
[0028] Technical effects obtained: 1. Enhanced data privacy protection: By using the improved distributed ElGamal encryption system IDEC algorithm, the present invention significantly enhances data privacy protection during model training and sharing, ensuring that the privacy data of the participants will not be leaked.
[0029] 2. Enhanced model security: The encryption and key management methods in the present invention improve the security of the model during transmission and storage, preventing the model from being accessed or tampered with by unauthorized parties.
[0030] 3. Improved aggregation efficiency and accuracy: By implementing aggregation calculations on the aggregation server, the present invention can efficiently process encrypted models from different participants, which not only improves the aggregation efficiency but also ensures the accuracy of model aggregation through precise encryption and decryption processes.
[0031] 4. Reliability of model updates and iterations: By allowing participants to download the aggregated model and perform decryption updates, the present invention supports continuous model iteration and optimization, improving the dynamic adaptability and long-term reliability of the entire federated learning system.
[0032] Through advanced encryption technology and key management methods, the present invention effectively solves the problems of data privacy, model security and aggregation efficiency in federated learning systems, and provides a safe, reliable and efficient model training and sharing solution for all walks of life. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 is a system flow chart provided by an embodiment of the present invention; Figure 2 is a prediction accuracy effect diagram provided by an embodiment of the present invention; Figure 3 It is a communication overhead diagram of each stage provided by an embodiment of the present invention; Figure 4 is a training loss and accuracy graph provided by an embodiment of the present invention; Figure 5 is a test loss and accuracy diagram provided by an embodiment of the present invention; Figure 6 is an encryption operation time diagram provided by an embodiment of the present invention; Figure 7 is a decryption operation time diagram provided by an embodiment of the present invention; Figure 8 This is the test effect diagram of the original method; Fig. 9 This is the test effect diagram of this method; Fig.10 This is the test effect diagram of all samples of this method; Fig.11 This is the training effect diagram under the conditions of this method; Fig.12 This is the test effect diagram under the conditions of this method. DETAILED DESCRIPTION
[0034] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0035] The following are two specific application examples of the model privacy protection method in the federated learning system: Application Example 1: Privacy Protection of Federated Learning in the Financial Field In the financial field, data protection and privacy rights are crucial because data involves user privacy and asset security. Suppose multiple banks want to work together to develop a risk assessment model, but do not want to share customer data directly. At this time, federated learning and model privacy protection methods can come in handy.
[0036] As a participant, each bank uses its own customer data to train the model locally. During the training process, the bank uses the improved distributed ElGamal encryption system IDEC algorithm to encrypt the model to ensure the security of the model during transmission. The encrypted model is uploaded to a central aggregation server. The aggregation server aggregates these encrypted models to form a global risk assessment model.
[0037] Each bank then downloads the aggregated model from the aggregation server. By interacting with the key server, the bank obtains the decryption key, decrypts the model, and updates it locally for the next round of training. In this way, banks can collaborate to develop a more accurate risk assessment model without directly sharing data, while protecting the privacy rights of customers.
[0038] Application Example 2: Privacy Protection of Federated Learning in Healthcare Field In the healthcare field, patients' medical records are highly sensitive private information. Suppose multiple hospitals want to collaborate on developing a disease prediction model, but also want to avoid directly sharing patient data.
[0039] As a participant, each hospital uses local patient data to train the model. The trained model is encrypted by the distributed ElGamal encryption system IDEC algorithm and uploaded to the central aggregation server. The aggregation server aggregates the encrypted model to form a global disease prediction model.
[0040] The hospital downloads the aggregated model from the aggregation server, obtains the decryption key through interaction with the key server, decrypts the model, and updates it locally for the next round of training. In this way, hospitals can work together to develop a more accurate disease prediction model while protecting patient privacy, which helps improve the quality and efficiency of medical services.
[0041] These two examples demonstrate the practical application of the model privacy protection method in the federated learning system. Whether in the financial field or the medical and health field, this method can effectively protect data privacy while promoting multi-party cooperation to jointly develop more accurate and useful models.
[0042] The following are two specific examples of model privacy protection in a federated learning system: Example
[0043] In a medical data federated learning scenario, multiple hospitals jointly participate in a model training task for disease prediction. Due to the high sensitivity of medical data, the participants need to ensure the privacy of the data during the model training process.
[0044] First, each hospital uses its own medical data to train the model locally. Then, the trained model is encrypted using the Improved Distributed ElGamal Encryption System (IDEC) technology. The encrypted model is securely uploaded to the aggregation server.
[0045] After receiving the encrypted models from various hospitals, the aggregation server performs aggregation calculations to obtain the aggregated encrypted model. During this process, the aggregation server cannot obtain any information about the original model or data, ensuring privacy protection.
[0046] Next, each hospital downloads the aggregated encrypted model from the aggregation server. By interacting with the key server, each hospital can obtain the decryption key and decrypt the aggregated model. The decrypted model is used to update the local model for the next round of model training and encrypted upload.
[0047] In this way, multiple hospitals can jointly train a disease prediction model with superior performance without disclosing their respective medical data. Example
[0048] In a cross-regional federated learning project for financial fraud detection, multiple financial institutions jointly participated in the training of the model. Due to the sensitivity of financial data, the participants need to ensure the privacy of the data throughout the training process.
[0049] First, each financial institution uses its own financial data to train the model locally. Then, the trained model is encrypted using the improved distributed ElGamal encryption system IDEC algorithm to ensure the security of the model during transmission.
[0050] The encrypted model is uploaded to the aggregation server. The aggregation server performs aggregation calculations on the received encrypted models to obtain the aggregated encrypted model. During this process, the aggregation server cannot obtain any information about the original model or data, thereby protecting the privacy of the participants.
[0051] Subsequently, each financial institution downloads the aggregated encrypted model from the aggregation server. Through interaction with the key server, each financial institution can obtain the decryption key and decrypt the aggregated model. The decrypted model is used to update the local model for the next round of model training and encrypted upload.
[0052] In this way, multiple financial institutions can jointly train an efficient and accurate financial fraud detection model without leaking their own financial data, thereby improving the security of the entire financial industry.
[0053] These two embodiments demonstrate the practical application of model privacy protection in a federated learning system. By adopting an improved distributed encryption system and a secure model aggregation method, the privacy of the data of the participants is ensured, while the joint training and optimization of the model is achieved.
[0054] Federated learning is a new type of distributed machine learning method that can effectively solve the problem of "data islands". However, the federated learning methods currently used in the field still have a large risk of privacy leakage and the problem of offline users affecting the training process. In response to these problems, this method makes innovative improvements to the existing large-scale privacy-preserving federated learning system based on homomorphic encryption, improves the security performance of the system, ensures the privacy data security of the participants, and improves the stability of the system. Even if there are offline users during the model training process, the model training and aggregation process can still be completed correctly. In addition, while maintaining low computational overhead, this method also reduces communication costs.
[0055] The entire federated learning system of the present invention includes three parts: participants, aggregation servers and key servers. Participants use private data to train models locally. After each round of model training is completed, the trained models are uploaded to the aggregation server. The server is responsible for aggregating the models uploaded by each participant. The key server is responsible for generating and distributing keys for the participants. The locally trained model is encrypted using homomorphic encryption technology, and then uploaded to the central server for model aggregation. After the model aggregation is completed, each participant downloads the aggregated model from the central server, and then interacts with the key server to realize secret sharing to obtain the decryption key. After decrypting the aggregated model, the local model is updated and the next round of model training is performed. The central service area is only responsible for aggregating the encrypted model, and the key server is only responsible for the generation and distribution of keys. All model information exists in a secret state in the learning system, which can ensure the security of the privacy information of the participants. This method improves the Lifted Distributed ElGamal Cryptosystem, and the upgraded encryption system is called the distributed ElGamal encryption system IDEC algorithm. The effect and security performance of its privacy protection are greatly improved. The innovative IDEC is used on the participant side to encrypt and decrypt the model.
[0056] This method superimposes the distributed ElGamal encryption system IDEC algorithm and secret sharing in a novel way to improve the privacy protection performance of the federated learning system.
[0057] This method solves the problem of private information leakage of participants when there are malicious participants or curious servers in large-scale federated learning systems. At the same time, it overcomes the problem that some current federated learning systems cannot complete the aggregation step normally when there are participants who accidentally disconnect during model training, which affects the final learning effect. In addition, while maintaining low computational costs, this method reduces the communication overhead of some privacy-preserving federated learning systems based on homomorphic encryption technology.
[0058] The improved distributed ElGamal encryption system IDEC algorithm provided by the present invention: 1. Key generation: Given a modulus that is prime The multiplicative cyclic group of and one of its generators , each participant From the multiplicative cyclic group Randomly select a private key , and then the public key Sent to the key server, which then calculates the joint public key of all participants ; where n is the number of participants.
[0059] 2. Encryption: In the multiplication cycle group The ciphertext is determined in and joint public key After that, all participants share the same random number , is from The joint public key is then used The encryption process is as follows: , where C is the encrypted ciphertext, , ; 3) Decryption: After receiving the secret sharing share from the key server, each participant uses the secret sharing to reconstruct the joint private key , each participant Get Message By the following equation: , ciphertext It can be quickly obtained by table lookup.
[0060] The security of this method is guaranteed by the ElGamal cryptosystem and IDEC under the assumption of Decisional Diffie-Hellman (DDH).
[0061] We establish the semantic security of IDEC and evaluate its security against collusion attacks while ensuring model privacy.
[0062] Theorem 1 (Semantic Security): If the ElGamal cryptosystem is semantically secure, then IDEC is also semantically secure.
[0063] Proof: Given and , the opponent cannot get , the opponent cannot be sure (g is the generator, c is the ciphertext, is a process value that the opponent can calculate), nor can it be calculated by Multiply by the inverse To calculate .in Since the semantic security of ElGamal is based on DDH and the discrete logarithm hard problem, IDEC is semantically secure.
[0064] Theorem 2 (Protection against collusion and dependence): If ElGamal and IDEC exhibit semantic security, they are also resistant to collusion and information harvesting.
[0065] Proof: In our ElGamal cryptosystem, the server outputs , the ciphertext can be obtained by When the server colludes with the attacker, or the attacker eavesdrops on the ciphertext data and all When the attacker can analyze the plaintext information The proposed IDEC method uses SS to obtain the decryption key. In addition, less than 80% of the participants cannot reconstruct the secret, so security is guaranteed. and When he could not obtain , and the ciphertext information cannot be obtained ;in Process values calculated for the server, A name for each process value.
[0066] Theorem 3 (Confidentiality of trained models): If the IDEC scheme guarantees semantic security, then the privacy of the model is also protected.
[0067] Proof: During the training phase, participants cipher the parameter Send it to the aggregation server, which aggregates the trained model to obtain the ciphertext result The data on the server is encrypted and the encrypted model is then returned to the participant. Due to the semantic security of the IDEC protocol, the privacy model is robustly protected.
[0068] Example 1: Federated Learning System Across Medical Institutions In a federated learning project across medical institutions, each medical institution (participant) hopes to jointly develop a machine learning model that can predict diseases, but does not want to share sensitive medical data they hold. Using the federated learning system of the present invention, each medical institution uses its own data training model locally, encrypts the model using the improved distributed ElGamal encryption system (IDEC) technology, and then uploads it to the aggregation server. The aggregation server aggregates the encrypted model, and then each participant downloads the aggregated model, and obtains the decryption key by interacting with the key server using a secret sharing mechanism to decrypt and update the local model. This embodiment allows each medical institution to jointly contribute to the development of a disease prediction model without leaking patient data, significantly improving the accuracy and reliability of the model while ensuring data privacy and security.
[0069] Example 2: Federated Learning of Credit Scoring Models Across Banks In the development of cross-bank credit scoring models, multiple banks (participants) hope to improve the performance of credit scoring models through cooperation, but at the same time, each bank is extremely sensitive to the financial information of its customers and is unwilling to share data directly. Using the federated learning system of the present invention, each bank uses its own customer data to train the credit scoring model locally, and then uploads the model to the aggregation server after encrypting it with IDEC. The server is responsible for aggregating the encrypted model, and each participant then downloads the aggregated model and decrypts it through the secret sharing mechanism provided by the key server to update the local model. This embodiment enables multiple banks to share model learning results while ensuring customer privacy, thereby improving the universal applicability and accuracy of the credit scoring model.
[0070] These two embodiments demonstrate how the federated learning system of the present invention protects data privacy and security in practical applications, while improving the efficiency and effectiveness of cross-domain cooperation, and promoting technological advancement and application expansion.
[0071] In the process of crop cultivation across scientific research institutions, this method is used to predict the growth of crops. The crop growth information of the cultivation base is used as a data set, and the BP neural network based on the frost ice optimization algorithm (RIME) is used for federated learning to predict crop growth. During the learning process, this method is used to protect the privacy data of each participant.
[0072] Figure 8This is the test result of the original method. From this picture, we can see that the final prediction effect of the federated learning system using the original method is quite close to the true value, and has reached an accuracy that can be applied in real scenarios.
[0073] Fig. 9 This is the test effect diagram of this method. Fig.10 This is the test result diagram of all samples of this method. Fig. 9 and Fig.10 It can be seen that the final prediction effect of the federated learning system using this method is quite close to the true value, and has reached an accuracy that can be applied in real scenarios.
[0074] Depend on Figures 8-10 It can be seen that when this method is applied, the final training effect is not affected, and the accuracy of the final training result is even improved. However, using the previous attack method, it is impossible to obtain the private data of the participants from this system. This method improves the security performance of the federated learning system.
[0075] Fig.11 This is the training effect diagram under the conditions of this method. Fig.11 It can be seen that under the conditions of this method, the results of the training process are very close to the actual values and can meet the usage standards in actual scenarios.
[0076] Fig.12 This is the test effect diagram under the conditions of this method. Fig.12 It can be seen that under the conditions of this method, the test results are basically consistent with the actual values, with only small errors in very few cases, which has reached the usage standards in actual scenarios.
[0077] Depend on Fig.11 and Fig.12 It can be seen that the prediction effect of the trained model is good and basically consistent with the real data. In addition, during the learning process, disconnected users do not affect the correctness of the federated learning process.
[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for protecting model privacy in a federated learning system, characterized in that: The method comprises the following steps: a. Participants use private data to train models locally; b. The participants use the improved distributed ElGamal cryptographic system IDEC algorithm to encrypt the trained model; c. The participants upload the encrypted model to the aggregation server; d. The aggregation server aggregates the received encryption model; e. The participants download the aggregated model from the aggregation server and obtain the decryption key by interacting with the key server to achieve secret sharing; f. The participants use the decryption key to decrypt the aggregated model and update the local model for the next round of training.
2. The method for protecting model privacy in a federated learning system according to claim 1, wherein: The participants use the improved distributed ElGamal encryption system IDEC algorithm to encrypt the model, including the following steps: a. The key server generates the encryption key for the participants; b. The participants use the generated key to encrypt the locally trained model using the distributed ElGamal encryption system IDEC algorithm; c. The encrypted model exists in ciphertext form to ensure security during transmission and storage.
3. The method for protecting model privacy data in a federated learning system according to claim 1, characterized in that: The participants upload the encrypted model to the aggregation server to aggregate the received encrypted model, including the following steps: a. The aggregation server receives the encrypted models from each participant; b. The aggregation server performs aggregation calculation on the received encryption model to obtain an aggregated encryption model; c. The aggregation server returns the aggregated encrypted model to the participants to ensure that the privacy data of any participant is not leaked during the entire aggregation process.
4. The method for protecting model privacy data in a federated learning system according to claim 1, characterized in that: The participants use the decryption key to decrypt the aggregated model and update the local model for the next round of training, including the following steps: a. The participants download the aggregated encrypted model from the aggregation server; b. The participants interact with the key server and obtain the decryption key through secret sharing; c. The participants use the decryption key to decrypt the aggregated encrypted model to obtain the decrypted model; d. The participants update the local model based on the decrypted model for the next round of model training and encrypted upload.
5. The method for protecting model privacy data in a federated learning system according to claim 1, characterized in that: The participants adopt the improved distributed ElGamal cryptographic system IDEC algorithm as follows: 1) Key generation: Given a modulus that is prime The multiplicative cyclic group of and one of its generators , each participant From the multiplicative cyclic group Randomly select a private key , and then the public key Sent to the key server, which then calculates the joint public key of all participants , where n is the number of participants; 2) Encryption: In the multiplication cycle group The ciphertext is determined in and joint public key After that, all participants share the same random number , is from The joint public key is then used The encryption process is as follows: , where C is the encrypted ciphertext, Represents a part of the ciphertext, Represents ciphertext data, It is an encrypted process value. It is part of the encryption result and one of the process values of the encryption operation; 3) Decryption: After receiving the secret sharing share from the key server, each participant uses the secret sharing to reconstruct the joint private key , each participant Get Message By the following equation: , ciphertext It can be quickly obtained by table lookup.
6. A federated learning system implementing the method according to any one of claims 1 to 5, characterized in that: The system includes: participants, who are responsible for using private data to train models locally and uploading the trained models after encryption to the aggregation server; the aggregation server, which is responsible for receiving encrypted models uploaded by each participant and aggregating these models; the key server, which is responsible for generating and distributing keys to participants so that the participants can decrypt the downloaded aggregated models; the model encryption adopts the distributed ElGamal encryption system IDEC algorithm to ensure the security and privacy protection of model information in the entire learning system.
7. The federated learning system according to claim 6, wherein: The model encryption and decryption method in the federated learning system adopts the improved distributed ElGamal encryption system IDEC algorithm. After the participants encrypt the local model, they upload it to the aggregation server for aggregation. After the aggregation is completed, the participants download the aggregated model from the aggregation server and obtain the decryption key through secret sharing interaction with the key server. After decrypting the aggregated model, the local model is updated for the next round of training.
8. A key management method for a federated learning system, characterized in that: This method uses a key server to generate and distribute keys to participants, and uses a secret sharing mechanism to ensure the secure distribution of decryption keys. Participants update local models by decrypting the obtained aggregated model. This method ensures the security of the participants' privacy information while supporting the model information to exist in a confidential state throughout the learning system.
9. A privacy protection method for a federated learning system using an improved distributed ElGamal encryption system IDEC algorithm combined with a secret sharing mechanism, characterized in that: This method uses the distributed ElGamal encryption system IDEC algorithm to encrypt the model at the participating parties, and uses the secret sharing mechanism to interact with the key server to obtain the decryption key, so as to improve the privacy protection performance of the federated learning system.
