Lightweight security decentration privacy protection method for medical system
By constructing a decentralized secure communication network and employing gradient-independent optimization strategies, the problems of privacy leakage and gradient computation difficulties in medical IoT are solved, achieving efficient privacy protection and model co-optimization, and improving the security and performance of medical AI systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2026-01-09
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies in the medical Internet of Things (IoT) face challenges in terms of privacy leakage, data silos, and limited computing resources. These challenges result in difficulties in security compliance and real-time performance when training high-precision diagnostic models across institutions, particularly in gradient calculation on terminal devices, making it difficult to achieve efficient decentralized collaborative optimization.
A decentralized secure communication network is constructed using a time-varying undirected graph model and a double random weight matrix. By combining EC-ElGamal encryption and ECDSA signature, the gradient direction is estimated through a single-point residual feedback mechanism, thereby achieving homomorphic encryption and privacy protection of model parameters and enabling distributed optimization iteration.
Without disclosing the original parameters, it significantly reduces the risk of privacy leakage, enhances the model's anti-interference ability and convergence accuracy, reduces communication and computing overhead, and meets the high availability and low latency requirements of medical AI systems.
Smart Images
Figure CN121902207A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of privacy protection, specifically to a lightweight, secure, decentralized privacy protection method for medical systems. Background Technology
[0002] Currently, deep learning-driven medical IoT has become a core support for smart healthcare, and the collaborative training of high-precision diagnostic models using cross-institutional medical data has become a global consensus. However, due to the high sensitivity of medical data, and the strict privacy protection of information such as images and medical records, large-scale collaborative modeling across hospitals and devices presents significant challenges to system security, compliance, and real-time performance due to inherent privacy risks, data silos, and the timeliness requirements of online streaming. A single medical institution has limited data volume and significant feature distribution bias, making it difficult to independently train a diagnostic model with high generalization ability. Therefore, adopting a decentralized federated architecture formed by geographically dispersed multiple medical institutions has become an emerging technology that can improve the accuracy, robustness, and fairness of models without sharing raw data. Mainstream federated learning schemes rely on a central server to aggregate gradient information from all participants for global updates. However, this centralized approach places extremely high demands on the computation and bandwidth of the central node, and the entire training process is interrupted if the central node fails or is attacked, fundamentally conflicting with the distributed requirements of high availability, high privacy, and low latency in medical scenarios. Furthermore, existing methods often assume that participants can obtain complete gradient or loss information in real time. However, in actual IoT terminals, computing resources are limited, and only single-point loss values can be observed, resulting in a "slot machine"-like partial information feedback, which further exacerbates the scheduling difficulty. Summary of the Invention
[0003] To address the aforementioned shortcomings in existing technologies, this invention provides a lightweight, secure, decentralized privacy protection method for medical systems.
[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A lightweight, secure, and decentralized privacy protection method for healthcare systems includes the following steps: S1. Construct a decentralized secure communication network, using a time-varying undirected graph model and a double random weight matrix to achieve local cooperation and information diffusion among clients, providing a network foundation for secure parameter exchange; S2, based on network communication of S1, performs EC-ElGamal homomorphic encryption and ECDSA digital signature on model parameters; S3. Under the encryption protection of S2, the gradient direction is estimated through a single-point residual feedback mechanism, and the loss function value is queried only to overcome the difficulty of gradient calculation. S4 integrates the gradient direction estimated by S3 and the safe aggregation mechanism of S2 to perform model update aggregation and convergence checks, completing the distributed optimization iteration with privacy protection.
[0005] Furthermore, step S1 specifically includes the following steps: S11. Construct a time-varying undirected graph model ,in , representing the set of all clients, express The communication link at any given moment; S12, Define the weight matrix It is a doubly random matrix and satisfies the connectivity assumption, i.e., there exist integers. , making each The graph is connected within a series of consecutive time steps; S13, Weighting coefficients Decomposed into the product of two independent factors generated independently by the client. ,in, By client For the neighbors Generate and keep confidential, By client For the neighbors Independently generated and kept confidential; S14. Calculate the self-weight To ensure the randomness of the matrix rows.
[0006] Furthermore, step S2 specifically includes the following steps: S21. Calculation model parameter update information In the formula, For local model parameters, For the neighbor model parameters, convert the model parameter update information into an integer vector. For vectors Each element in Calculate its integer representation The calculation method is as follows: ,in, For quantification factors; S22. Encrypt M, calculate and send it to the neighbor. j The ciphertext:
[0007]
[0008]
[0009] In the formula, r It is a random number.G The base point of the elliptic curve For the neighbors j The public key for encryption; S23, Send to neighbor j ciphertext Perform a signature and generate a signature:
[0010] Among them, H is the private key for signing, and H is the hash function; S24, tuple Send to neighbor client Client Received from neighbor After receiving the information, first use The signature public key is used to verify the validity of the signature. If the verification fails, the message is discarded; if successful, the message is decrypted using the signature private key and converted into a real number vector.
[0011] Furthermore, S3 includes the following steps: S31, Client From a unit sphere Uniformly sample a unit vector ; S32, at the current decision point Nearby query loss function value:
[0012] in, For smoothing parameters; S33. Utilize the loss function value queried at the current time. The loss function value queried at the previous time step Construct the residuals and perform gradient estimation based on the constructed residuals, specifically as follows:
[0013] in, For parameter dimensions.
[0014] Furthermore, step S4 specifically includes the following steps: S41, Receive all neighbors The encrypted message is sent, verified, and decrypted to obtain the neighbor's model update information. :
[0015] In the formula, For local model parameters, For neighbor model parameters, Weighting coefficients Independent factors derived from decomposition; S42. Calculate the privacy-preserving weighted average term using the weighting factor of its own will and the model update information of the decrypted point neighbors. S43. Based on the local gradient estimate Perform model updates and project the updated model parameters onto the constraint set. ; S44. Return to step S2 and iterate until the model converges or the preset maximum number of iterations is reached.
[0016] Furthermore, the specific method for calculating the privacy-protected weighted average term in S42 is as follows:
[0017] In the formula, This is the weighted average term.
[0018] Furthermore, the specific method of model updating in S43 is as follows:
[0019] In the formula, For the updated local model parameters, For the weighted average term, For learning rate, This is the estimated gradient.
[0020] Furthermore, the projection operation in S43 is represented as follows:
[0021] In the formula, For the updated local model parameters, Let it be its projection onto the constraint set. Let be the projection operator, where For the search variables in the constraint set.
[0022] The present invention has the following beneficial effects: This method defends against both honest but curious participants and external eavesdropping / tampering attacks in a partially information feedback environment. It uses elliptic curve cryptography to homomorphically encrypt and digitally sign model parameters, and utilizes single-point residual feedback to estimate the zero-order gradient. This enables decentralized collaborative optimization without revealing any original parameters, significantly reducing the impact of privacy leakage risks on the system, enhancing the anti-interference ability and convergence accuracy of medical AI models, and reducing overall communication-computation costs. Attached Figure Description
[0023] Figure 1This is a flowchart of the decentralized federated learning fusion EC-ElGamal encryption algorithm of the present invention.
[0024] Figure 2 This is a schematic diagram of the privacy protection model update process according to an embodiment of the present invention. Detailed Implementation
[0025] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0026] A lightweight, secure, decentralized privacy protection method for healthcare systems, such as Figure 1 As shown, it includes the following steps: 1) Construct a decentralized secure communication network: Employ a time-varying undirected graph model and a double random weight matrix to achieve local cooperation and information diffusion among clients, avoid dependence on a central server, and improve system robustness.
[0027] 2) Integrating lightweight cryptographic mechanisms: EC-ElGamal encryption is used to achieve homomorphic protection of model parameters, and ECDSA signature is combined to verify the source and integrity of data, effectively resisting internal privacy inference and external eavesdropping and tampering attacks.
[0028] 3) Design gradient-independent optimization strategy: In partial information feedback scenarios, the gradient direction is estimated through the single-point residual feedback (OPRF) mechanism. Only the loss function value needs to be queried, which overcomes the problem of gradient calculation difficulty of deep models on terminal devices.
[0029] 4) Achieving a balance between privacy and efficiency: While ensuring the privacy and security of medical data, the algorithm's computational and communication overhead is significantly lower than that of traditional homomorphic encryption schemes, and its model accuracy and practical feasibility have been verified through medical image datasets.
[0030] Specifically, a lightweight, secure, decentralized, online federated learning method for healthcare systems is proposed, and its implementation process is as follows: The first step is to build a decentralized communication topology and security foundation, starting with constructing a time-varying undirected communication network graph. ,in , representing the set of all clients, express The communication link at any given moment. Each client Only with its neighbor set To enable communication, this topology must satisfy the following assumptions: 1. Randomness: The associated weight matrix It is double random.
[0031] 2. Connectivity: There exists an integer , making each The graph is connected within a series of consecutive event steps.
[0032] This step is fundamental to achieving decentralized collaboration and privacy protection, and its core lies in the decomposition design of the weight matrix. At each moment... The system dynamically generates a weight matrix based on network connectivity. To achieve fully distributed and privacy-preserving aggregation, the weight coefficients are... Decompose into the product of two independent factors: .in By client For the neighbors Independently generated and kept confidential. By client For the neighbors Independently generated and kept confidential. The two factors are uniformly and randomly selected from the interval to ensure that the assumptions of the weight matrix are met. (Client) Self weight Determined by the following formula to ensure Randomness of rows: The second step integrates lightweight cryptographic mechanisms: EC-ElGamal encryption is used to achieve homomorphic protection of model parameters, and ECDSA signatures are combined to verify the source and integrity of data, effectively resisting internal privacy inference and external eavesdropping and tampering attacks.
[0033] Specifically, the first step is to perform information calculation and integerization: the calculation needs to be sent to the neighbor. The model update information. This is typically the difference between the local model and its neighbors' models. To facilitate ECC encryption, it is necessary to first... 3D real vector Convert to integer vector For vectors Each element in Calculate its integer representation: in As a quantization factor, it converts floating-point numbers to integers with controllable precision loss.
[0034] Secondly, EC-ElGamal encryption is applied to the data: the integerized information vector... Encryption is performed. Use neighbors. public key of encryption Random selection Calculate the ciphertext:
[0035]
[0036] Here For operations on a point group, the actual implementation is to... After mapping the points onto the curve, scalar multiplication and addition of the points are performed to obtain the result sent to the curve. ciphertext .
[0037] To further prevent messages from being tampered with or forged, ESDSA signatures are used on the client side. Use your own signing private key , for ciphertext Perform signing, generate signature
[0038] Further tuples Send to neighbor client Client Received from neighbor After receiving the information, first use The signature public key is used to verify the validity of the signature. If verification fails, the message is discarded; if successful, the message is decrypted using the user's private encryption key.
[0039] Then Converting back to a real number vector ensures the confidentiality, integrity, and authentication of communications.
[0040] The third step is to design a gradient-independent optimization strategy: In scenarios with partial information feedback, the gradient direction is estimated through the single-point residual feedback (OPRF) mechanism, which only requires querying the loss function value, thus overcoming the problem of gradient calculation difficulties for deep models on terminal devices.
[0041] Specifically, for clients unable to calculate precise gradients (Bandit feedback scenarios), follow these steps: First, at the time step Client From the unit sphere Uniformly sample a unit vector Secondly, at the current decision point Nearby query loss function value:
[0042] Using the current query value Compared to the query value at the previous time step:
[0043] Construct the residual.
[0044] Gradient estimation is performed using the single-point residual feedback (OPRF) mechanism:
[0045] This estimate is a smoothed version of the loss function. gradient The unbiased estimate, i.e.:
[0046] Furthermore, its variance is bounded, making it an effective alternative when the true gradient cannot be obtained.
[0047] The fourth step achieves a balance between privacy and efficiency. While ensuring the privacy and security of medical data, the algorithm's computational and communication overhead is significantly lower than that of traditional homomorphic encryption schemes. The model's accuracy and practical feasibility have been verified through a medical image dataset.
[0048] Each client Update its local model using the following steps: Figure 2 As shown: First, receive and decrypt: Receive all neighbors. The received encrypted message is verified and decrypted as described in step two to obtain the neighbor's model update information. .
[0049] Further local aggregation is performed: utilizing the known weighting factors. and the neighbor information obtained by decryption Calculate the privacy-protected weighted average:
[0050] Client have no idea and However, the consensus process was completed indirectly through encrypted communication and local computation, without leaking any node's private information.
[0051] Further model updates: incorporating local gradient estimates (or precise gradient) (If available) Update:
[0052] This update formula is equivalent to standard distributed gradient descent, but all communication links are cryptographically protected.
[0053] Further projection operations are performed: the updated parameters are projected onto the constraint set. Above, ensure the feasibility of the solution:
[0054] Finally, perform iterative loops: repeat steps two through four until the model converges and satisfies the formula:
[0055] Or reach the preset maximum number of iterations .
[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0058] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0059] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
[0060] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. A lightweight, secure, decentralized privacy protection method for medical systems, characterized in that, Includes the following steps: S1. Construct a decentralized secure communication network, using a time-varying undirected graph model and a double random weight matrix to achieve local cooperation and information diffusion among clients, providing a network foundation for secure parameter exchange; S2, based on network communication of S1, performs EC-ElGamal homomorphic encryption and ECDSA digital signature on model parameters; S3. Under the encryption protection of S2, the gradient direction is estimated through a single-point residual feedback mechanism, and the loss function value is queried only to overcome the difficulty of gradient calculation. S4 integrates the gradient direction estimated by S3 and the safe aggregation mechanism of S2 to perform model update aggregation and convergence checks, completing the distributed optimization iteration with privacy protection.
2. The lightweight, secure, decentralized privacy protection method for medical systems according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Construct a time-varying undirected graphical model ,in Represents the set of all clients. express The communication link at any given moment; S12, Define the weight matrix It is a doubly random matrix and satisfies the connectivity assumption, i.e., there exist integers. , making each The graph is connected within a series of consecutive time steps; S13, Weighting coefficients Decomposed into the product of two independent factors generated independently by the client. ,in, By client For the neighbors Generate and keep secret. By client For the neighbors Independently generated and kept confidential; S14. Calculate the self-weight To ensure the row randomness of the matrix, in the formula, To calculate the total sum.
3. The lightweight, secure, decentralized privacy protection method for medical systems according to claim 1, characterized in that, S2 specifically includes the following steps: S21. Calculation model parameter update information In the formula, For local model parameters, For the neighbor model parameters, convert the model parameter update information into an integer vector. For vectors Each element in Calculate its integer representation The calculation method is as follows: ,in, For quantification factors; S22. Encrypt M, calculate and send it to the neighbor. j The ciphertext: In the formula, r It is a random number. G The base point of the elliptic curve For the neighbors j The public key for encryption; S23, Send to neighbor j ciphertext Perform a signature and generate a signature: Among them, H is the private key for signing, and H is the hash function; S24, tuple Send to neighbor client Client Received from neighbor After receiving the information, first use The signature public key is used to verify the validity of the signature. If the verification fails, the message is discarded; if successful, the message is decrypted using the signature private key and converted into a real number vector.
4. The lightweight, secure, decentralized privacy protection method for medical systems according to claim 1, characterized in that, S3 includes the following steps: S31, Client From the unit sphere Uniformly sample a unit vector ; S32, at the current decision point Nearby query loss function value: in, For smoothing parameters; S33. Utilize the loss function value queried at the current time. The loss function value queried at the previous time step Construct the residuals and perform gradient estimation based on the constructed residuals, specifically as follows: in, For parameter dimensions.
5. The lightweight, secure, decentralized privacy protection method for medical systems according to claim 1, characterized in that, S4 specifically includes the following steps: S41, Receive all neighbors The encrypted message is sent, verified, and decrypted to obtain the neighbor's model update information. : In the formula, For local model parameters, For neighbor model parameters, Weighting coefficients Independent factors derived from decomposition; S42. Calculate the privacy-preserving weighted average term using the weighting factor of its own will and the model update information of the decrypted point neighbors. S43. Based on the local gradient estimate Perform model updates and project the updated model parameters onto the constraint set. ; S44. Return to step S2 and iterate until the model converges or the preset maximum number of iterations is reached.
6. The lightweight, secure, decentralized privacy protection method for medical systems according to claim 5, characterized in that, The specific method for calculating the privacy-protected weighted average in S42 is as follows: In the formula, This is the weighted average term.
7. The lightweight, secure, decentralized privacy protection method for medical systems according to claim 5, characterized in that, The specific method for model updating in S43 is as follows: In the formula, For the updated local model parameters, For the weighted average term, For learning rate, This is the estimated gradient.
8. The lightweight, secure, decentralized privacy protection method for medical systems according to claim 5, characterized in that, The projection operation in S43 is represented as follows: In the formula, For the updated local model parameters, Let its projection onto the constraint set, Let be the projection operator, where For the search variables in the constraint set.