A privacy preserving collaborative decision making system and method

Abstract

The application relates to the technical field of data security, and discloses a privacy protection collaborative decision system and method. The system comprises: a plurality of proposal generation and encryption modules, which are used for generating a probabilistic action proposal vector representing a decision tendency and an encrypted proposal vector based on local private data and received feedback signals; an encryption aggregation module, which performs weighted aggregation in a ciphertext domain on a plurality of encrypted proposal vectors to generate an encrypted aggregated decision potential vector; a decision analysis and feedback module, which judges whether a decision consensus is reached according to a consensus degree check; and a decision output module, which outputs a privacy protection collaborative decision when the decision reaches a consensus. The application performs homomorphic encryption on action proposals and completes aggregation and analysis in a ciphertext domain, thereby not only protecting local data of each participant and intermediate intentions in a multi-round decision process, but also avoiding behavior strategy inference risks caused by information leakage.

Description

Technical Field

[0001] This invention relates to the field of data security technology, specifically to a privacy-protecting collaborative decision-making system and method. Background Technology

[0002] With the rapid development of technologies such as artificial intelligence, the Internet of Things, and blockchain, in numerous distributed application scenarios, multiple intelligent agents or participants need to make collaborative decisions based on their respective local private data. These scenarios, such as cross-institutional financial risk assessment, model selection in federated learning, or task collaboration in multi-robot systems, fundamentally require achieving a globally optimal or universally agreed-upon decision outcome without disclosing sensitive information from any party.

[0003] To resolve the conflict between collaboration and privacy, existing technological solutions typically rely on a trusted centralized server that collects and processes the raw data from all participants. However, this architecture places all responsibility for privacy protection on a single entity, creating a significant security bottleneck and trust risk. If the centralized server is attacked or acts maliciously, the sensitive data of all participants faces the risk of complete leakage.

[0004] To overcome the drawbacks of centralized trust, secure multi-party computation or cryptographic aggregation schemes are introduced. While these methods improve the confidentiality of data transmission and processing to some extent, they still expose a series of deep-seated problems in complex dynamic decision-making scenarios. Existing methods often lack effective guidance mechanisms when dealing with disagreements among participants. When a consensus cannot be reached in the short term, the system usually gets bogged down in multiple rounds of inefficient, blind iterations, or directly declares the decision a failure. The root cause is the failure to deeply analyze the internal structure of the disagreements, thus failing to provide targeted adjustment strategies.

[0005] Furthermore, existing technologies typically employ homogenized methods such as simple averaging or static weighting when evaluating the contributions of each participant. This mechanism ignores the differences in expertise, data quality, or historical performance among different participants, allowing low-quality or even malicious proposals to influence the final outcome as much as high-quality proposals. This not only reduces the accuracy of decision-making but also lowers the overall robustness of the system. Moreover, even the publication of processed intermediate aggregated results during multiple rounds of interaction carries the risk of information leakage. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a privacy-preserving collaborative decision-making system and method, which solves the problems of privacy leakage in the decision-making process, inefficient consensus guidance, and insufficient accuracy of decision results that exist simultaneously in existing technologies.

[0007] To achieve the above objectives, the first aspect of the present invention provides a privacy-preserving collaborative decision-making system, which includes: multiple proposal generation and encryption modules, an encryption aggregation module, a decision analysis and feedback module, and a decision output module.

[0008] Each of the multiple proposal generation and encryption modules is deployed locally on one of the participating parties. In the decision-making process... In this round, the proposal generation and encryption module is based on its local private data. And the feedback signal received in the previous round To execute its internal decision-making strategies Generate probabilistic action proposal vectors The calculation formula is:

[0009] ;

[0010] In the formula, As a participant In the Probabilistic action proposal vectors generated in each round; As a participant The decision strategy function; As a participant Local private data or observations held; For the system in the first The feedback signal broadcast at the end of the cycle; The total number of discrete options in the decision action space; Indicates the participating parties In the Round proposal selection The vector contains the probability values ​​of each decision option. Each element of this vector satisfies the probability distribution constraints, i.e., and... , .

[0011] The encryption aggregation module is used to perform computations within the ciphertext domain. It receives encrypted proposal vectors from all proposal generation and encryption modules, and obtains the reputation weight assigned to each module. This module utilizes the homomorphic addition and homomorphic scalar multiplication properties of homomorphic encryption schemes to perform weighted aggregation operations, generating an encrypted aggregation decision potential vector. The calculation formula is as follows:

[0012] ;

[0013] This vector, in its encrypted state, incorporates the decision-making tendencies and reputational influences of all participants.

[0014] The decision analysis and feedback module is used to process the encrypted aggregated decision potential vector. First, without decrypting the data, this module performs a consensus check to determine whether the concentration of opinions among all participants has reached a preset threshold.

[0015] If the consensus check results in no consensus being reached, the module initiates a topology analysis procedure. This procedure, also executed within the encrypted domain, calculates the statistical characteristics of the encrypted aggregated decision potential vector, identifying the decision distribution pattern it represents as one of a preset category, such as a unimodal, multimodal, or flat distribution. Based on this analysis, the module generates a structured feedback message without specific numerical values, such as a classification label identifying the distribution pattern, and uses this as the feedback signal for this round. Send to all proposal generation and encryption modules.

[0016] In one technical solution, the decision-making strategy of the proposal generation and encryption module is a reinforcement learning strategy. After receiving the structured feedback information, the module adjusts its decision-making strategy by incorporating the structured feedback information as part of the environment state of the reinforcement learning strategy and calculating the reward function of the reinforcement learning strategy based on the information, thereby generating probabilistic action proposal vectors with different tendencies in the next round.

[0017] The decision output module is triggered when the decision analysis and feedback module determines that a consensus has been reached. This module first determines a final decision. In one specific implementation, the decision analysis and feedback module decrypts the final encrypted aggregated decision potential vector to obtain the plaintext aggregated decision potential vector, and selects the decision option corresponding to the maximum value as the final decision. Subsequently, the decision output module updates the corresponding reputation weight based on the consistency between the probabilistic action proposal vector submitted by each proposal generation and encryption module and the final decision. The updated reputation weight will be used in subsequent decision-making processes.

[0018] A second aspect of the present invention provides a privacy-preserving collaborative decision-making method, the method comprising the following steps:

[0019] Step 1: Generating and Encrypting Proposals: Based on local private data and received feedback signals, the computing devices of multiple participating parties generate probabilistic action proposal vectors representing decision-making tendencies, and encrypt the probabilistic action proposal vectors to generate encrypted proposal vectors.

[0020] Step 2, Encrypted Aggregation: A central processing unit or a distributed computing network performs weighted aggregation of the received multiple encrypted proposal vectors in the ciphertext domain based on the reputation weight set for each participant, generating an encrypted aggregation decision potential vector.

[0021] Step 3, Consensus Check: Perform a consensus check based on the encrypted aggregated decision potential vector to determine whether a decision consensus has been reached;

[0022] Step 4: Output Decision: When the consensus check results in reaching a consensus, a final privacy-protection collaborative decision is determined and output.

[0023] In one embodiment, after step 3, if the consensus check result is that no decision consensus has been reached, the method further includes: performing topological structure analysis on the encrypted aggregate decision potential vector in the ciphertext domain to generate structured feedback information that does not contain specific values; and sending the structured feedback information as the feedback signal to the multiple participants for their use in step 1 of the next round of decision-making, thereby gradually reaching consensus through iterative adjustments.

[0024] This invention provides a privacy-preserving collaborative decision-making system and method. It has the following beneficial effects:

[0025] 1. This invention uses homomorphic encryption to encrypt probabilistic action proposal vectors; furthermore, subsequent weighted aggregation, consensus checking, and topology analysis are all performed within the ciphertext domain by the encryption aggregation module and the decision analysis and feedback module. Therefore, it not only protects the local private data of each participant, but also protects the intermediate decision intentions in the multi-round dynamic game process, avoiding the risk of behavioral strategy inference due to the leakage of interactive information.

[0026] 2. By performing topological analysis, this invention can generate structured feedback information that represents the current global decision distribution when no consensus is reached. This feedback information provides non-numerical directional guidance for each proposal generation and encryption module to adjust its internal decision-making strategy, avoiding disordered random exploration, thereby reducing the number of iterations required to reach a consensus and accelerating the convergence of the decision-making process.

[0027] 3. This invention introduces reputation weights, and the decision output module updates the final decision based on the consistency between the proposals of each participant and the final decision after the final decision is determined. This positively enhances the influence of participants who consistently provide high-quality proposals that are consistent with the direction of collective consensus in subsequent decisions, while weakening the influence of participants who exhibit random or malicious behavior. This improves the fairness and credibility of the final decision result. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0029] Figure 2 This is a schematic diagram of the method flow of the present invention.

[0030] The module comprises: 100, Proposal Generation and Encryption Module; 110, Decision Strategy Unit; 120, Encryption Unit; 130, Communication Interface; 200, Encryption Aggregation Module; 210, Data Receiving Unit; 220, Ciphertext Calculation Unit; 230, Data Sending Unit; 300, Decision Analysis and Feedback Module; 310, Data Interface Unit; 320, Consensus Check Unit; 330, Topology Analysis Unit; 340, Feedback Generation Unit; 400, Decision Output Module; 410, Data Receiving Unit; 420, Decryption Unit; 430, Decision Determination Unit; and 440, Weight Update Unit. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] To better understand the present invention, the above content will be described in detail below with reference to specific embodiments.

[0033] Please see the appendix Figure 1 , Figure 1 This is a schematic diagram of the system architecture of a privacy-preserving collaborative decision-making system according to an embodiment of the present invention; the privacy-preserving collaborative decision-making system of artificial intelligence includes:

[0034] Multiple proposal generation and encryption modules 100 output encrypted proposal vectors, which are then sent to an encryption aggregation module. Simultaneously, this module receives feedback signals from a decision analysis and feedback module. The purpose of this connection is to securely submit the local decision intentions of each participant to the system in encrypted form and to receive global feedback for policy adjustments.

[0035] The encrypted aggregation module 200 receives data from all proposal generation and encryption modules and sends the encrypted aggregated decision potential vector generated after aggregation to the decision analysis and feedback module. The purpose of this connection is to merge dispersed, independent decision proposals into a single encrypted data structure representing the global decision state without decryption.

[0036] The decision analysis and feedback module 300 receives data from the encrypted aggregation module. This module has an internal information processing loop: if no consensus is reached, it generates structured feedback information and sends it back to all proposal generation and encryption modules; if a consensus is reached, it sends the final encrypted aggregated decision potential vector to the decision output module. The purpose of this connection is to analyze the encrypted global state and, based on the analysis results, control the direction of the decision-making process, either guiding a new round of iteration or terminating the iteration and initiating decision output.

[0037] The decision output module 400 receives data from the decision analysis and feedback module and outputs the final decision result after processing. Furthermore, this module feeds back updated reputation weight information to the encrypted aggregation module for use in subsequent decision-making tasks. The purpose of this connection is to complete the final analysis of the decision after consensus is confirmed, and to adjust the reputation evaluations of each participant based on the results of this round, thereby achieving dynamic self-adaptation of the system.

[0038] In one specific embodiment, the multiple proposal generation and encryption modules 100 may include a decision strategy unit 110, an encryption unit 120, and a communication interface 130.

[0039] Specifically, the decision-making strategy unit 110 acquires the local private data of the participants and receives feedback signals from the previous round of decision-making via the communication interface 130, and generates probabilistic action proposal vectors based on these signals.

[0040] In one specific embodiment, the system predefines a subset of... Decision action space of discrete options In the decision-making process Wheel, for the participating parties Its decision strategy unit 110 executes the internal decision strategy function. This function uses local private data Feedback signals from the previous round As input, the output is a probabilistic action proposal vector. The specific form of this vector is as follows:

[0041] ;

[0042] In the formula, As a participant In the Probabilistic action proposal vectors generated in each round; As a participant The decision strategy function; As a participant Local private data or observations held; For the system in the first The feedback signal broadcast at the end of the cycle; The total number of discrete options in the decision action space; Indicates the participating parties In the Round proposal selection The vector contains the probability values ​​of each decision option. Each element of this vector satisfies the probability distribution constraints, i.e., and... , .

[0043] Decision strategy function This can be implemented using various models; in this embodiment, it can be implemented as a reinforcement learning model. This model is trained to adapt to the input state (including...) and Output the optimal action probability distribution.

[0044] The encryption unit 120 encrypts the probabilistic action proposal vector generated by the decision strategy unit 110 element by element.

[0045] In one specific embodiment, the encryption unit 120 employs an additive homomorphic encryption scheme, such as the Paillier cryptosystem. The system pre-generates a public-private key pair. and the public key The public key is distributed to all proposal generation and encryption modules 100. Encryption unit 120 uses this public key. For probabilistic action proposal vectors Each probability value in Perform encryption operations to generate encrypted proposal vectors. The encryption vector takes the following form:

[0046] ;

[0047] in, Indicates the use of a public key The homomorphic encryption function is executed. Because homomorphic encryption is used, specific algebraic operations can be performed on these encrypted values ​​without decryption.

[0048] Therefore, communication interface 130 sends the encrypted proposal vector generated by encryption unit 120 to encryption aggregation module. Simultaneously, communication interface 130 is also used to receive feedback signals sent by decision analysis and feedback module at the end of each round of decision-making, for use in the next round of decision-making.

[0049] In one specific embodiment, the encryption aggregation module 200 may include a data receiving unit 210, a ciphertext calculation unit 220, and a data sending unit 230.

[0050] Data receiving unit 210 from Each of the 100 proposal generation and encryption modules receives its own generated encrypted proposal vector. It also obtains the set of reputation weights determined by all participants at the end of the previous round from the decision output module (or a separate weight management unit). .

[0051] The ciphertext computation unit 220 performs a weighted aggregation operation within the ciphertext domain on the received data to generate an encrypted aggregation decision potential vector. This operation utilizes the homomorphic addition and homomorphic scalar multiplication properties of homomorphic encryption schemes.

[0052] In one specific embodiment, the operation first processes each encrypted proposal vector. Perform homomorphic scalar multiplication. Specifically, use the participating parties. Corresponding reputation weight As scalars, each component of the encrypted vector is operated on. Then, homomorphic addition is performed on all encrypted vectors after scalar multiplication, aggregating them into a single encrypted vector. The mathematical expression for this process is as follows:

[0053] ;

[0054] In the formula, In the first The encrypted aggregate decision potential vector generated in the round; The total number of participants; As a participant In the The reputation weight determined at the end of the round; As a participant In the The encrypted proposal vector submitted in each round; This represents the homomorphic scalar multiplication operation.

[0055] When using additive homomorphic encryption schemes such as Paillier, the homomorphic scalar multiplication in the above formula... Homomorphic addition and sine addition correspond to specific operations in cryptography. Ultimately, the aggregated vector... In the encrypted state, this is equivalent to the weighted sum of the proposal vectors of all participants according to their reputation weights, that is:

[0056] ;

[0057] Furthermore, each component of the vector is a ciphertext, and its decrypted value corresponds to the weighted sum of the support for the decision option by all participants.

[0058] Subsequently, the data sending unit 230 transmits the encrypted aggregation decision potential vector generated by the ciphertext calculation unit 220. Send it to the decision analysis and feedback module for subsequent consensus checks and decision analysis.

[0059] Subsequently, the decision analysis and feedback module 300 receives the encrypted aggregated decision potential vector. The decision analysis and feedback module 300 includes a data interface unit 310, a consensus checking unit 320, a topology analysis unit 330, and a feedback generation unit 340.

[0060] Specifically, data interface unit 310 receives data from encryption aggregation module 200 in the decision-making stage. The encrypted aggregate decision potential vector generated by the round The vector is then transmitted to consensus checking unit 320.

[0061] Consensus checking unit 320 checks the received data. A consensus check is performed to determine whether the decision-making tendency of all participants has reached a preset level of concentration. This process is completed without decrypting the vector.

[0062] In one specific embodiment, the check is achieved by calculating an encrypted decision concentration index and performing a secure comparison with a preset encryption threshold. Let the undecrypted aggregated decision potential vector be... Then the feasible decision concentration index It can be defined as the sum of squares of the components of the vector, and its formula is:

[0063] ;

[0064] In the formula, The first vector representing the aggregation decision potential vector. One plaintext component. This plaintext vector is the encrypted aggregation decision potential vector. The decrypted vector, where the value of this component represents the value of all participants for the first... The sum of weighted support for each decision option.

[0065] Consensus Check Unit 320 Holding The encrypted value for calculating this index. This unit performs a specific secure computation protocol that relies on a homomorphic encryption scheme. For example, in a system that includes a secure multiplication protocol, this unit can compute... Without knowing any The specific value.

[0066] The consensus checking unit 320 will calculate the encrypted decision concentration index. With the preset, equally encrypted consensus threshold A security comparison is performed. This comparison process is completed by executing a security comparison protocol, the output of which is a Boolean value indicating... Whether it is valid or not, and without disclosing information during the process. and Any information.

[0067] When the consensus checking unit 320 determines that no consensus has been reached, it will trigger the topology analysis unit 330 to process the encrypted aggregation decision potential vector. Perform the analysis.

[0068] After the topology analysis unit 330 is triggered, it obtains the encrypted aggregation decision potential vector from the consensus check unit 320 or the data interface unit 310. This analysis aims to identify the decision distribution pattern represented by the vector without decrypting the data.

[0069] The topology analysis unit 330 calculates the difference between adjacent components within the encrypted domain for subsequent local extremum point determination. Let the encrypted aggregation decision potential vector be:

[0070] ;

[0071] Using the properties of homomorphic encryption, the encrypted value of the difference between adjacent components can be calculated:

[0072] ;

[0073] In the formula, where, , For the first The component and the first The plaintext difference of each component, The symbol represents a homomorphic subtraction operation.

[0074] Subsequently, the topology analysis unit 330 determines the encryption difference for each encryption using a secure comparison protocol. The relationship between the magnitude of zero and the value of zero, i.e., the judgment. It can be positive, negative, or zero, without being leaked during the process. The specific values ​​are then calculated. Based on the comparison results, the number of peak values ​​is determined.

[0075] Furthermore, to distinguish between unimodal and flat distributions, the topology analysis unit 330 calculates the encrypted value of the variance of the aggregation decision potential vector within the encrypted domain. The variance The calculation formula is as follows:

[0076] ;

[0077] In the formula, The variance of each component of the aggregate decision potential vector; Let the dimension be the vector. For the vector of the first One plaintext component. This computation process is completed by executing several secure multi-party computation protocols (including homomorphic addition, homomorphic scalar multiplication, and secure multiplication protocols), ultimately yielding... Without revealing intermediate values.

[0078] Subsequently, the topology analysis unit 330 calculates the encryption variance value. The same encrypted flatness threshold preset by the system Perform a safety comparison to determine the degree of dispersion of the decision distribution.

[0079] Finally, based on the number of peak values Based on the comparison results with variance, the topology analysis unit 330 categorizes the decision distribution pattern into one of the preset categories. For example: if If it is classified as a multimodal distribution; if If the variance is greater than the flatness threshold, it is classified as a unimodal distribution; if If the variance is not greater than the flatness threshold, it is classified as a "flat distribution". This classification result will be sent to the feedback generation unit 340.

[0080] After the topology analysis unit 330 completes the classification of the decision distribution pattern, it sends the classification result to the feedback generation unit 340. The feedback generation unit 340 receives the classification result from the topology analysis unit 330, which is an identifier, including: unimodal distribution, multimodal distribution, or flat distribution.

[0081] The feedback generation unit 340 converts the received classification results into structured feedback information that does not contain specific numerical values, i.e., the feedback signal for this round, according to a preset mapping rule. The signal is a discrete, standardized code; for example, integers are used to represent different distribution patterns.

[0082] Furthermore, the feedback generation unit 340 transmits the generated feedback signal through the data interface unit 310. The proposal generation and encryption module 100 broadcasts the proposal to all participants for their decision-making in the next round (the...). Each team adjusts its decision-making strategy in each round.

[0083] In this embodiment, the proposed generation is related to the decision-making strategy within the encryption module 100. It is implemented as a reinforcement learning strategy. This strategy utilizes the received feedback signals in the following way. :

[0084] First, the feedback signal Used as a reinforcement learning strategy in the Part of the environmental state of the wheel. Participants The state observed by the reinforcement learning policy is defined as containing its local private data. and global feedback signals The combination of .

[0085] Secondly, based on this feedback information For the participants Calculate reward value The reward function Used to evaluate its in the The contribution of each proposal submitted in a round to the system's consensus. The reward function is defined as follows:

[0086] ;

[0087] In the formula, As a participant In the The reward value obtained at the end of the round; For the first Feedback signals generated by the wheel; It is a predefined mapping function.

[0088] The mapping function Specifically, it can be set as follows:

[0089] ;

[0090] In the formula, , and It is a preset positive real constant. This reward value It was subsequently used to update the participants. Decision-making strategies For example, by updating the parameters of the Q-table or policy network, it can be made to tend to generate probabilistic action proposal vectors in subsequent rounds that can guide the system to evolve toward a "unimodal distribution" state.

[0091] In this embodiment, the decision output module 400 includes a data receiving unit 410, a decryption unit 420, a decision determination unit 430, and a weight update unit 440.

[0092] Specifically, once the decision analysis and feedback module 300 determines that a consensus has been reached, it will trigger the decision output module 400. The decision output module 400 determines the final decision through the following steps:

[0093] The data receiving unit 410 receives the consensus signal and the final encrypted aggregated decision potential vector from the decision analysis and feedback module 300. The received vector is then transmitted to the decryption unit 420.

[0094] Decryption unit 420 uses the public key held by the system for encryption. Paired private keys The aggregation decision potential vector for encryption Perform the decryption operation. This operation recovers the plaintext aggregation decision potential vector. Its form is as follows:

[0095] ;

[0096] In the formula, To use the private key The decryption function executed; For the vector of the first Each component has a value equal to the weighted sum of the corresponding components of the proposal vector of all participants, based on their reputation weights.

[0097] Subsequently, the decision-making unit 430 receives the plaintext vector output by the decryption unit 420. And based on this, a final decision is determined. This determination process uses vectors... This is achieved by finding the element with the maximum value in the set. The final decision option's index is then used. Determined by the following formula:

[0098] ;

[0099] in, The function returns an expression Index to get the maximum value .

[0100] Finally, the decision-making unit 430 will index... The corresponding decision options serve as the final output of this collaborative decision-making process. Simultaneously, the vector... and final decision index It is sent to the weight update unit 440 for subsequent reputation weight updates.

[0101] In this embodiment, after the decision-making unit 430 determines the final decision, it will trigger the weight update unit 440 to update the reputation weight. The weight update function is implemented through the following steps:

[0102] The weight update unit 440 receives the index of the final decision option from the decision determination unit 430. To assess contribution, the weight update unit 440 also needs to obtain the consensus reached by all participants on the [number]th [period / phase]. In one specific embodiment, this step involves requiring the proposal generation and encryption module 100 of each participant to submit its plaintext proposal vector for the last round after the decision result is announced. This is achieved by sending the data to the decision output module 400.

[0103] The weight update unit 440 calculates the contribution of each participant based on their support for the final decision option. In this embodiment, the participants... The contribution was quantified as its contribution in the first quarter. proposal vector of the round In the middle, assigned to the final decision options probability value .

[0104] The weight update unit 440 updates the reputation weight of each participant based on the calculated contribution. The weight update uses a smooth update function, and its update rules are as follows:

[0105] ;

[0106] In the formula, As a participant Updated, unnormalized reputation weights; As a participant Reputation weight before update; This is a preset learning rate parameter, with a value between 0 and 1, used to control the magnitude of updates; As a participant In the consensus reached In each round, the proposal vector corresponds to the final decision option. The probability value.

[0107] Subsequently, to ensure that the updated weight set satisfies the constraint that the sum of all weights is 1, the weight update unit 440 normalizes all newly calculated unnormalized weights. The normalization formula is as follows:

[0108] ;

[0109] In the formula, As a participant The final updated and normalized new reputation weights; As a participant The updated, unnormalized new reputation weights; This represents the total number of participating parties.

[0110] Finally, the weight update unit 440 stores the normalized new reputation weights for this set. It is then provided to the encrypted aggregation module 200 as a reputation weight in the next collaborative decision-making task.

[0111] Please see the appendix Figure 2 , Figure 2 This is a schematic flowchart of a privacy-preserving collaborative decision-making method according to another embodiment of the present invention; the privacy-preserving collaborative decision-making method includes the following steps:

[0112] Step 1: Generating and Encrypting Proposals: Each participant generates a probabilistic action proposal vector representing their decision-making tendency based on their local private data and received feedback signals. The probabilistic action proposal vector is then encrypted to generate an encrypted proposal vector.

[0113] In this step, each participant performs this step independently. Specifically, each participant, based on its locally owned private data that is not visible to outsiders, and the global feedback signals received in the previous decision round, generates a probabilistic action proposal vector using a reinforcement learning-based decision-making strategy. Each component of this vector represents the participant's support for a specific decision option. To protect decision privacy, the participant encrypts the generated proposal vector using the public key of a homomorphic encryption scheme, thus outputting an encrypted proposal vector.

[0114] Step 2, Encrypted Aggregation: Based on the reputation weight set for each participant, the received multiple encrypted proposal vectors are weighted and aggregated in the ciphertext domain to generate an encrypted aggregation decision potential vector.

[0115] In this step, encrypted proposal vectors from all participants are collected, and a reputation weight representing the influence of each participant is pre-defined. Leveraging the ciphertext operation capability of homomorphic encryption schemes, a weighted aggregation operation within the ciphertext domain is performed on the received multiple encrypted proposal vectors without decrypting any data. The result of this operation is a single encrypted vector, namely the encrypted aggregation decision potential vector, which, in its encrypted state, represents the weighted collective decision-making tendency of all participants.

[0116] Step 3, Consensus Check: Perform a consensus check based on the encrypted aggregated decision potential vector to determine whether a decision consensus has been reached;

[0117] In this step, the encrypted aggregate decision potential vector is received, and the vector is analyzed in the ciphertext domain by executing one or more secure computing protocols to check whether the decision-making tendency of all participants has reached a preset level of concentration, thereby determining whether a decision consensus has been reached.

[0118] Step 4: Output Decision: When the consensus check result indicates that a consensus has been reached, output the privacy protection collaborative decision.

[0119] In this step, if the consensus check results in a decision consensus, then this step is executed. In this step, the final encrypted aggregated decision potential vector is decrypted to recover the plaintext aggregated result. Subsequently, by locating the decision option corresponding to the component with the largest value in the plaintext vector, the final result of this collaborative decision is determined and output.

[0120] If the consensus check results in no consensus on decision-making, then a topological structure analysis is performed on the encrypted aggregate decision potential vector in the ciphertext domain to generate structured feedback information that does not contain specific numerical values.

[0121] The structured feedback information is then sent as a feedback signal to multiple participants for their use in step 1 of the next round of decision-making.

[0122] In this step, if the consensus check results in no consensus being reached, a topological structure analysis is performed on the encrypted aggregate decision potential vector within the encrypted domain. This analysis aims to identify the shape of the current collective decision distribution, such as a unimodal, multimodal, or flat distribution. Based on the analyzed shape, structured feedback information without specific numerical values ​​is generated. This information, as a standardized feedback signal, is sent back to all participants. Upon receiving this feedback signal, each participant returns to step 1 and uses it as part of updating the environment state in the next round of decision-making to adjust their respective reinforcement learning strategies, thereby generating new proposal vectors aimed at driving the system towards consensus, and initiating a new round of iteration.

[0123] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A privacy-preserving collaborative decision-making system, characterized in that, include: Multiple proposal generation and encryption modules are provided. Each of the proposal generation and encryption modules is used to generate a probabilistic action proposal vector representing a decision tendency based on local private data and received feedback signals, and to encrypt the probabilistic action proposal vector to generate an encrypted proposal vector. The encrypted aggregation module, based on the reputation weight set for each of the proposal generation and encryption modules, performs weighted aggregation within the ciphertext domain on multiple encrypted proposal vectors received by the multiple proposal generation and encryption modules to generate an encrypted aggregation decision potential vector. The decision analysis and feedback module performs a consensus check based on the encrypted aggregated decision potential vector to determine whether a decision consensus has been reached. The decision output module outputs a privacy-protected collaborative decision when a consensus is reached on the decision. The decision analysis and feedback module is also used for: When the consensus check result is that no decision consensus has been reached, a topological structure analysis is performed on the encrypted aggregate decision potential vector within the ciphertext domain. Based on the results of the topology analysis, structured feedback information without specific numerical values ​​is generated; The structured feedback information is sent as the feedback signal to the plurality of proposal generation and encryption modules; The generation of probabilistic action proposal vectors representing decision-making tendencies uses local private data. and feedback signals As input, a probabilistic action proposal vector is calculated using the probabilistic action proposal vector calculation formula, which is: ； In the formula, As a participant In the Probabilistic action proposal vectors generated in each round; As a participant The decision strategy function; As a participant Local private data or observations held; For the system in the first The feedback signal broadcast at the end of the cycle; The total number of discrete options in the decision action space; Indicates the participating parties In the Round proposal selection The vector contains the probability values ​​of several decision options, and each element of the vector satisfies the probability distribution constraints. , ; The decision output module is specifically used for: When the consensus check results in a consensus on the decision, a final decision is determined. Based on the consistency between the probabilistic action proposal vector submitted by each of the proposal generation and encryption modules and the final decision as input, the reputation weight set by the corresponding proposal generation and encryption module is updated using the reputation weight update formula. The reputation weight is used by the encryption aggregation module when generating the encrypted aggregation decision potential vector. The reputation weight update formula is expressed as follows: ; In the formula, As a participant Updated, unnormalized reputation weights; As a participant Reputation weight before update; The preset learning rate parameter is a constant between 0 and 1, used to control the magnitude of updates; As a participant In the consensus reached In a round, the proposal vector corresponds to the final decision option. The probability value.

2. The privacy-protecting collaborative decision-making system according to claim 1, characterized in that, Each of the aforementioned proposal generation and encryption modules is further used for: The feedback signal is received, and the decision-making strategy of the proposal generation and encryption module is adjusted based on the feedback signal to generate the probabilistic action proposal vector for the next round.

3. The privacy-protecting collaborative decision-making system according to claim 1, characterized in that, When the consensus check results in a consensus on decision-making, the decision analysis and feedback module is further used for: Perform a decryption operation on the encrypted aggregation decision potential vector to obtain the plaintext aggregation decision potential vector; The final decision is determined based on the maximum value in the aggregated decision potential vector of the plaintext.

4. The privacy-protecting collaborative decision-making system according to claim 1, characterized in that, The encrypted aggregated decision potential vector is used to characterize the decision distribution shape, which includes one of the following: unimodal distribution, multimodal distribution, or flat distribution.

5. A privacy-preserving collaborative decision-making system according to claim 2, characterized in that, The decision-making strategy includes a reinforcement learning strategy; the proposal generation and encryption module adjusts the decision-making strategy in the following ways: The structured feedback information is used as part of the environment state of the reinforcement learning policy, and the reward function of the reinforcement learning policy is calculated based on the structured feedback information.

6. The privacy-protecting collaborative decision-making system according to claim 1, characterized in that, In the proposal generation and encryption module, the encrypted proposal vector is generated through homomorphic encryption, and the weighted aggregation within the ciphertext domain performed by the encryption aggregation module is generated using the homomorphic encryption and homomorphic scalar multiplication.

7. A privacy-preserving collaborative decision-making method, based on the privacy-preserving collaborative decision-making system according to any one of claims 1-6, characterized in that, Includes the following steps: Step 1: Generating and Encrypting Proposals: Each of the participating parties generates a probabilistic action proposal vector representing a decision-making tendency based on its local private data and received feedback signals, and then encrypts the probabilistic action proposal vector to generate an encrypted proposal vector. Step 2, Encrypted Aggregation: Based on the reputation weight set for each participant, the received multiple encrypted proposal vectors are weighted and aggregated in the ciphertext domain to generate an encrypted aggregation decision potential vector. Step 3, Consensus Check: Perform a consensus check based on the encrypted aggregated decision potential vector to determine whether a decision consensus has been reached; Step 4: Output Decision: When the consensus check result indicates that a consensus has been reached, output the privacy protection collaborative decision.

8. The privacy-preserving collaborative decision-making method according to claim 7, characterized in that, Step 3 also includes: If the consensus check results in no consensus being reached, then a topological analysis is performed on the encrypted aggregate decision potential vector within the ciphertext domain to generate structured feedback information that does not contain specific numerical values. The structured feedback information is then sent as a feedback signal to the multiple participants for their use in step 1 of the next round of decision-making.

Images