Text privacy data exchange method and system based on multi-party secure computation

Abstract

The application discloses a kind of text privacy data exchange method and system based on multi-party secure computation, belong to information security access control technical field;Method includes obtaining platform discrete request data computing queuing congestion state, generates asynchronous queue delay index;Obtain node history access log extraction topological structure difference, generate abnormal behavior score;With both as input, carry out cross-domain tensor product calculation, generate mixed defense control instruction by penalty activation mapping;Parse the instruction to extract buffer queue delay duration and decoy ciphertext generation proportion;Accordingly adjust data request release rule, and synchronously mix into the same state sum offset state decoy ciphertext;The application is jointly coordinated by space-time dimension, effectively resists the time sequence sniffing and ciphertext analysis to access mode, under the premise of guaranteeing legal business high throughput and joint calculation result absolute lossless, constructs high robustness dynamic data access immune barrier.

Description

Technical Field

[0001] This invention relates to the field of information security access control technology, specifically to a text privacy data exchange method and system based on multi-party secure computation. Background Technology

[0002] With the accelerated flow of data elements, multi-party secure computation, as a key technology for breaking down "data silos," has been widely applied in cross-organizational scenarios for joint modeling and exchange of privacy-preserving data. Systems often rely on stringent access control rules to ensure the security of protected objects. However, as application scenarios evolve towards high-concurrency, low-latency real-time interaction, traditional static key distribution and offline access control strategies are no longer sufficient to handle the highly dynamic data interaction needs within platforms, posing profound challenges to system-level access scheduling security.

[0003] In existing multi-party secure computation platforms, the access control of text privacy data faces the following limitations: In terms of temporal defense, to counter malicious nodes' attempts to sniff out encrypted text tags by analyzing data release patterns through side-channel packet capture, current techniques typically employ mechanically superimposed random delays to disrupt temporal characteristics. The limitation of this approach is that static, indiscriminate delays severely degrade the concurrent throughput of multi-party computation protocols, leading to system computational load imbalance. In terms of spatial defense, existing techniques attempt to inject fake data into the data stream for interference, but conventional fake traffic is easily cleaned and eliminated by advanced persistent threats based on queuing theory models. More seriously, disordered poisoning can easily contaminate the accuracy of the joint computation results of legitimate nodes. These two types of techniques are mutually exclusive in engineering implementation, making it difficult to establish a collaborative defense loop that balances timeliness and data integrity.

[0004] To this end, the present invention provides a method and system for exchanging text privacy data based on multi-party secure computation. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a text privacy data exchange method and system based on multi-party secure computation, thereby solving the technical problems described in the background section.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A text privacy data exchange method based on multi-party secure computation includes the following steps: S1. Obtain discrete request data in the current multi-party secure computing platform, calculate the queuing congestion status of discrete request data in the platform network, and generate an asynchronous queue latency index. S2. Obtain the historical access logs of the target computing node, calculate the degree of deviation between the access trajectory of the target computing node and the normal business baseline, and generate an abnormal behavior score. S3. Using the asynchronous queue delay index and abnormal behavior score as input, perform cross-domain tensor product calculation, and perform feature mapping through the penalty activation function to generate a hybrid defense control command that represents the spatiotemporal joint scheduling strategy. S4. Parse the hybrid defense control instructions, extract the buffer queue delay duration used to control the request release time, and the decoy ciphertext generation ratio used to determine the amount of fake data mixed in. S5. Send the buffer queue delay duration and the decoy ciphertext generation ratio to the access control module. The access control module adjusts the data request permission rules for the target computing node according to the buffer queue delay duration, and drives the fake data generation module to mix in the decoy ciphertext that conforms to the homomorphic summation cancellation state according to the generation ratio, thereby performing dynamic access protection for text privacy data.

[0007] A text privacy data exchange system based on multi-party secure computation includes a queuing congestion and delay analysis module, an abnormal behavior scoring generation module, a spatiotemporal joint scheduling decision module, a defense instruction dimensionality reduction parsing module, and a dynamic access protection execution module. The queuing congestion and delay analysis module is configured to acquire discrete request data in the current multi-party secure computing platform, calculate the queuing congestion status of discrete request data in the platform network, and generate an asynchronous queue delay index. The abnormal behavior scoring generation module is configured to obtain the historical access logs of the target computing node, calculate the degree of deviation between the access trajectory of the target computing node and the normal business baseline, and generate an abnormal behavior score. The spatiotemporal joint scheduling decision module is configured to perform cross-domain tensor product calculation with asynchronous queue delay index and abnormal behavior score as input, and perform feature mapping through penalty activation function to generate hybrid defense control instructions that characterize the spatiotemporal joint scheduling strategy. The defense command dimensionality reduction parsing module is configured to parse hybrid defense control commands, extract the buffer queue delay duration used to control the request release time, and extract the decoy ciphertext generation ratio used to determine the amount of fake data mixed in. The dynamic access protection execution module contains an access control unit and a fake data generation unit. The dynamic access protection execution module is configured to send the buffer queue delay duration and the decoy ciphertext generation ratio to the access control unit. The access control unit adjusts the data request release rules for the target computing node according to the buffer queue delay duration, and drives the fake data generation unit to mix in decoy ciphertext that conforms to the homomorphic summation cancellation state according to the decoy ciphertext generation ratio, thereby performing dynamic access protection for text privacy data.

[0008] (III) Beneficial Effects This invention obtains the asynchronous queue latency index and abnormal behavior score, and generates hybrid defense control instructions through cross-domain tensor product calculation. It overcomes the resistance of mutual exclusion between delay strategy and poisoning strategy in the prior art. With only the high-dimensional tensor fusion core, it can synchronously trigger time-series release discretization and decoy directional obfuscation in the access control engine. The time dimension delay provides a computing power window for generating high-overhead homomorphic decoys, while the decoy ciphertext volume fills the traffic trough caused by the delay, achieving a highly concealed collaborative defense effect. After analyzing the decoy cipher generation ratio, this invention drives the fake data generation module to mix in positive and negative cancellation ghost shares that conform to the homomorphic encryption algebra rules. This causes the injected pair of positive and negative decoy ciphers to force high-risk nodes to consume huge amounts of computing power to decrypt massive amounts of fake tags, effectively increasing the computing power exhaustion multiplier of reverse reasoning. When legitimate nodes perform multi-party calculations and summaries, the opposite numbers automatically add up to zero under homomorphic addition. This solves the technical bottleneck of traditional fake data injection easily polluting the final joint calculation result, achieving global lossless mixing of fake data. This invention forcibly introduces a capacity truncation verification step based on the maximum length threshold of the underlying gateway when calculating the algebraic difference between the instantaneous request arrival rate and the historical processing rate based on the delay differential equation; it integrates the open-loop accumulation of the theoretical differential equation with the physical capacity of the objective network queue in a closed loop, effectively blocking the risk of differential value overflow and physical memory penetration caused by extreme unbounded traffic surges, and ensuring the system-level high availability of the platform under high concurrency requests; In the process of generating all control parameters, this system employs dual read / write cursors and extracts the API endpoint call topology, physically isolating the data probes within the network protocol stack and API log layer. This enables the system to accurately locate unknown sniffing probes under the stringent conditions of zero-knowledge proofs and absolutely no contact with the original plaintext content. From the code architecture level, it cuts off the physical path for using business data to create user profiles or monitor violations, achieving a deep decoupling between technical efficiency and legal privacy compliance. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the steps of the text privacy data exchange method based on multi-party secure computation of the present invention. Figure 2 This is a schematic diagram of the system architecture of the text privacy data exchange system based on multi-party secure computation of the present invention. Detailed Implementation

[0010] The technical solutions of 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.

[0011] Please see Figure 1 This invention provides a text privacy data exchange method based on multi-party secure computation, comprising the following steps: S1. Obtain discrete request data in the current multi-party secure computing platform, calculate the queuing congestion status of discrete request data in the platform network, and generate an asynchronous queue latency index. S2. Obtain the historical access logs of the target computing node, calculate the degree of deviation between the access trajectory of the target computing node and the normal business baseline, and generate an abnormal behavior score. S3. Using the asynchronous queue delay index and abnormal behavior score as input, perform cross-domain tensor product calculation, and perform feature mapping through the penalty activation function to generate a hybrid defense control command that represents the spatiotemporal joint scheduling strategy. S4. Parse the hybrid defense control instructions, extract the buffer queue delay duration used to control the request release time, and the decoy ciphertext generation ratio used to determine the amount of fake data mixed in. S5. Send the buffer queue delay duration and the decoy ciphertext generation ratio to the access control module. The access control module adjusts the data request permission rules for the target computing node according to the buffer queue delay duration, and drives the fake data generation module to mix in the decoy ciphertext that conforms to the homomorphic summation cancellation state according to the generation ratio, thereby performing dynamic access protection for text privacy data.

[0012] The steps for generating the asynchronous queue latency index are as follows: synchronously extract the API call entry timestamp sequence representing the arrival of the request and the API call completion timestamp sequence representing the completion of the processing, call the initial sliding time window boundary configured, determine the concurrent request density within the current time window through the delay differential equation calculation logic, calculate the queuing congestion state of discrete request data in the platform network based on the concurrent request density, and generate the asynchronous queue latency index.

[0013] The specific steps for calculating the delay differential equation are as follows: 1. Count the total number of timestamp sequences entering the current time window and divide it by the time span of the current time window to calculate the instantaneous request arrival rate at the current moment. 2. Backtrack the historical timeline by a preset time delay to locate the historical reference time window and count the total number of completed timestamp sequences within the historical reference time window, dividing it by the corresponding time span to calculate the request processing rate at the historical moment. 3. Calculate the algebraic difference between the instantaneous request arrival rate and the request processing rate. 4. Multiply the algebraic difference by a preset congestion feedback coefficient and add the resulting product as a differential update amount to the density base value of the previous period to output the current concurrent request density.

[0014] In a more preferred implementation, to address the risks of differential value overflow and physical memory overflow caused by extreme unbounded traffic surges, a buffer capacity truncation and verification step is forcibly introduced before outputting the concurrent request density: The maximum length threshold of the pre-allocated physical buffer queue within the multi-party secure computing platform gateway node is obtained; after algebraically summing the differential update amount with the density base value of the previous period to obtain the intermediate computation density, a memory boundary determination is performed on the intermediate computation density; if it is determined that the intermediate computation density absolutely exceeds the maximum length threshold, a high-bit overflow circuit breaker mechanism is forcibly triggered, discarding the original accumulation result and directly truncating the current concurrent request density and assigning it the maximum length threshold. The value is set, and the underlying gateway defense rate limiting command is triggered in parallel. If the intermediate calculation density is determined to be less than the absolute zero value, the low-bit penetration error prevention compensation is triggered, and the current concurrent request density is forcibly set to 0. The maximum length threshold mentioned above is at the physical level that characterizes the upper limit of the actual buffer pool space resources of the system. In this embodiment, it is forcibly set to 10000 based on the gateway memory page allocation size. By integrating the open-loop accumulation of the theoretical differential equation with the closed-loop of the objective network queue physical capacity, the forced boundary truncation and zero-reconstruction are strictly performed to completely block the kernel calculation crash or memory penetration caused by the differential evolution deviating from the actual physical capacity limit, and effectively ensure the high availability and robustness of the congestion assessment model.

[0015] The steps for generating an abnormal behavior score are as follows: Read the access state machine transition records of the platform gateway, use the API endpoints in the access state machine transition records as network nodes, and use the order and frequency of calls between API endpoints as the weights of directed edges to construct a dynamic call graph representing the current access trajectory; call the configured view neural network topology confidence threshold, use the view neural network to compare the topology differences between the dynamic call graph and the pre-stored global security baseline graph model, and calculate the degree of deviation between the access trajectory of the target computing node and the normal business baseline based on the topology differences, i.e., the abnormal behavior score.

[0016] The specific steps of the view neural network comparison are as follows: the current adjacency feature matrix is ​​used as the forward propagation input and injected into the graph convolutional hidden layer of the view neural network; message passing of node features and aggregation of neighbor features are performed to output the dimension-reduced current graph embedding representation vector; the baseline graph embedding representation vector pre-calculated by the global safe baseline graph model is obtained simultaneously; then the spatial Euclidean distance between the current graph embedding representation vector and the baseline graph embedding representation vector is calculated, and the spatial Euclidean distance is used as a quantitative indicator of the topological structure difference.

[0017] Before comparing the topological differences between the dynamic call graph and the pre-stored global security baseline graph model, the view neural network undergoes offline pre-training and baseline encoding on offline high-computing cluster nodes. This involves extracting historical access logs and signaling handshake records from a pre-defined time period under clean business conditions to construct multiple sample call graphs without abnormal node contamination. These sample call graphs are then input into the view neural network, using the reconstruction error of the context node topology as the loss function. Iterative optimization through forward propagation and backpropagation continues until the reconstruction error converges below a pre-defined loss threshold, thus solidifying the network node aggregation weight parameters in physical storage. Each input to the view neural network with solidified parameters undergoes manual security auditing. The benchmark-level access call graph is recognized, and the low-dimensional spatial vector output by the hidden layer of the network is extracted and solidified as the benchmark graph embedding representation vector of the pre-stored global safe baseline graph model, which resides in the online working memory for a long time. The preset loss threshold is set in the physical judgment system that characterizes the lower limit of the network's tolerance for the accuracy of baseline topology reconstruction. In this embodiment, the preferred value is 0.05, but it can be any real number between [0.01, 0.1] under the actual system signal-to-noise ratio conditions. By introducing the offline reconstruction learning model training step and the benchmark embedding residence strategy, the engineering pre-mechanism of the neural network evolving from a random blank board to a usable feature extractor is filled. While ensuring the purity of the global benchmark, the online high-concurrency latency is greatly compressed, and the operation of calculating feature differences is reduced to Euclidean distance retrieval.

[0018] In a preferred embodiment, the system acquires the aforementioned indicators using two isolated dual read / write cursors to avoid interference from subsequent timing disruptions on the initial baseline assessment. To achieve technical decoupling between the core algorithm logic and specific platform business strategies, all configurable operating parameters (in this embodiment, including the initial sliding time window boundary of the delay differential equation and the topological confidence threshold of the view neural network, etc.) are predefined and stored in a structured external data carrier. The system reads the file from the external data carrier, deserializes the content, and loads it into the program's working memory dictionary. When S1 and S2 calculate the concurrent request density and topological differences, the system directly addresses and retrieves the required baseline parameters from the memory data structure to ensure low-latency computation in a high-concurrency network environment. By separately collecting congestion status and behavioral deviation, the system completely separates objective environmental factors from subjective malicious factors. The extracted asynchronous queue latency index accurately reflects the current system's computing power load capacity. The abnormal behavior score, under the stringent conditions of zero-knowledge proof and without touching the original text content, accurately identifies internal unauthorized sniffing requests disguised as normal API calls, achieving early locking of unknown sniffing probes.

[0019] The steps for generating hybrid defense control commands are as follows: Extract asynchronous queue latency indices generated over N consecutive historical periods, including the current time, from the dynamic sliding buffer queue, and concatenate them sequentially to construct a multi-dimensional time-domain state vector; synchronously extract N consecutive abnormal behavior scores within the corresponding time window and construct them into a multi-dimensional spatial-domain behavior vector; perform Kronecker product operations on the time-domain state vector and the spatial-domain behavior vector to output a spatiotemporal entanglement feature matrix; perform matrix eigenvalue decomposition on the spatiotemporal entanglement feature matrix, extract the largest eigenvalue as the core eigenvalue, and determine whether the core eigenvalue exceeds a preset safety baseline threshold; in response to the core eigenvalue exceeding the safety baseline threshold, trigger the nonlinear amplification factor within the penalty activation function to exponentially amplify the core eigenvalue and serialize it into a hybrid defense control command.

[0020] The steps for responding to a core feature value exceeding a safety baseline threshold, triggering a nonlinear amplification factor within the penalty activation function, exponentially amplifying the core feature value, and serializing it into a hybrid defense control command are as follows: Extracting multiple discrete snapshot data points of the core feature value within a preset historical time window; calculating the dispersion of the multiple discrete snapshot data points to generate a fluctuation evolution rate; extracting a preset maximum expected fluctuation constant, and using logarithmic normalization mapping logic to calculate the logarithmic ratio of the fluctuation evolution rate to the maximum expected fluctuation constant, mapping the fluctuation evolution rate to a dynamic tolerance attenuation coefficient ranging from 0 to 1; obtaining a baseline amplification base, obtaining a preset anti-overflow positive constant, adding the dynamic tolerance attenuation coefficient to the preset anti-overflow positive constant to obtain a safety denominator, then dividing the baseline amplification base by the safety denominator to obtain the modulated amplification base; replacing the default base of the nonlinear amplification factor with the modulated amplification base to exponentially amplify the core feature value, generating amplified feature parameters; encapsulating the amplified feature parameters in a format and serializing them into a hybrid defense control command.

[0021] The calculation and fusion steps of the dynamic tolerance attenuation coefficient are as follows: First, obtain multiple discrete core feature values ​​within the current time and historical time windows as the input set. Second, calculate the mathematical variance of the input set, using the obtained value as an intermediate result, defined as the fluctuation evolution rate. Third, add a constant of 1 to the fluctuation evolution rate, and then perform a natural logarithmic operation to obtain a logarithmic smoothing value. Fourth, obtain the preset maximum expected fluctuation constant by parsing the system configuration pre-stored in an external spreadsheet file (in this embodiment, the preferred value of the maximum expected fluctuation constant is set to 4.6. Under the theoretical expectation that the maximum tolerable variance mutation extreme value is 100, adding 1 and taking the natural logarithm is approximately 4.6, thus ensuring smooth mapping for subsequent normalization). Fifth, divide the logarithmic smoothing value by the maximum expected fluctuation constant, and forcefully truncate the result to the absolute range of 0 to 1 using a threshold truncation function to obtain the dynamic tolerance attenuation coefficient. Sixth, obtain the configured baseline amplification base, divide the baseline amplification base by the dynamic tolerance attenuation coefficient to obtain the modulated amplification base. Seventh, use the modulated amplification base as the base and the core feature value exceeding the threshold as the exponent for power operation to obtain the final amplified feature parameter.

[0022] By constructing an asynchronous queue latency index reflecting temporal congestion and an abnormal behavior score reflecting spatial anomalies into a vector and then performing tensor product operations, the system accurately identifies the correlation characteristics between network congestion and abnormal access concurrency. Instead of directly physically disconnecting high-risk connections, it constructs a spatiotemporally distorted local defense domain based on the tensor product results. When the penalty activation function is triggered, the generated instructions ensure that the spatial dimension of the decoy ciphertext release concentration closely matches the temporal dimension of the dark pool queuing latency, achieving precise targeted poisoning. The fusion process generates a clear and quantifiable indicator: the malicious computing power exhaustion multiplier. By monitoring the output of the comprehensive decision tensor, the system can amplify the average computation time consumed by a malicious node in a single attempt to perform text tag reverse reasoning to 10 times that under normal business conditions. 3 Orders of magnitude higher; after aggregating two independent monitoring dimensions into a single high-dimensional decision valve, the system will neither mistakenly kill legitimate requests due to normal business peak congestion nor miss insiders conducting low-frequency covert sniffing during off-peak periods; the feature product amplification mechanism successfully creates a computational network with extremely high information density but no logical value for malicious nodes at the access control protocol layer.

[0023] The steps for dimensionality reduction of the hybrid defense control command are as follows: The single amplified feature parameter contained within the hybrid defense control command is used as input. A dual-channel mapping coefficient configured in the working memory is introduced. The amplified feature parameter is multiplied by a preset delay conversion coefficient. The buffer queue delay duration used to control request release time is extracted through linear calculation. The amplified feature parameter is substituted into a logarithmic proportional function with a preset saturation limit as the denominator to extract the decoy ciphertext generation ratio used to determine the amount of spurious data.

[0024] Furthermore, the latency conversion coefficient and the preset saturation limit are determined based on the extreme game calibration and deduction of the physical operating conditions of the target computing node: The service level agreement signed by the multi-party secure computing platform for the current business flow is read and parsed in advance, the highest tolerable physical timeout limit duration under the business scenario is extracted, and this duration is converted by security attenuation to obtain the latency conversion coefficient, ensuring that the amplified latency penalty will never trigger a brute-force disconnection by the front-end gateway; the underlying hardware performance polling monitoring interface is read to obtain the peak encryption / decryption throughput capacity index of the cryptographic machine or encryption coprocessor within the current secure computing unit, and the throughput limit is converted into the absolute maximum number that can be imperceptibly carried and mixed with false shares within a unit time window, and this objective capacity is set as the preset saturation limit; latency In the objective feedback loop that controls the penalty intensity in the time dimension, the conversion coefficient is preferably 15.5 milliseconds / unit feature value in this embodiment, and the reasonable range that can be dynamically adapted is [5, 50] milliseconds / unit feature value; in the objective computing power boundary that characterizes the physical carrying capacity limit of the system against the data deluge, the preset saturation limit value is preferably 2048 concurrent fake shares in this embodiment, and its actual carrying capacity range is [512, 4096] discrete shares; by introducing parameter deduction and extreme value calibration methods that are completely anchored to the real hardware physical upper limit and service bottom line, the extreme disasters of blind calculation causing normal requests to be circuit-broken on a large scale due to timeout, or encryption hardware to be overloaded and crash due to excessive issuance of fake decoys are completely blocked, and a dynamic balance of high business availability is achieved while ensuring high-intensity security suppression.

[0025] The steps for generating the decoy ciphertext in the homomorphic summation cancellation state are as follows: Receive the buffer queue delay duration and the decoy ciphertext generation ratio; based on the buffer queue delay duration, forcibly extend the polling wait period window of the corresponding target computing node; generate the first decoy ciphertext data according to the decoy ciphertext generation ratio; and synchronously generate the second cancellation ciphertext data with absolutely equal values ​​but opposite signs under the homomorphic encryption algebra rules; forcibly bind the first decoy ciphertext data and the second cancellation ciphertext data within the same polling wait period window, and distribute them as a pair of ghost shares mixed into the real multi-party computing task sequence.

[0026] Furthermore, to ensure the theoretical concept of automatic ciphertext cancellation is implemented on deterministic cryptographic architectures, a Paillier public-key infrastructure system with additive isomorphism is forcibly adopted for transmission. When generating the first decoy ciphertext data, the squared modulus parameter of the global public key allocated to the secure channel is extracted, and the randomly generated decoy plaintext is mapped and encapsulated into the underlying Paillier homomorphic ciphertext field. When synchronously calculating the second cancellation ciphertext data, based on the number-theoretical closure property of the additive homomorphic space, and under the constraint of maintaining complete consistency of the underlying encryption random number configuration factor, large-number modular arithmetic operations based on the global public key modulus are used to obtain the additive inverse plaintext data in the modulus space for the first decoy plaintext. Then, the same encryption mechanism is used to homomorphically encapsulate and encrypt the additive inverse plaintext data to generate... The second offset ciphertext data ensures that the ciphertext product of the phantom share, after being decrypted and restored by the homomorphic private key in the hardware arithmetic logic unit at the receiving end, is always mapped to the absolute value 0 in the plaintext group. The preferred engineering deployment configuration of this scheme is at the 2048-bit binary length level, and the key length range that can be extended in actual scenarios is [1024, 4096] bits. By sinking and binding the broadly conceptualized homomorphic algebra operation to the specific Paillier ciphertext field modular inverse transformation code rule execution level, the illegal questioning of cryptographic objects caused by purely mathematical transformations is eliminated. This not only makes the false share implanted in the circulation channel have extremely high concealment at the ciphertext capture level, but also provides a mathematically unshakable absolute zeroing guarantee, protecting the backend statistical summary system from data contamination.

[0027] The monitoring steps for decoy ciphertext in the homomorphic summation cancellation state are as follows: When it is detected that the node computing power load saturation index exceeds the preset saturation defense limit due to the continuous generation of ghost shares, the decoy ciphertext generation ratio is forcibly cleared to stop the generation of false shares, and the buffer queue delay time is increased proportionally; relying on the distribution of mixed task sequences and the change of platform release rules, the positive and negative decoy values ​​are reduced to 0 and the real accuracy is not lost when the legitimate summary is made, while the malicious nodes are forced to exhaust computing power, thereby completing the dynamic access protection for text privacy data.

[0028] In S4, the high-dimensional hybrid defense control instructions are parsed down to extract two specific execution parameters (delay duration and generation ratio). In S5, the underlying platform rule scheduling and ghost loop cancellation algorithm are executed. The buffer queue delay duration is injected into the network scheduler, forcibly lengthening the response time of the target node. During this extended time interval, the fake data generation module begins to operate. To ensure the homomorphic summation cancellation state, for each generated decoy ciphertext, the underlying operator must also calculate the opposite cancellation ciphertext. The two are inserted into the same computation batch as a pair of ghost shares. When the hardware probe detects that the GPU memory usage rate of the computing node exceeds the set saturation defense limit (in this embodiment, the saturation defense limit is set to 85%) due to the generation of a large amount of homomorphic fake data, the arm-breaking self-defense mechanism is activated. The system automatically resets the decoy ciphertext generation ratio to zero, stops the generation of all fake shares, and proportionally increases the buffer queue delay time. The platform's defense mode is downgraded from poisoning and delay smoothing to pure extreme traffic denial delay. By changing the platform's access rules, it achieves perfect dual defense and business-free operation. The forced waiting caused by modifying the access rules completely breaks the causal relationship of the request packet on the timeline, cracking the temporal feature sniffing. The mixed-in positive and negative decoy ciphertexts force high-risk nodes to consume extremely large amounts of local computing power to decrypt massive amounts of fake tags. When the legitimate multi-party calculation results are summarized, the two opposite numbers are automatically added to zero under homomorphic addition. This fundamentally ensures that the decoy injection behavior does not interfere with the calculation accuracy of real text privacy exchange, achieving the extraordinary effect of single-point blinding and global losslessness.

[0029] Please see Figure 2 A text privacy data exchange system based on multi-party secure computation includes a queuing congestion and delay analysis module, an abnormal behavior scoring generation module, a spatiotemporal joint scheduling decision module, a defense instruction dimensionality reduction parsing module, and a dynamic access protection execution module. The queuing congestion and delay analysis module is configured to acquire discrete request data in the current multi-party secure computing platform, calculate the queuing congestion status of discrete request data in the platform network, and generate an asynchronous queue delay index. The abnormal behavior scoring generation module is configured to obtain the historical access logs of the target computing node, calculate the degree of deviation between the access trajectory of the target computing node and the normal business baseline, and generate an abnormal behavior score. The spatiotemporal joint scheduling decision module is configured to perform cross-domain tensor product calculation with asynchronous queue delay index and abnormal behavior score as input, and perform feature mapping through penalty activation function to generate hybrid defense control instructions that characterize the spatiotemporal joint scheduling strategy. The defense command dimensionality reduction parsing module is configured to parse hybrid defense control commands, extract the buffer queue delay duration used to control the request release time, and extract the decoy ciphertext generation ratio used to determine the amount of fake data mixed in. The dynamic access protection execution module contains an access control unit and a fake data generation unit. The dynamic access protection execution module is configured to send the buffer queue delay duration and the decoy ciphertext generation ratio to the access control unit. The access control unit adjusts the data request release rules for the target computing node according to the buffer queue delay duration, and drives the fake data generation unit to mix in decoy ciphertext that conforms to the homomorphic summation cancellation state according to the decoy ciphertext generation ratio, thereby performing dynamic access protection for text privacy data.

[0030] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A text privacy data exchange method based on multi-party secure computation, characterized in that, Includes the following steps: S1. Obtain discrete request data in the current multi-party secure computing platform, calculate the queuing congestion status of discrete request data in the platform network, and generate an asynchronous queue latency index. S2. Obtain the historical access logs of the target computing node, calculate the degree of deviation between the access trajectory of the target computing node and the normal business baseline, and generate an abnormal behavior score. S3. Using the asynchronous queue delay index and abnormal behavior score as input, perform cross-domain tensor product calculation, and perform feature mapping through the penalty activation function to generate a hybrid defense control command that represents the spatiotemporal joint scheduling strategy. S4. Parse the hybrid defense control instructions, extract the buffer queue delay duration used to control the request release time, and the decoy ciphertext generation ratio used to determine the amount of fake data mixed in. S5. Send the buffer queue delay duration and the decoy ciphertext generation ratio to the access control module. The access control module adjusts the data request permission rules for the target computing node according to the buffer queue delay duration, and drives the fake data generation module to mix in the decoy ciphertext that conforms to the homomorphic summation cancellation state according to the generation ratio, thereby performing dynamic access protection for text privacy data.

2. The text privacy data exchange method based on multi-party secure computation according to claim 1, characterized in that: The steps for generating the asynchronous queue latency index are as follows: synchronously extract the API call entry timestamp sequence representing the arrival of the request and the API call completion timestamp sequence representing the completion of the processing, call the initial sliding time window boundary configured, determine the concurrent request density within the current time window through the delay differential equation calculation logic, calculate the queuing congestion state of discrete request data in the platform network based on the concurrent request density, and generate the asynchronous queue latency index.

3. The text privacy data exchange method based on multi-party secure computation according to claim 2, characterized in that: The specific steps for calculating the delay differential equation are as follows: count the total number of timestamp sequences entering the current time window, divide by the time span of the current time window to calculate the instantaneous request arrival rate at the current moment; backtrack to the historical timeline by a preset time delay, locate the historical reference time window, and count the total number of timestamp sequences completed within the historical reference time window, divide by the corresponding time span to calculate the request processing rate at the historical moment. Calculate the algebraic difference between the instantaneous request arrival rate and the request processing rate; multiply the algebraic difference by a preset congestion feedback coefficient, and add the resulting product as a differential update amount to the density base value of the previous period, thereby outputting the current concurrent request density.

4. The text privacy data exchange method based on multi-party secure computation according to claim 1, characterized in that: The steps for generating an abnormal behavior score are as follows: Read the access state machine transition records of the platform gateway, use the API endpoints in the access state machine transition records as network nodes, and use the order and frequency of calls between API endpoints as the weights of directed edges to construct a dynamic call graph representing the current access trajectory; call the configured view neural network topology confidence threshold, use the view neural network to compare the topology differences between the dynamic call graph and the pre-stored global security baseline graph model, and calculate the degree of deviation between the access trajectory of the target computing node and the normal business baseline based on the topology differences, i.e., the abnormal behavior score.

5. The text privacy data exchange method based on multi-party secure computation according to claim 4, characterized in that: The specific steps of the view neural network comparison are as follows: the current neighbor feature matrix is ​​used as the forward propagation input and injected into the graph convolutional hidden layer of the view neural network; message passing of node features and aggregation of neighbor features are performed, and the current graph embedding representation vector with reduced dimension is output; The baseline graph embedding representation vector, which is pre-computed and output by the global security baseline graph model, is acquired synchronously. Then, the spatial Euclidean distance between the current graph embedding representation vector and the baseline graph embedding representation vector is calculated, and the spatial Euclidean distance is used as a quantitative indicator of the difference in topology.

6. The text privacy data exchange method based on multi-party secure computation according to claim 1, characterized in that: The steps for generating hybrid defense control commands are as follows: Extract asynchronous queue latency indices generated over N consecutive historical periods, including the current time, from the dynamic sliding buffer queue, and concatenate them sequentially to construct a multi-dimensional time-domain state vector; synchronously extract N consecutive abnormal behavior scores within the corresponding time window and construct them into a multi-dimensional spatial-domain behavior vector; perform Kronecker product operations on the time-domain state vector and the spatial-domain behavior vector to output a spatiotemporal entanglement feature matrix; perform matrix eigenvalue decomposition on the spatiotemporal entanglement feature matrix, extract the maximum eigenvalue as the core eigenvalue, and determine whether the core eigenvalue exceeds a preset safety baseline threshold. In response to the core feature value exceeding the safety baseline threshold, a non-linear amplification factor within the penalty activation function is triggered, exponentially amplifying the core feature value and serializing it into a hybrid defense control command.

7. The text privacy data exchange method based on multi-party secure computation according to claim 6, characterized in that: The calculation and fusion steps of the dynamic tolerance attenuation coefficient are as follows: Obtain multiple discrete core feature values ​​within the current time and historical time windows as the input set; calculate the mathematical variance of the input set, and use the obtained value as an intermediate result, defined as the fluctuation evolution rate; add a constant of 1 to the fluctuation evolution rate, and then perform a natural logarithmic operation to obtain a logarithmically smoothed value; obtain the preset maximum expected fluctuation constant by parsing the system configuration pre-stored in an external spreadsheet file; divide the logarithmically smoothed value by the maximum expected fluctuation constant, and forcefully truncate the calculation result within the absolute range of 0 to 1 using a threshold truncation function, thereby obtaining the dynamic tolerance attenuation coefficient; obtain the configured baseline amplification base, and divide the baseline amplification base by the dynamic tolerance attenuation coefficient to obtain the modulated amplification base; Using the modulated amplified base as the base, and the core feature value exceeding the threshold as the exponent, a power operation is performed to obtain the final amplified feature parameter.

8. The text privacy data exchange method based on multi-party secure computation according to claim 1, characterized in that: The steps for dimensionality reduction of the hybrid defense control command are as follows: The single amplified feature parameter contained within the hybrid defense control command is used as input. A dual-channel mapping coefficient configured in the working memory is introduced. The amplified feature parameter is multiplied by a preset delay conversion coefficient. The buffer queue delay duration used to control request release time is extracted through linear calculation. The amplified feature parameter is substituted into a logarithmic proportional function with a preset saturation limit as the denominator to extract the decoy ciphertext generation ratio used to determine the amount of spurious data.

9. The text privacy data exchange method based on multi-party secure computation according to claim 8, characterized in that: The steps for generating the decoy ciphertext in the homomorphic summation cancellation state are as follows: receive the buffer queue delay duration and the decoy ciphertext generation ratio; based on the buffer queue delay duration, forcibly extend the polling waiting period window of the corresponding target computing node; generate the first decoy ciphertext data according to the decoy ciphertext generation ratio. And under the homomorphic encryption algebra rules, a second offset ciphertext data with absolutely equal values ​​but opposite signs is generated synchronously; The first decoy ciphertext data and the second offset ciphertext data are forcibly bound within the same polling waiting period window and distributed as a pair of ghost shares in the real multi-party computation task sequence.

10. A text privacy data exchange system based on multi-party secure computation, applied to the text privacy data exchange method based on multi-party secure computation as described in any one of claims 1 to 9; characterized in that, It includes a queuing congestion and delay analysis module, an abnormal behavior scoring generation module, a spatiotemporal joint scheduling decision module, a defense command dimensionality reduction parsing module, and a dynamic access protection execution module. The queuing congestion and delay analysis module is configured to acquire discrete request data in the current multi-party secure computing platform, calculate the queuing congestion status of discrete request data in the platform network, and generate an asynchronous queue delay index. The abnormal behavior scoring generation module is configured to obtain the historical access logs of the target computing node, calculate the degree of deviation between the access trajectory of the target computing node and the normal business baseline, and generate an abnormal behavior score. The spatiotemporal joint scheduling decision module is configured to perform cross-domain tensor product calculation with asynchronous queue delay index and abnormal behavior score as input, and perform feature mapping through penalty activation function to generate hybrid defense control instructions that characterize the spatiotemporal joint scheduling strategy. The defense command dimensionality reduction parsing module is configured to parse hybrid defense control commands, extract the buffer queue delay duration used to control the request release time, and extract the decoy ciphertext generation ratio used to determine the amount of fake data mixed in. The dynamic access protection execution module contains an access control unit and a fake data generation unit. The dynamic access protection execution module is configured to send the buffer queue delay duration and the decoy ciphertext generation ratio to the access control unit. The access control unit adjusts the data request release rules for the target computing node according to the buffer queue delay duration, and drives the fake data generation unit to mix in decoy ciphertext that conforms to the homomorphic summation cancellation state according to the decoy ciphertext generation ratio, thereby performing dynamic access protection for text privacy data.

Images