An artificial intelligence-based cross-platform data interaction sharing method and system

Abstract

This invention belongs to the field of artificial intelligence technology, specifically relating to a cross-platform data interaction and sharing method and system based on artificial intelligence. It includes: constructing a health-aware edge cloud collaborative architecture to establish a dynamic health baseline for nodes; constructing a modified reward function with node health as a stability constraint, and using a multi-agent reinforcement learning model to train and obtain the optimal Nash equilibrium policy; decoding the optimal policy into an executable cross-platform data interaction planning graph through a policy interpreter network, and using game theory to coordinate resource conflicts; based on the planning graph, performing context-aware agent re-encryption and hierarchical homomorphic encryption, and establishing a power token chain to achieve fine-grained permission tracking; and performing closed-loop optimization of the policy and system self-evolution through multi-dimensional quality scoring, multi-layer source analysis, and elastic weight solidification technology; achieving synergistic gains in efficiency, security, and long-term stability of cross-platform data interaction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a cross-platform data interaction and sharing method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of technologies such as the Internet of Things and the Industrial Internet, the demand for data interaction and sharing across organizations and regions is becoming increasingly urgent. Such interactions are usually carried out on a heterogeneous collaborative architecture consisting of edge nodes and cloud platforms.

[0003] Existing technologies mainly face the following pain points: Current scheduling and routing strategies are mostly based on static rules or simple optimization algorithms, which cannot make globally coordinated optimal decisions under multiple dynamic changes in network topology, node load, and business needs. This leads to low data transmission efficiency and uneven resource utilization. Traditional security solutions (such as VPNs and static access control lists) are deployed independently of the business logic of data interaction. This "two-layer" model results in either excessive security, imposing high-overhead encryption on insensitive data, affecting efficiency, or insufficient security, failing to implement fine-grained, dynamic security control based on the specific flow and usage intent of data, posing a risk of data misuse or leakage. When existing systems experience node failures, performance degradation, or new security threats, they heavily rely on manual intervention for fault location, policy adjustment, and system upgrades, unable to learn from historical operating experience, leading to long-term performance degradation and high maintenance costs.

[0004] In conclusion, there is an urgent need for a cross-platform data interaction and sharing method and system based on artificial intelligence. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a cross-platform data interaction and sharing method and system based on artificial intelligence, to solve the following technical problem:

[0006] It is difficult to use node health indicators simultaneously as the basis for infrastructure monitoring, the negative feedback term of the reinforcement learning reward function, and the priority weight of experience replay sampling; it is also difficult to use the interaction planning graph generated by AI policy decoding as the sole blueprint for secure execution, driving the context-aware operation of agent re-encryption and power token chain; it lacks an incremental optimization mechanism based on elastic weight solidification technology and a security update mechanism driven by simulation verification, and it also lacks adaptive optimization scheduling based on feedback intensity.

[0007] To address the aforementioned problems, a first aspect of the present invention provides a cross-platform data interaction and sharing method based on artificial intelligence, comprising the following steps:

[0008] S1: Identify the multiple platforms involved in data interaction and sharing, as well as the data types, scale, and service quality requirements of their interactions. Simultaneously, build a collaborative architecture that includes multiple edge nodes and cloud platforms, construct a dynamic health baseline for each edge node, and calculate the node health in real time using a dynamic weighted algorithm.

[0009] S2: Construct a multivariate reward function that includes latency, bandwidth cost, data freshness and computational load, introduce the average health of nodes as a stability negative feedback to form a modified reward function, use the MADDPG algorithm to construct a multi-agent reinforcement learning model, train it through a health-weighted priority experience replay strategy, and obtain the optimal Nash equilibrium strategy.

[0010] S3: Input the optimal strategy into the strategy interpreter network, decode to generate a cross-platform data interaction planning graph, and form the final executable planning graph through a game theory-based resource conflict coordination mechanism, and decompose it into local execution subgraphs for each platform.

[0011] S4: Based on the interactive planning graph, the original data is encrypted using ciphertext policy attribute base encryption on the data source platform; interactive context-aware re-encryption agents are deployed on edge nodes; a hierarchical homomorphic encryption scheme is used to achieve cross-platform computation in ciphertext state for the collaborative computing task flow in the planning graph; a power token chain is established for each piece of shared data and stored on multiple edge nodes through lightweight consensus.

[0012] S5: Collect multi-dimensional quality indicators in real time during the interaction process and calculate the comprehensive quality score. When the comprehensive quality score fails to meet the standard after N consecutive interactions, locate the root cause through multi-level source tracing analysis and generate strategy optimization guidance signals.

[0013] Preferably, in step S1, constructing a dynamic health baseline and calculating node health in real time using a dynamic weighted algorithm includes the following steps:

[0014] Collect health data of edge nodes in different time periods over the past 90 days, extract trend and periodic items through time series decomposition algorithm, and generate dynamic health baseline by linear superposition function of the trend and periodic items.

[0015] Define a task execution efficiency index, and calculate the Spearman rank correlation coefficient between each monitoring index and the task execution efficiency index within a 30-day sliding time window; after normalizing the correlation coefficient, obtain the dynamic weight of the index.

[0016] Calculate the real-time health of edge nodes ,in, For the first Dynamic weights of each monitoring indicator To make the first Raw collected values ​​of each monitoring indicator A function that maps negative scores to the interval [0,1].

[0017] Preferably, step S1 further includes an edge node intelligent migration mechanism:

[0018] When the real-time health of the edge node meets When this occurs, the migration assessment process is triggered, in which... This is the preset health decay tolerance coefficient. This is the baseline value of the dynamic health of the edge node at the current moment;

[0019] For each candidate target node, the node's current health, resource availability, migration task profile, and local network state are input into a pre-trained gradient boosting decision tree model to output two predicted values, including predicted health and migration risk score.

[0020] Calculate the overall score for each candidate node. ,in, For the first The current health of each candidate node. For the first Risk score of the local network stability of each candidate node To migrate to the first The communication and synchronization overhead of each candidate node, , and These are the weighting coefficients;

[0021] choose highest and The candidate nodes are selected as the final migration targets.

[0022] Preferably, in step S2, modifying the reward function includes the following steps:

[0023] Based on the aforementioned service quality requirements, the four optimization objectives—latency, bandwidth cost, data freshness, and computational load—are quantified into reward components and then weighted and summed. The basic multivariate reward function is as follows: ,in, For negative delayed rewards, As a negative bandwidth cost incentive, To reward positive data freshness, To calculate the load reward negatively, , , and These are the initial weighting coefficients for each reward component;

[0024] Based on the aforementioned fundamental multivariate reward function, the real-time health of edge nodes is introduced as a global stability constraint to construct a modified reward function. ,in, The corrected reward function value. The health impact coefficient. This represents the average real-time health of all edge nodes involved in the current data interaction task.

[0025] Preferably, in step S2, a multi-agent reinforcement learning model is constructed using the MADDPG algorithm and trained using a health-weighted priority experience replay strategy, including the following steps:

[0026] The multi-agent reinforcement learning model constructs an actor network and a centralized critic network for each platform agent. Each actor network takes into account the local observation vector of the corresponding platform and outputs the action policy of the platform. The centralized critic network takes into account the concatenation of the local observation vectors and action vectors of all platforms and outputs a joint Q-value estimate of the current global state-joint action.

[0027] At each training time step, each agent generates an experience tuple based on its action strategy output by its actor network and its interaction with the environment. Stored in a globally shared experience replay pool; where, For local state observation, The action performed by the intelligent agent. To perform the action The environment then transitions to the next state observation vector;

[0028] When sampling in the experience replay pool for neural network parameter updates, a health-based priority sampling strategy is adopted, specifically: a corresponding sampling priority is calculated for each experience, specifically: ,in, For the first The time difference error of the empirical rule, This is the health bias coefficient. For the first The average health of relevant edge nodes during the experience period;

[0029] The policy network and value network of each agent are based on small batches of sampled experience data to maximize the cumulative correction reward for the future. The predicted value is the objective, and it is updated using the policy gradient method;

[0030] After training, the policy networks of all agents will converge collaboratively to a steady-state Nash equilibrium policy, which is the optimal policy.

[0031] Preferably, in step S3, the resource conflict coordination mechanism includes the following steps:

[0032] When multiple interactive planning graphs conflict over the same resource, the relevant agents conduct multiple rounds of bidding negotiations based on the task's criticality level and the resources they already hold.

[0033] A temporary Nash equilibrium for resource allocation is formed through a limited number of rounds of game playing;

[0034] The interaction planning diagram of each agent is dynamically updated based on the negotiation results.

[0035] Preferably, in step S4, deploying an interaction context-aware re-encryption proxy at the edge node includes the following steps:

[0036] When the re-encryption proxy receives a data request, it queries the real-time interaction context associated with this request;

[0037] Session-limiting parameters are generated based on the real-time interaction context, and a single-use valid re-encryption conversion key is synthesized using a lattice-based cryptographic algorithm.

[0038] The initial ciphertext is transformed using the re-encryption conversion key to generate a personalized re-encryption ciphertext;

[0039] The encrypted digest of the real-time interaction context is embedded in the generated ciphertext metadata to form a traceable interaction binding ciphertext.

[0040] Submit this conversion event to the Power Token Chain for evidence storage.

[0041] Preferably, in step S4, the power token chain includes the following steps:

[0042] Record the complete history of data changes in permissions from its generation to each access, calculation, and transfer.

[0043] When abnormal platform behavior or authorization expiration is detected, the relevant permissions are invalidated through a consensus protocol.

[0044] Supports fine-grained access auditing and accountability based on token chain.

[0045] Preferably, step S5 includes the following steps:

[0046] During each cross-platform data interaction, data utility indicators, collaboration efficiency indicators, and strategy compliance indicators are collected. These indicators are standardized and weighted and integrated to generate a comprehensive quality score for the data interaction.

[0047] The comprehensive quality score is compared with the expected value of the service quality requirement to calculate the quality gap; if the quality gap is ≤ the preset quality threshold in N consecutive interaction cycles, a systematic source analysis is initiated to locate the root cause of the quality gap from three dimensions: data content layer, interaction coordination layer, and security execution layer.

[0048] Based on the source tracing results, generate strategy optimization guidance signals;

[0049] The strategy optimization guidance signal and the corresponding complete state trajectory of the interaction process are jointly encapsulated into enhanced training samples. The elastic weight solidification technology is used to perform targeted incremental optimization on the multi-agent reinforcement learning model. During the optimization process, the corresponding guiding reward item is introduced or strengthened in the reward function according to the optimization direction indicated by the strategy optimization guidance signal.

[0050] When the improvement of the simulation interaction quality evaluation value compared with the baseline value obtained by the model before optimization in the same simulation environment is greater than the set improvement threshold, the optimization is deemed effective and the coordinated evolution of the entire system strategy is triggered.

[0051] The simulation interaction quality evaluation value is obtained by collecting various indicators, standardizing and weighting the results after fully executing the strategy generated by the optimized multi-agent reinforcement learning model in a high-fidelity simulation interaction environment.

[0052] A second aspect of the present invention provides a cross-platform data interaction and sharing system based on artificial intelligence, comprising the following modules:

[0053] Collaborative architecture initialization and health awareness module: Identify the multiple platforms involved in data interaction and sharing, as well as the data types, scale, and service quality requirements of their interactions. Simultaneously, build a collaborative architecture that includes multiple edge nodes and cloud platforms, construct a dynamic health baseline for each edge node, and calculate the node health in real time through a dynamic weighted algorithm.

[0054] Multi-agent cooperative policy learning module: Construct a multivariate reward function that includes latency, bandwidth cost, data freshness and computational load, introduce the average health of nodes as a stability negative feedback to form a modified reward function, use the MADDPG algorithm to construct a multi-agent reinforcement learning model, train it through a health-weighted priority experience replay strategy, and obtain the optimal Nash equilibrium policy;

[0055] Strategy Decoding and Executable Blueprint Generation Module: Inputs the optimal strategy into the strategy interpreter network, decodes and generates a cross-platform data interaction planning graph, and forms the final executable planning graph through a game theory-based resource conflict coordination mechanism, and decomposes it into local execution subgraphs for each platform;

[0056] The planning graph-driven adaptive secure execution module encrypts the original data on the data source platform using ciphertext policy attribute-based encryption based on the interactive planning graph; deploys interactive context-aware re-encryption agents on edge nodes; and uses a hierarchical homomorphic encryption scheme to achieve cross-platform computation in ciphertext state for the collaborative computing task flow in the planning graph; and establishes a power token chain for each piece of shared data, which is stored on multiple edge nodes through lightweight consensus.

[0057] Quality closed-loop and autonomous evolution module: Collects multi-dimensional quality indicators in real time during the interaction process and calculates the comprehensive quality score. When the comprehensive quality score fails to meet the standard after N consecutive interactions, it locates the root cause through multi-level source tracing analysis and generates strategy optimization guidance signals.

[0058] The beneficial effects of this invention are:

[0059] This invention decomposes time series data into nodes to establish dynamic health expectations that change according to their own operational patterns, making health assessments more accurate. It also introduces the average health of nodes as a negative feedback term into the reward function of reinforcement learning, forcing AI strategies to consider the impact on the health of the underlying infrastructure when optimizing business metrics (latency or cost, etc.). Furthermore, during model training, it prioritizes sampling interaction experiences generated during periods of higher system health, guiding the agent to learn and solidify system-friendly collaborative behaviors. This achieves an optimal dynamic trade-off between single-interaction efficiency and long-term system sustainability, significantly improving the reliability of data interaction services.

[0060] This invention decodes the optimal AI strategy into an executable cross-platform data interaction blueprint, clarifying the data flow, computational tasks, and security requirements. The re-encryption agent synthesizes a single-use valid conversion key based on the real-time interaction context generated by the planning diagram, achieving dynamic and precise authorization of ciphertext. It also establishes an immutable "digital passport" for each piece of data, fully recording its permission changes and usage history throughout its entire lifecycle. This achieves a precise improvement in security strength without sacrificing business flexibility and establishes a traceable and non-repudiable trust foundation.

[0061] This invention enables the system to automatically locate the root cause from three dimensions—data layer, coordination layer, and security layer—when the interaction quality consistently fails to meet standards. Based on the source tracing results, it generates optimization guidance signals and uses elastic weight solidification technology to perform targeted fine-tuning of the core AI model to prevent knowledge forgetting. The optimized model must be verified as effective in a high-fidelity simulation environment before triggering a collaborative update of the entire system's strategy, planning map, and security configuration; thus achieving an efficient and sustainable intelligent ecosystem for data interaction. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0063] Figure 2 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0064] 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.

[0065] Please see Figure 1 As shown, this invention is a cross-platform data interaction and sharing method based on artificial intelligence, comprising the following steps:

[0066] S1: Identify the multiple platforms involved in data interaction and sharing, as well as the data types, scale, and service quality requirements of their interactions. Simultaneously, build a collaborative architecture that includes multiple edge nodes and cloud platforms, construct a dynamic health baseline for each edge node, and calculate the node health in real time using a dynamic weighted algorithm.

[0067] S2: Construct a multivariate reward function that includes latency, bandwidth cost, data freshness and computational load, introduce the average health of nodes as a stability negative feedback to form a modified reward function, use the MADDPG algorithm to construct a multi-agent reinforcement learning model, train it through a health-weighted priority experience replay strategy, and obtain the optimal Nash equilibrium strategy.

[0068] S3: Input the optimal strategy into the strategy interpreter network, decode to generate a cross-platform data interaction planning graph, and form the final executable planning graph through a game theory-based resource conflict coordination mechanism, and decompose it into local execution subgraphs for each platform.

[0069] S4: Based on the interactive planning graph, the original data is encrypted using ciphertext policy attribute base encryption on the data source platform; interactive context-aware re-encryption agents are deployed on edge nodes; a hierarchical homomorphic encryption scheme is used to achieve cross-platform computation in ciphertext state for the collaborative computing task flow in the planning graph; a power token chain is established for each piece of shared data and stored on multiple edge nodes through lightweight consensus.

[0070] S5: Collect multi-dimensional quality indicators in real time during the interaction process and calculate the comprehensive quality score. When the comprehensive quality score fails to meet the standard after N consecutive interactions, locate the root cause through multi-level source tracing analysis and generate strategy optimization guidance signals.

[0071] Specifically, the process begins by clarifying the platforms, data types, and quality requirements involved in the interaction, deploying an edge-cloud architecture, establishing a dynamic health baseline for each edge node, and dynamically calculating node health based on the real-time correlation between metrics and task efficiency. Then, a reward function is constructed targeting latency, cost, freshness, and load, incorporating average node health as a stability negative feedback to form a modified reward function. The MADDPG algorithm is used to model each platform as an agent for training. During training, experience generated during periods of high system health is prioritized to guide the agent in learning stable and friendly collaborative behaviors, ultimately converging to the optimal Nash equilibrium strategy. This optimal strategy is then input into a policy interpreter network to decode it into a cross-platform data interaction planning graph, clarifying data flow paths and resource allocation. If resource conflicts occur across multiple planning graphs, a fast negotiation mechanism based on game theory is used. A temporary allocation agreement is reached, ultimately generating local execution subgraphs for each platform. At the data source, data is encrypted according to the planning graph strategy. When an interaction request is triggered, the re-encryption agent on the path converts the ciphertext into personalized ciphertext that only the requester can decrypt, based on the real-time interaction context. For task flows requiring collaborative computation, homomorphic encryption is used to perform ciphertext computation. Simultaneously, a power token chain is established for each piece of data to record the complete operation history, supporting real-time auditing and revocation of permissions. The quality score of the interaction process is calculated in real-time. When the score consistently fails to meet the standard, multi-layered source analysis is initiated to locate the root cause. An optimization guidance signal is generated based on the root cause, and this signal is used as a sample for targeted incremental optimization of the strategy model. The optimized model needs to be verified in a simulation environment. After confirmation of effectiveness, a collaborative update of the entire system's strategy, planning graph, and security configuration is triggered, enabling the system's autonomous evolution.

[0072] In one embodiment of the present invention, step S1 involves constructing a dynamic health baseline and calculating node health in real time using a dynamic weighted algorithm, including the following steps:

[0073] Collect health data of edge nodes in different time periods over the past 90 days, extract trend and periodic items through time series decomposition algorithm, and generate dynamic health baseline by linear superposition function of the trend and periodic items.

[0074] Define a task execution efficiency index, and calculate the Spearman rank correlation coefficient between each monitoring index and the task execution efficiency index within a 30-day sliding time window; after normalizing the correlation coefficient, obtain the dynamic weight of the index.

[0075] Calculate the real-time health of edge nodes ,in, For the first Dynamic weights of each monitoring indicator To make the first Raw collected values ​​of each monitoring indicator A function that maps negative scores to the interval [0,1].

[0076] Specifically, the system collects performance snapshot data of edge nodes at different times during the business cycle, and uses seasonal time-series decomposition algorithms (such as STL) to separate long-term trend items reflecting hardware aging and software load upgrades from noisy monitoring data. And periodic items that capture daily / weekly business fluctuations. The linear superposition of the two ,in, and These are the combined weighting coefficients for the trend and cycle items, with values ​​set according to the nodes of interest, generating a dynamic health baseline. This baseline is not a fixed threshold, but a dynamically adjusted expected health state according to the evolution of the nodes themselves, which enables the system to distinguish between normal business peaks and abnormal node degradation, providing a time-aware decision basis for subsequent interactive scheduling.

[0077] Define task execution efficiency E as the core indicator (e.g., E = number of successfully processed requests / unit time); within a sliding time window (e.g., the past 30 days), calculate the Spearman rank correlation coefficient between each monitoring indicator (CPU utilization, memory usage, and network latency) and task execution efficiency in real time. The absolute value of this coefficient quantifies the strength of the indicator's influence on task execution efficiency in the current period; after normalizing the absolute value of this coefficient, obtain dynamic weights, so that the health assessment can adapt to the differences in resource sensitivity of different interaction task types, ensuring that the assessment focus always remains consistent with the characteristics of the currently dominant interaction business.

[0078] In one embodiment of the present invention, step S1 further includes an edge node intelligent migration mechanism:

[0079] When the real-time health of the edge node meets When this occurs, the migration assessment process is triggered, in which... This is the preset health decay tolerance coefficient. This is the baseline value of the dynamic health of the edge node at the current moment;

[0080] For each candidate target node, the node's current health, resource availability, migration task profile, and local network state are input into a pre-trained gradient boosting decision tree model to output two predicted values, including predicted health and migration risk score.

[0081] Calculate the overall score for each candidate node. ,in, For the first The current health of each candidate node. For the first Risk score of the local network stability of each candidate node To migrate to the first The communication and synchronization overhead of each candidate node, , and These are the weighting coefficients;

[0082] choose highest and The candidate nodes are selected as the final migration targets.

[0083] Specifically, the preset health decay tolerance coefficient is set to 0.85 in this embodiment; where, , and These are weighting coefficients, allowing the system to adjust decision preferences based on the criticality level of the interactive task (e.g., high real-time tasks pay more attention to weighting coefficients). In this embodiment, The value is 0.3. The value is 0.2. The value is 0.5;

[0084] Among them, the intelligent migration mechanism ensures that when the data interaction carrying node is at risk of failure, it can provide early warning and migration based on dynamic baseline, and complete the switch before the interaction is interrupted, thus protecting the user experience. At the same time, the weights in the comprehensive scoring formula can be dynamically adjusted according to the type of interaction task and SLA (service quality requirements), so that the migration strategy is deeply aligned with business objectives.

[0085] In one embodiment of the present invention, step S2, modifying the reward function, includes the following steps:

[0086] Based on the aforementioned service quality requirements, the four optimization objectives—latency, bandwidth cost, data freshness, and computational load—are quantified into reward components and then weighted and summed. The basic multivariate reward function is as follows: ,in, For negative delayed rewards, As a negative bandwidth cost incentive, To reward positive data freshness, To calculate the load reward negatively, , , and These are the initial weighting coefficients for each reward component;

[0087] Based on the aforementioned fundamental multivariate reward function, the real-time health of edge nodes is introduced as a global stability constraint to construct a modified reward function. ,in, The corrected reward function value. The health impact coefficient. This represents the average real-time health of all edge nodes involved in the current data interaction task.

[0088] Specifically, the system first quantifies the core service quality requirements of cross-platform data interaction into four calculable reward components, forming a basic multivariate reward function; delayed rewards are negative rewards, for example, ,in, For actual delay, For target latency, the lower the actual latency (the better the performance), the smaller its negative value, and the smaller the drag on the total reward; bandwidth cost reward is a negative reward, for example, ,in, For actual bandwidth used, It is the unit cost; data freshness reward is a positive reward, for example, ,in, The reward value is calculated based on the time from data generation to use; the fresher the data, the higher the reward value, encouraging rapid data transfer. Computational load rewards are negative rewards, for example... ,in, For node CPU utilization, the squared term is used to penalize high loads and encourage load balancing. , , and The values ​​are set to 0.2, 0.2, 0.3 and 0.3 respectively; the basic multivariate reward function guides the agent to learn how to efficiently complete a single data interaction;

[0089] in, The health impact coefficient is a constant greater than zero, used to control the strength of the influence of stability constraints on the overall reward. In this embodiment, The value is set to 0.5;

[0090] in, It is a global, aggregate metric; the behavior of all agents participating in the same interactive task, regardless of their individual actions, affects... The effects will all be through The function provides feedback equally to all agents, incentivizing them to coordinate their actions spontaneously and jointly maintain the health of the shared infrastructure they depend on, thereby giving rise to cooperative interaction strategies that are conducive to the long-term stable operation of the system.

[0091] In one embodiment of the present invention, step S2 involves constructing a multi-agent reinforcement learning model using the MADDPG algorithm and training it through a health-weighted priority experience replay strategy, including the following steps:

[0092] The multi-agent reinforcement learning model constructs an actor network and a centralized critic network for each platform agent. Each actor network takes into account the local observation vector of the corresponding platform and outputs the action policy of the platform. The centralized critic network takes into account the concatenation of the local observation vectors and action vectors of all platforms and outputs a joint Q-value estimate of the current global state-joint action.

[0093] At each training time step, each agent generates an experience tuple based on its action strategy output by its actor network and its interaction with the environment. Stored in a globally shared experience replay pool; where, For local state observation, The action performed by the intelligent agent. To perform the action The environment then transitions to the next state observation vector;

[0094] When sampling in the experience replay pool for neural network parameter updates, a health-based priority sampling strategy is adopted, specifically: a corresponding sampling priority is calculated for each experience, specifically: ,in, For the first The time difference error of the empirical rule, This is the health bias coefficient. For the first The average health of relevant edge nodes during the experience period;

[0095] The policy network and value network of each agent are based on small batches of sampled experience data to maximize the cumulative correction reward for the future. The predicted value is the objective, and it is updated using the policy gradient method;

[0096] After training, the policy networks of all agents will converge collaboratively to a steady-state Nash equilibrium policy, which is the optimal policy.

[0097] Specifically, the system constructs an actor network for each platform agent participating in data interaction. The input of this network is the agent's local observation vector, which encapsulates its private interaction context, such as the state of the data queue to be sent, the list of available neighbor nodes and their health and estimated link quality. The output of the actor network is the action strategy that the agent should take in the current state, such as which target platform to choose, how much bandwidth to request, and what priority to set. At the same time, the system constructs a centralized commentator network. The input of this network is the concatenation of all agents' local observations and all agents' actions. Its output is a joint Q-value estimate of the current "global state-joint action" pair. This design enables the commentator to have a global perspective and accurately assess the long-term value of any set of cooperative actions.

[0098] in, The health bias coefficient is a positive-zero hyperparameter used to control the strength of the influence of health on sampling priority; in the early stages of training, the system can statistically analyze the experience pool. Typical ranges (e.g., mean and standard deviation); mean, mean, and standard deviation; mean, mean, and standard deviation. The range is between [0, 1]; assuming initial statistics The mean μ_δ ≈ 0.5, and the standard deviation σ_δ ≈ 0.3. To ensure that the maximum impact of health status is comparable to the typical error impact, the mean μ_δ ≈ 0.5, and the standard deviation σ_δ ≈ 0.3 ... standard deviation σ_δ ≈ 0.3. The initial value is set in the range of 0.1 to 1.0; for example, set... The value is 0.5, ensuring that the health factor has a significant but not overwhelming weight in the sampling decision.

[0099] Health-weighted priority experience replay is an intelligent bridge connecting the stability-aware reward function and the final collaborative strategy. By influencing the learning process itself, it is deeply internalized into the decision-making model of each agent, enabling the entire system to learn how to operate in the "best times" in history to guide how to conduct better data interaction in the future. It is the core training guarantee for realizing intelligent and sustainable cross-platform collaboration.

[0100] In one embodiment of the present invention, step S3, the resource conflict coordination mechanism, includes the following steps:

[0101] When multiple interactive planning graphs conflict over the same resource, the relevant agents conduct multiple rounds of bidding negotiations based on the task's criticality level and the resources they already hold.

[0102] A temporary Nash equilibrium for resource allocation is formed through a limited number of rounds of game playing;

[0103] The interaction planning diagram of each agent is dynamically updated based on the negotiation results.

[0104] Specifically, when the strategy decoding and executable blueprint generation modules generate their respective interaction planning diagrams for multiple concurrent data interaction tasks, the system will automatically perform resource demand overlap analysis. If it finds that two or more planning diagrams make mutually exclusive requests for the same scarce resource (e.g., a specific computing unit of an edge node, or reserved bandwidth of a high-bandwidth link) within the same time window, it is determined to be a resource conflict. The system will immediately send a negotiation request to all agents involved in the conflict (each agent represents an interaction task) and attach the identifier of the conflicting resource.

[0105] Each participating agent will make a bid based on two core attributes of its task, including the task's criticality level. (This is a preset or dynamically calculated priority, for example) =1, Real-time video analysis, millisecond-level latency requirement; =2, batch log synchronization, minute-level latency requirement; =3, Non-emergency data backup, hourly latency requirement) and existing resources (Intelligent agent) The total value of other resources already secured through its planning map in this interaction (quantifiable as a weighted sum of computing power and bandwidth) reflects the "bargaining chips" already acquired for the task; bid. Calculated by the agent based on its internal policy network, typically with Positive correlation, and may be considered To achieve strategic balancing, a modified Vickrey auction (VCG) mechanism is used for negotiation. In each round: each agent submits its sealed bid; a lightweight consensus module (which can be rotated among participating nodes) calculates the highest bidder as the temporary winner; the winner's identity and bid are announced, but other bids are not disclosed; to avoid delays caused by infinite negotiation, the system presets a maximum negotiation round K (e.g., 3-5 rounds). After each round, agents that have not won resources can increase their bids, modify their alternatives, or withdraw from the competition; after a finite number of rounds of game, the system will reach a temporary Nash equilibrium: no agent can obtain a better benefit (i.e., obtain resources at a lower cost, or obtain resources to a greater extent to improve the overall efficiency of the task) by unilaterally changing its strategy (bid or alternative). At this point, the resources are allocated to the agent with the highest bid (or the best overall benefit) in the current equilibrium state.

[0106] After the negotiation results are generated: The winning agent: its demand for conflicting resources in its interaction planning graph is confirmed, and it sends a final locking command to the resource manager; Other agents: must dynamically and in real time update their local interaction planning graphs according to the alternatives they committed to in the negotiation; for example, migrating computing tasks from conflicting edge nodes to alternative nodes in their planning graphs, or adjusting data transmission paths. The updated planning graphs need to pass a consistency check again.

[0107] In one embodiment of the present invention, step S4, deploying an interaction context-aware re-encryption proxy at the edge node, includes the following steps:

[0108] When the re-encryption proxy receives a data request, it queries the real-time interaction context associated with this request;

[0109] Session-limiting parameters are generated based on the real-time interaction context, and a single-use valid re-encryption conversion key is synthesized using a lattice-based cryptographic algorithm.

[0110] The initial ciphertext is transformed using the re-encryption conversion key to generate a personalized re-encryption ciphertext;

[0111] The encrypted digest of the real-time interaction context is embedded in the generated ciphertext metadata to form a traceable interaction binding ciphertext.

[0112] Submit this conversion event to the Power Token Chain for evidence storage.

[0113] Specifically, when the re-encryption agent located at the edge node receives a data request from the requesting platform, its primary action is not to directly perform re-encryption, but to initiate a query to the interaction state manager in the system. The query content includes the data resource identifier and the requester's identity. The agent uses this information to query the real-time interaction context currently associated with this request. The context content is a structured information tuple, including at least: an interaction graph identifier (Graph_ID): the global interaction graph to which this request belongs; a task phase code (Phase_Code): identifying the specific phase of this request within the overall graph process (e.g., "data prefetching," "collaborative computation input," "result feedback"); a list of authorizing platforms: the set of target platforms explicitly authorized to access this data in the graph; and a usage constraint hash: an encrypted digest of the "usage intent policy" (e.g., "only allow aggregation computation," "prohibit secondary distribution") defined for this data in the graph. This ensures the legitimacy verification of the prerequisites for secure execution.

[0114] After successful verification, the re-encryption proxy performs the following core operations: Generating a session constraint parameter: Based on the queried interaction context, the proxy calculates a session constraint parameter, for example, SBP = Hash(Graph_ID || Phase_Code || Timestamp || Nonce), which uniquely identifies the specific interaction session instance; Synthesizing a single-use valid key: The proxy uses a lattice-based cryptographic algorithm (based on the RLWE problem) to synthesize the long-term re-encryption key factor pre-distributed by the data owner with the newly generated session constraint parameter SBP, generating a single-use valid re-encryption conversion key; Performing ciphertext conversion: Using this OT-RK, the proxy converts the initial ciphertext generated by the data owner using attribute-based encryption, generating a personalized re-encrypted ciphertext, which can only be decrypted by the current requester's private key.

[0115] To ensure the auditability and non-repudiation of the security process: Interaction Binding Ciphertext: In the metadata area of ​​the generated re-encrypted ciphertext, the agent uses the requester's public key to encrypt a cryptographic digest embedded in the interaction context. This makes the ciphertext itself "carry" the genes of its respective interaction task. After the receiver decrypts it, it can verify this metadata to ensure that the data obtained is exactly what its current interaction task expected, realizing end-to-end usage compliance verification; Power Token Chain Evidence Storage: The agent generates a record for the key information of this conversion event (including: data ID, requester ID, Graph_ID, Phase_Code, SBP hash, timestamp, and agent node ID) and submits it to the power token chain. After lightweight consensus, this record is permanently stored as tamper-proof evidence, providing complete audit traceability for data flow.

[0116] In one embodiment of the present invention, step S4 of the power token chain includes the following steps:

[0117] Record the complete history of data changes in permissions from its generation to each access, calculation, and transfer.

[0118] When abnormal platform behavior or authorization expiration is detected, the relevant permissions are invalidated through a consensus protocol.

[0119] Supports fine-grained access auditing and accountability based on token chain.

[0120] Specifically, when a piece of data is created on the source platform and ready to be shared, the system generates a root power token for that data and anchors it to the power token chain. This root token is a structured record that includes at least: a data fingerprint (a cryptographic hash of the data, such as SHA-256), serving as a unique identifier for the data; owner information (the identity identifier of the data source platform); an initial policy hash (a cryptographic digest of the "use intent policy" defined for this data in the interaction planning diagram of step S3, such as "allowing platforms A and B to use it for computation C within time period T"); and a creation timestamp and transaction hash (recording the creation event and linking it to the blockchain).

[0121] Throughout the entire data interaction lifecycle, any critical event that leads to a change in the right to use the data is added to the chain as a new power token, forming a chronologically ordered token chain. Key change events include: Access Authorization: When an interaction context-aware re-encryption agent successfully performs a transformation, granting access to the data to a requester, an access token is generated; Computation Authorization: When data is used as input in a collaborative computation task defined in a planning graph, a computation token is generated; and Flow Authorization: When data is processed by one platform and then flows to the next platform as new data or an intermediate result, a flow token is generated, establishing a clear data lineage. Each new token contains the hash value of the previous token, forming a strict chain structure to ensure the immutability of historical records.

[0122] The power token chain is not only a recording system but also an execution system. When the system (through the anomaly detection module in step S5) detects the following situations, it will trigger the permission revocation process: abnormal platform behavior (e.g., a platform is detected to be performing unauthorized large-scale data downloads or launching attacks); authorization expiration (the data usage time limit defined in the interaction planning diagram has expired); and policy violation (an audit reveals that a data usage violated the initial policy, such as using data limited to "aggregate calculation" for "individual profiling").

[0123] The revocation process is executed among multiple trusted edge nodes storing the power token chain copy via a lightweight consensus protocol (such as RAFT or PoA-based consensus): the module or supervisory node that detects the anomaly initiates a revocation proposal, specifying the particular power token to be invalidated (or all tokens associated with the anomaly platform); the nodes participating in the consensus verify the validity of the proposal (e.g., checking for evidence of the anomaly, whether the authorization has expired); after consensus is reached, a global invalidation instruction is issued on the chain. Thereafter, any re-encryption agent or computing node will refuse service when it receives a data request involving the invalidated token; the invalidation instruction itself is recorded on the chain as a special revocation token, completing the audit loop.

[0124] The Power Token Chain provides unprecedented transparency and accuracy for post-audit, including full-view traceability, precise accountability, and compliance proof; the Power Token Chain becomes a trusted data hub connecting AI-driven interactive intents, context-aware secure execution, and full lifecycle compliance governance.

[0125] In one embodiment of the present invention, step S5 includes the following steps:

[0126] During each cross-platform data interaction, data utility indicators, collaboration efficiency indicators, and strategy compliance indicators are collected. These indicators are standardized and weighted and integrated to generate a comprehensive quality score for the data interaction.

[0127] The comprehensive quality score is compared with the expected value of the service quality requirement to calculate the quality gap; if the quality gap is ≤ the preset quality threshold in N consecutive interaction cycles, a systematic source analysis is initiated to locate the root cause of the quality gap from three dimensions: data content layer, interaction coordination layer, and security execution layer.

[0128] Based on the source tracing results, generate strategy optimization guidance signals;

[0129] The strategy optimization guidance signal and the corresponding complete state trajectory of the interaction process are jointly encapsulated into enhanced training samples. The elastic weight solidification technology is used to perform targeted incremental optimization on the multi-agent reinforcement learning model. During the optimization process, the corresponding guiding reward item is introduced or strengthened in the reward function according to the optimization direction indicated by the strategy optimization guidance signal.

[0130] When the improvement of the simulation interaction quality evaluation value compared with the baseline value obtained by the model before optimization in the same simulation environment is greater than the set improvement threshold, the optimization is deemed effective and the coordinated evolution of the entire system strategy is triggered.

[0131] The simulation interaction quality evaluation value is obtained by collecting various indicators, standardizing and weighting the results after fully executing the strategy generated by the optimized multi-agent reinforcement learning model in a high-fidelity simulation interaction environment.

[0132] Specifically, after each data interaction task is completed, the system automatically collects three types of core indicators: data utility indicators (used to quantify the value of data usage, such as the success rate of data being referenced by downstream tasks); collaboration efficiency indicators (used to quantify the smoothness and economy of the interaction process, such as end-to-end task completion time, average idle rate of resources on each platform, or bandwidth and computing costs, etc.); and strategy compliance indicators (used to quantify the deviation between actual execution and AI planning, such as the matching degree between the actual data flow path and the interaction planning diagram, the consistency between the execution of security measures and planning requirements, etc.). After Z-score standardization to eliminate the influence of dimensions, all indicators are weighted and fused according to the business objectives to calculate the comprehensive quality score of this interaction, with the score ranging from 0 to 100 points.

[0133] The system compares the calculated comprehensive quality score with the expected value of the service quality requirements defined when the task was initiated to obtain the quality gap. The system sets an observation window (N consecutive interactions) and a preset quality threshold. Only when the quality gap exceeds the preset quality threshold in N consecutive interactions will the system determine that there is a systemic quality problem rather than an accidental fluctuation, and start the source analysis engine.

[0134] The preset quality threshold is obtained by adding the average quality difference to twice the standard deviation;

[0135] The tracing engine performs in-depth analysis from three levels: Data Content Layer: It checks whether semantic distortion, format errors, or integrity loss occur during data flow by comparing the feature vector hashes of the data sender and receiver, or verifying the integrity proof in homomorphic encryption computation; Interaction Coordination Layer: It analyzes whether there are design flaws in the interaction planning diagram, for example, checking whether the resource allocation in the plan has caused persistent bottlenecks or conflicts under real network conditions; it analyzes whether the game coordination mechanism has achieved a suboptimal equilibrium in a specific scenario; Security Execution Layer: It examines the power token chain and re-encrypted proxy logs to determine whether high-strength security operations (such as fully homomorphic encryption) have introduced unacceptable latency overhead, or whether the policy is too strict, causing legitimate requests to be mistakenly blocked.

[0136] Based on the root cause identified, the system generates a policy optimization guidance signal. This signal is a structured instruction, such as: {Optimization objective: reduce latency, Problematic stage: the path from edge node X to Y, Suggested action: strengthen the penalty for the health of node X in the reward function}. The system then encapsulates the guidance signal and the complete interaction process state trajectory leading to the quality problem (including all agents' observations, action sequences, and environmental states) into an enhanced training sample, labeling it as "needs improvement". The system employs elastic weight solidification technology to fine-tune the original multi-agent reinforcement learning model. The EWC algorithm calculates the importance of each parameter in the model to the learned knowledge. During fine-tuning, "resistance" is applied to important parameters to prevent new knowledge from overwriting (forgetting) old, still effective policies, thus achieving precise improvement without compromising existing capabilities. During optimization, the guidance signal is used to dynamically adjust the reward function, for example, temporarily increasing the reward or penalty weights for indicators related to the problem stage, guiding the model to explore new policy directions.

[0137] The optimized model is not directly deployed to the production environment. Instead, it is first placed in a high-fidelity simulation environment that replicates the network topology, node performance model, and historical traffic patterns of the real system. The optimized model runs multiple complete interactive scenarios in this environment. The system collects the same quality metrics during simulation, calculates the simulation interaction quality evaluation value, and compares it with the baseline value obtained by the unoptimized model in the same simulation environment. The optimization is considered effective only when the improvement exceeds a set improvement threshold (in this embodiment, the improvement threshold is set to 5%). After verification, the system triggers a system-wide collaborative evolution, including: policy update: replacing the original model with the optimized multi-agent model; planning graph generator update: the policy interpreter network uses the new model to generate a new interactive planning graph; and security policy synchronization: updating the context verification rules of the re-encryption proxies of each edge node according to the intent implied in the new planning graph.

[0138] Please see Figure 2 As shown, this invention is a cross-platform data interaction and sharing system based on artificial intelligence, comprising the following modules:

[0139] Collaborative architecture initialization and health awareness module: Identify the multiple platforms involved in data interaction and sharing, as well as the data types, scale, and service quality requirements of their interactions. Simultaneously, build a collaborative architecture that includes multiple edge nodes and cloud platforms, construct a dynamic health baseline for each edge node, and calculate the node health in real time through a dynamic weighted algorithm.

[0140] Multi-agent cooperative policy learning module: Construct a multivariate reward function that includes latency, bandwidth cost, data freshness and computational load, introduce the average health of nodes as a stability negative feedback to form a modified reward function, use the MADDPG algorithm to construct a multi-agent reinforcement learning model, train it through a health-weighted priority experience replay strategy, and obtain the optimal Nash equilibrium policy;

[0141] Strategy Decoding and Executable Blueprint Generation Module: Inputs the optimal strategy into the strategy interpreter network, decodes and generates a cross-platform data interaction planning graph, and forms the final executable planning graph through a game theory-based resource conflict coordination mechanism, and decomposes it into local execution subgraphs for each platform;

[0142] The planning graph-driven adaptive secure execution module encrypts the original data on the data source platform using ciphertext policy attribute-based encryption based on the interactive planning graph; deploys interactive context-aware re-encryption agents on edge nodes; and uses a hierarchical homomorphic encryption scheme to achieve cross-platform computation in ciphertext state for the collaborative computing task flow in the planning graph; and establishes a power token chain for each piece of shared data, which is stored on multiple edge nodes through lightweight consensus.

[0143] Quality closed-loop and autonomous evolution module: Collects multi-dimensional quality indicators in real time during the interaction process and calculates the comprehensive quality score. When the comprehensive quality score fails to meet the standard after N consecutive interactions, it locates the root cause through multi-level source tracing analysis and generates strategy optimization guidance signals.

[0144] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A cross-platform data interaction and sharing method based on artificial intelligence, characterized in that, Includes the following steps: S1: Determine the multiple platforms involved in data interaction and sharing, as well as the data types, scale, and service quality requirements of their interactions. At the same time, build a collaborative architecture that includes multiple edge nodes and cloud platforms, construct a dynamic health baseline for each edge node, and calculate the node health in real time through a dynamic weighted algorithm. S2: Construct a multivariate reward function that includes latency, bandwidth cost, data freshness and computational load, introduce the average health of nodes as a stability negative feedback to form a modified reward function, use the MADDPG algorithm to construct a multi-agent reinforcement learning model, train it through a health-weighted priority experience replay strategy, and obtain the optimal Nash equilibrium strategy. S3: Input the optimal strategy into the strategy interpreter network, decode to generate a cross-platform data interaction planning graph, and form the final executable planning graph through a game theory-based resource conflict coordination mechanism, and decompose it into local execution subgraphs for each platform. S4: Based on the interaction planning graph, the original data is encrypted using the ciphertext policy attribute base encryption on the data source platform; an interaction context-aware re-encryption proxy is deployed on the edge node; For the collaborative computing task flow in the planning diagram, a hierarchical homomorphic encryption scheme is adopted to realize cross-platform computing in ciphertext state; a power token chain is established for each piece of shared data, and stored on multiple edge nodes through lightweight consensus; S5: Collect multi-dimensional quality indicators in real time during the interaction process and calculate the comprehensive quality score. When the comprehensive quality score fails to meet the standard for N consecutive interactions, locate the root cause through multi-level source tracing analysis and generate strategy optimization guidance signals. In step S1, a dynamic health baseline is constructed, and the node health is calculated in real time using a dynamic weighted algorithm, including the following steps: Collect health data of edge nodes in different time periods over the past 90 days, extract trend and periodic items through time series decomposition algorithm, and generate dynamic health baseline by linear superposition function of the trend and periodic items. Define a task execution efficiency index, and calculate the Spearman rank correlation coefficient between each monitoring index and the task execution efficiency index within a 30-day sliding time window; after normalizing the correlation coefficient, obtain the dynamic weight of the index. Calculate the real-time health of edge nodes ,in, For the first Dynamic weights of each monitoring indicator To make the first Raw collected values ​​of each monitoring indicator A function that maps negative scores to the interval [0,1].

2. The cross-platform data interaction and sharing method based on artificial intelligence according to claim 1, characterized in that, It also includes an intelligent migration mechanism for edge nodes: When the real-time health of the edge node meets When this occurs, the migration assessment process is triggered, in which... This is the preset health decay tolerance coefficient. This is the baseline value of the dynamic health of the edge node at the current moment; For each candidate target node, the node's current health, resource availability, migration task profile, and local network state are input into a pre-trained gradient boosting decision tree model to output two predicted values, including predicted health and migration risk score. Calculate the overall score for each candidate node. ,in, For the first The current health of each candidate node. For the first Risk score of the local network stability of each candidate node To migrate to the first The communication and synchronization overhead of each candidate node, , and These are the weighting coefficients; choose highest and The candidate nodes are selected as the final migration targets.

3. The cross-platform data interaction and sharing method based on artificial intelligence according to claim 1, characterized in that, In step S2, the reward function is modified, including the following steps: Based on the aforementioned service quality requirements, the four optimization objectives—latency, bandwidth cost, data freshness, and computational load—are quantified into reward components and then weighted and summed. The basic multivariate reward function is as follows: ,in, For negative delayed rewards, As a negative bandwidth cost incentive, To reward positive data freshness, To calculate the load reward negatively, , , and These are the initial weighting coefficients for each reward component; Based on the aforementioned fundamental multivariate reward function, the real-time health of edge nodes is introduced as a global stability constraint to construct a modified reward function. ,in, The corrected reward function value. The health impact coefficient. This represents the average real-time health of all edge nodes involved in the current data interaction task.

4. The cross-platform data interaction and sharing method based on artificial intelligence according to claim 1, characterized in that, In step S2, a multi-agent reinforcement learning model is constructed using the MADDPG algorithm and trained using a health-weighted priority experience replay strategy, including the following steps; The multi-agent reinforcement learning model constructs an actor network and a centralized critic network for each platform agent. Each actor network takes into account the local observation vector of the corresponding platform and outputs the action policy of the platform. The centralized critic network takes into account the concatenation of the local observation vectors and action vectors of all platforms and outputs a joint Q-value estimate of the current global state-joint action. At each training time step, each agent generates an experience tuple based on its action strategy output by its actor network and its interaction with the environment. Stored in a globally shared experience replay pool; where, For local state observation, The action performed by the intelligent agent. To perform the action The environment then transitions to the next state observation vector; When sampling in the experience replay pool for neural network parameter updates, a health-based priority sampling strategy is adopted, specifically: a corresponding sampling priority is calculated for each experience, specifically: ,in, For the first The time difference error of the empirical rule, This is the health bias coefficient. For the first The average health of relevant edge nodes during the experience period; The policy network and value network of each agent are based on small batches of sampled experience data to maximize the cumulative correction reward for the future. The predicted value is the objective, and it is updated using the policy gradient method; After training, the policy networks of all agents will converge collaboratively to a steady-state Nash equilibrium policy, which is the optimal policy.

5. The cross-platform data interaction and sharing method based on artificial intelligence according to claim 1, characterized in that, In step S3, the resource conflict coordination mechanism includes the following steps: When multiple interactive planning graphs conflict over the same resource, the relevant agents conduct multiple rounds of bidding negotiations based on the task's criticality level and the resources they already hold. A temporary Nash equilibrium for resource allocation is formed through a limited number of rounds of game playing; The interaction planning diagram of each agent is dynamically updated based on the negotiation results.

6. The cross-platform data interaction and sharing method based on artificial intelligence according to claim 1, characterized in that, In step S4, deploying an interaction context-aware re-encryption proxy at the edge node includes the following steps: When the re-encryption proxy receives a data request, it queries the real-time interaction context associated with this request; Session-limiting parameters are generated based on the real-time interaction context, and a single-use valid re-encryption conversion key is synthesized using a lattice-based cryptographic algorithm. The initial ciphertext is transformed using the re-encryption conversion key to generate a personalized re-encryption ciphertext; The encrypted digest of the real-time interaction context is embedded in the generated ciphertext metadata to form a traceable interaction binding ciphertext. Submit this conversion event to the Power Token Chain for evidence storage.

7. The cross-platform data interaction and sharing method based on artificial intelligence according to claim 1, characterized in that, In step S4, the power token chain includes the following steps: Record the complete history of data changes in permissions from its generation to each access, calculation, and transfer. When abnormal platform behavior or authorization expiration is detected, the relevant permissions are invalidated through a consensus protocol. Supports fine-grained access auditing and accountability based on token chain.

8. The cross-platform data interaction and sharing method based on artificial intelligence according to claim 1, characterized in that, Step S5 includes the following steps: During each cross-platform data interaction, data utility indicators, collaboration efficiency indicators, and strategy compliance indicators are collected. These indicators are standardized and weighted and integrated to generate a comprehensive quality score for the data interaction. The comprehensive quality score is compared with the expected value of the service quality requirement to calculate the quality gap; if the quality gap is ≤ the preset quality threshold in N consecutive interaction cycles, a systematic source analysis is initiated to locate the root cause of the quality gap from three dimensions: data content layer, interaction coordination layer, and security execution layer. Based on the source tracing results, generate strategy optimization guidance signals; The strategy optimization guidance signal and the corresponding complete state trajectory of the interaction process are jointly encapsulated into enhanced training samples. The elastic weight solidification technology is used to perform targeted incremental optimization on the multi-agent reinforcement learning model. During the optimization process, the corresponding guiding reward item is introduced or strengthened in the reward function according to the optimization direction indicated by the strategy optimization guidance signal. When the improvement of the simulation interaction quality evaluation value compared with the baseline value obtained by the model before optimization in the same simulation environment is greater than the set improvement threshold, the optimization is deemed effective and the coordinated evolution of the entire system strategy is triggered. The simulation interaction quality evaluation value is obtained by collecting various indicators, standardizing and weighting the results after fully executing the strategy generated by the optimized multi-agent reinforcement learning model in a high-fidelity simulation interaction environment.

9. A cross-platform data interaction and sharing system based on artificial intelligence, used to implement the cross-platform data interaction and sharing method based on artificial intelligence as described in any one of claims 1-8, characterized in that, Includes the following modules: Collaborative architecture initialization and health awareness module: Identify the multiple platforms involved in data interaction and sharing, as well as the data types, scale, and service quality requirements of their interactions. Simultaneously, build a collaborative architecture that includes multiple edge nodes and cloud platforms, construct a dynamic health baseline for each edge node, and calculate the node health in real time through a dynamic weighted algorithm. Multi-agent cooperative policy learning module: Construct a multivariate reward function that includes latency, bandwidth cost, data freshness and computational load, introduce the average health of nodes as a stability negative feedback to form a modified reward function, use the MADDPG algorithm to construct a multi-agent reinforcement learning model, train it through a health-weighted priority experience replay strategy, and obtain the optimal Nash equilibrium policy; Strategy Decoding and Executable Blueprint Generation Module: Inputs the optimal strategy into the strategy interpreter network, decodes and generates a cross-platform data interaction planning graph, and forms the final executable planning graph through a game theory-based resource conflict coordination mechanism, and decomposes it into local execution subgraphs for each platform; The planning graph-driven adaptive security execution module encrypts the original data on the data source platform using ciphertext policy attribute-based encryption based on the interactive planning graph; and deploys an interactive context-aware re-encryption proxy on the edge nodes. For the collaborative computing task flow in the planning diagram, a hierarchical homomorphic encryption scheme is adopted to realize cross-platform computing in ciphertext state; a power token chain is established for each piece of shared data, and stored on multiple edge nodes through lightweight consensus; Quality closed-loop and autonomous evolution module: Collects multi-dimensional quality indicators in real time during the interaction process and calculates the comprehensive quality score. When the comprehensive quality score fails to meet the standard after N consecutive interactions, it locates the root cause through multi-level source tracing analysis and generates strategy optimization guidance signals.

Images