A federated trusted management platform based on cross-domain information sharing

Abstract

The application discloses a kind of federal trusted management platform based on cross-domain intelligence sharing, it is related to information security technical field, including: federal trusted management platform is connected with multiple cross-domain intelligence systems;Fusion security layer is based on TIPP algorithm and is carried out privacy protection processing to the multi-source cross-domain intelligence data of multiple cross-domain intelligence systems accessed;Federal learning scheduling layer carries out federal learning modeling task scheduling according to the multi-source cross-domain intelligence data after privacy protection processing, and agent management layer carries out resource allocation to federal learning modeling task in federal learning scheduling layer by AACA adaptive cooperation mechanism, obtains the federal learning global model for sharing.This application solves the technical problem that multi-source cross-domain intelligence faces multi-level privacy leakage risk in federal learning process in prior art, achieves the technical effect that cross-domain intelligence is trusted shared and jointly modeled while ensuring data privacy security.

Description

Technical Field

[0001] This invention relates to the field of information security technology, specifically to a federated trusted management platform based on cross-domain intelligence sharing. Background Technology

[0002] In cross-domain intelligence sharing, all participating parties need to achieve joint modeling while maintaining local data storage. While federated learning, as a key technology, avoids centralized transmission of raw data, privacy risks still exist in multiple stages such as training, computation, and communication. Attackers could potentially reverse-engineer the original data through model updates, steal intermediate computation results, or eavesdrop on communications to obtain sensitive information. As data dimensionality increases and the number of participating nodes grows, the risk of privacy breaches multiplies at different levels, making comprehensive protection difficult with a single technology. This has become a key obstacle hindering the implementation of cross-domain federated intelligent collaboration. Summary of the Invention

[0003] This application provides a federated trust management platform based on cross-domain intelligence sharing, which addresses the technical problem of multi-layered privacy leakage risks faced by multi-source cross-domain intelligence in the federated learning process in existing technologies.

[0004] In view of the above problems, this application provides a federal trust management platform based on cross-domain intelligence sharing.

[0005] This application provides a federal trust management platform based on cross-domain intelligence sharing, the platform comprising:

[0006] The federated trust management platform connects to multiple cross-domain intelligence systems. The platform includes a primary core architecture comprising a fusion security layer, a federated learning scheduling layer, and an agent management layer. The fusion security layer performs privacy protection processing on the multi-source cross-domain intelligence data from the multiple connected cross-domain intelligence systems based on the TIPP algorithm. This fusion security layer includes a differential privacy module, a homomorphic encryption module, and a zero-knowledge proof module. The federated learning scheduling layer schedules federated learning modeling tasks according to the privacy-protected multi-source cross-domain intelligence data. The agent management layer allocates resources for the federated learning modeling tasks within the federated learning scheduling layer through the AACA adaptive collaboration mechanism, obtaining a shared global federated learning model.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] The federated trusted management platform described in this application connects to multiple cross-domain intelligence systems. The federated trusted management platform includes a primary core architecture comprising a fusion security layer, a federated learning scheduling layer, and an agent management layer. The fusion security layer performs privacy protection processing on the multi-source cross-domain intelligence data from the multiple connected cross-domain intelligence systems based on the TIPP algorithm. This fusion security layer includes a differential privacy module, a homomorphic encryption module, and a zero-knowledge proof module. The federated learning scheduling layer schedules federated learning modeling tasks according to the privacy-protected multi-source cross-domain intelligence data. The agent management layer allocates resources for the federated learning modeling tasks within the federated learning scheduling layer through an AACA adaptive collaboration mechanism, obtaining a shared global federated learning model. This invention addresses the technical problem of multi-layered privacy leakage risks faced by multi-source cross-domain intelligence in the federated learning process in existing technologies. By introducing the TIPP triple fusion privacy protection mechanism into the fusion security layer and collaboratively integrating differential privacy, homomorphic encryption, and zero-knowledge proof modules, it achieves the technical effect of trusted sharing and joint modeling of cross-domain intelligence while ensuring data privacy and security. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of a federal trusted management platform structure based on cross-domain intelligence sharing, provided for an embodiment of this application. Detailed Implementation

[0011] This application provides a federated trusted management platform based on cross-domain intelligence sharing. It addresses the technical problem of multi-layered privacy leakage risks faced by multi-source cross-domain intelligence in federated learning in existing technologies. By introducing the TIPP triple fusion privacy protection mechanism into the fusion security layer, and coordinating differential privacy, homomorphic encryption and zero-knowledge proof modules, it achieves the technical effect of trusted sharing and joint modeling of cross-domain intelligence while ensuring data privacy and security.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, platform, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown in the embodiment of this application, a federal trust management platform based on cross-domain intelligence sharing is provided. The platform includes:

[0015] The federated trust management platform is connected to multiple cross-domain intelligence systems. The federated trust management platform includes a primary core architecture, which includes a converged security layer, a federated learning scheduling layer, and an agent management layer.

[0016] In this embodiment, the Federal Trust Management Platform establishes connections with multiple cross-domain intelligence systems to enable access to and unified processing of multi-source cross-domain intelligence data distributed across different organizations, networks, or trust domains.

[0017] The Federated Trust Management Platform employs a primary core architecture, functionally divided into three layers: the Fusion Security Layer, the Federated Learning Scheduling Layer, and the Agent Management Layer. The Fusion Security Layer, a key component for privacy protection, utilizes the TIPP triple fusion privacy protection mechanism, integrating differential privacy, homomorphic encryption, and zero-knowledge proof modules. It prevents multi-layered privacy leaks during modeling by perturbing model parameters with noise, encrypting the data, and verifying computational correctness. The Federated Learning Scheduling Layer performs federated modeling tasks on the encrypted data output from the Fusion Security Layer. Based on the FedAvg+ adaptive aggregation algorithm, it dynamically integrates the weights of multi-node training results, improving model accuracy and convergence speed in non-independent, identically distributed environments. The Agent Management Layer uses the AACA adaptive collaboration mechanism to perceive the status information of each node in terms of computing resources, data quality, and network status. Based on the MADDPG reinforcement learning strategy, it assigns agent roles and schedules resources, ensuring efficient collaborative operation of the entire system in heterogeneous environments.

[0018] Furthermore, the platform provided in the application embodiments also includes:

[0019] The primary core architecture also includes a trusted computing layer, through which a trusted execution environment is constructed, and the federated learning scheduling layer executes federated learning modeling tasks in the trusted execution environment.

[0020] In this embodiment, within the primary core architecture, the federated trust management platform enhances the execution trustworthiness of the modeling process by introducing a trusted computing layer. The trusted computing layer employs hardware security technologies such as Intel SGX, ARM TrustZone, or AMD SEV to construct a protected Trusted Execution Environment (TEE). This environment provides isolation protection for code and data execution, preventing the operating system or other applications from accessing its internal runtime content. The platform integrates a remote attestation mechanism, utilizing a hardware root of trust to generate a trust report, proving the integrity and authenticity of the current execution environment, enabling other participants to verify that their computing environment has not been tampered with.

[0021] Within the Trusted Execution Environment (TEE), the Federated Learning Scheduling Layer initiates model training tasks. All related computational operations are completed within the TEE, including local model training, gradient calculation, and update operations, ensuring that intermediate computation results and parameters are not leaked or maliciously manipulated. To enhance privacy protection, the Trusted Computing Layer is also configured with a Privacy Budget Manager, which globally controls and dynamically allocates the differential privacy budget available in each training round, avoiding reuse or excessive consumption of privacy resources. This manager works in conjunction with the differential privacy module in the fusion security layer to ensure that the noise addition strategy maintains training performance while ensuring security.

[0022] Through the above steps, the federated trust management platform constructs a TEE through the trusted computing layer, combines a remote proof mechanism to ensure the trustworthiness of the execution environment, and completes the training task of the federated learning scheduling layer in this environment, thereby achieving secure execution, verifiability and privacy isolation of the modeling process.

[0023] Furthermore, the platform provided in the application embodiments also includes:

[0024] The federal trust management platform also includes a two-tier infrastructure, which comprises an infrastructure layer, a resource management layer, a data management layer, and a network communication layer.

[0025] In this embodiment, the federal trusted management platform also includes a two-tier infrastructure, which includes an infrastructure layer, a resource management layer, a data management layer, and a network communication layer.

[0026] The Infrastructure Layer uses a unified Resource Description Language (RDL) to abstractly model heterogeneous hardware resources. It includes three types of interfaces: ComputeResource, StorageResource, and NetworkResource, which describe computing performance, storage capacity, and network metrics, respectively. To enable cross-domain resource interoperability, this layer employs a Distributed HashTable (DHT) to establish a resource registration and discovery mechanism, allowing physical resources from different intelligence systems to be uniformly identified and indexed by the platform. It also integrates Docker containers and the Kubernetes orchestration platform to achieve containerized deployment and dynamic migration of intelligent agents, supporting rapid startup and scalability in heterogeneous environments.

[0027] Secondly, the Resource Management Layer introduces a multi-objective optimization-based scheduling algorithm to intelligently allocate resources during task execution. Specifically, it employs an improved Non-Dominated Sorting Genetic Algorithm (NSGA-II) to jointly optimize task execution time (makespan), energy consumption, and resource cost, dynamically outputting the Pareto optimal solution based on the scheduling objectives. Simultaneously, a Long Short-Term Memory (LSTM) network is used to predict historical load data, anticipating resource pressure and making advance scheduling adjustments. This layer also features an auto-scaling mechanism, automatically scaling computing resources up or down based on runtime monitoring data to improve resource utilization and responsiveness.

[0028] Next, the Data Management Layer provides a comprehensive data management system to address the quality and compliance issues of multi-source, cross-domain data. This layer constructs a Data Lineage Tracking System to record the metadata chain of data throughout its entire process, from collection and storage to processing and modeling, to meet regulatory audit requirements. Regarding data quality control, a Quality Score Model is introduced, constructing a scoring system based on four dimensions: completeness, accuracy, consistency, and timeliness.

[0029] The Quality Score is calculated using the formula: Quality Score = w1 × Completeness + w2 × Accuracy + w3 × Consistency + w4 × Timeliness. Completeness measures the proportion of non-missing values ​​in the data, calculated as the number of non-nullable fields divided by the total number of fields, expressed as Completeness = Number of non-nullable values ​​÷ Total number of values. Accuracy represents the correctness and accuracy of data records, calculated as 1 minus the ratio of erroneous records to the total number of records, expressed as Accuracy = 1 - (Number of erroneous records ÷ Total number of records). Consistency detects the proportion of logical conflicts or inconsistent formats in the data, calculated as 1 minus the ratio of inconsistent records to the total number of records, expressed as Consistency = 1 - (Number of inconsistent records ÷ Total number of records). Timeliness reflects the freshness of data. The formula is Timeliness equal to a natural exponential function with the exponent being the ratio of negative age to a timeliness threshold, expressed as Timeliness = e^(-age ÷ timeliness threshold)^(-exp(-age ÷ freshness threshold)). The four weights w1, w2, w3, and w4 are adjusted by technical experts according to the scenario. For problematic data, an intelligent repair module driven by machine learning automatically completes and corrects errors. Simultaneously, this layer supports distributed storage backends such as HDFS, Ceph, and MinIO, achieving high availability and redundancy disaster recovery capabilities.

[0030] Finally, the network communication layer employs the platform's self-developed Cross-domain Federated Agent Protocol (CFAP) to achieve efficient communication between agents across different protocol stacks. The CFAP protocol supports automatic adaptation to mainstream communication protocols such as HTTP / HTTPS, WebSocket, gRPC, and MQTT, and incorporates an intelligent routing mechanism that dynamically selects the optimal path based on metrics such as bandwidth, latency, and packet loss rate. Simultaneously, an adaptive flow control mechanism is introduced during communication to adjust the transmission rate in real time according to network congestion, preventing communication blockages or training failures caused by network fluctuations.

[0031] In summary, the federated trust management platform, through a two-tier infrastructure architecture consisting of an infrastructure layer, a resource management layer, a data management layer, and a network communication layer, enables intelligent, standardized, and automated management of underlying computing power, data, and communication, providing an operational foundation for cross-domain intelligence sharing and trusted federated modeling.

[0032] Furthermore, the platform provided in the application embodiments also includes:

[0033] The federal trust management platform also includes a three-tier application architecture, which includes a blockchain auxiliary layer and an application service layer.

[0034] In this embodiment, the federated trust management platform further includes a three-tier application architecture, comprising a Blockchain Assistance Layer and an Application Service Layer. The Blockchain Assistance Layer integrates Blockchain Enhanced Trust Management (BETM), deploying agent registration contracts, behavior recording contracts, trust assessment contracts, and task scheduling contracts based on the Ethereum smart contract runtime environment. This enables trusted registration, automatic recording, and traceable management of the identities, behaviors, and interaction results of all agents within the platform. The platform constructs a Trust Score model to dynamically quantify the trustworthiness of agents. This scoring model consists of four core dimensions: Behavior Score, Computation Score, Data Quality Score, and Validation Score.

[0035] Historical behavior score is calculated using the Behavior Decay Function, with the formula: Behavior Score = ∑(rating i ×decay factor^age i ), where rating i Age represents the score of the i-th task behavior. iThe time interval between the current action and the present is represented by the decay factor (e.g., 0.95), which is used to exponentially decay the impact of older actions, emphasizing the importance of recent actions. The contribution score is calculated by weighting and normalizing three parameters: Normalized Task Count (NT), Cumulative Computation Time (CT), and Resource Efficiency (RE). The ComputationScore = α × NT + β × CT + γ × RE, where α, β, and γ are weighting coefficients (e.g., 0.4, 0.3, 0.3). The data quality score is calculated based on four sub-indicators: Completeness (C), Accuracy (A), Consistency (S), and Timeliness (T). The four scores are calculated as follows: C = Number of valid fields ÷ Total number of fields, A = 1 - Error record ratio, S = 1 - Conflicting record ratio, T = exp(-Data age ÷ Timeliness threshold). The overall score is calculated as follows: Data Quality Score = w5 × C + w6 × A + w7 × S + w8 × T, where w5 to w8 are pre-set weights. The validation accuracy score is calculated based on the agent's accuracy in historical validation tasks, using the formula: Validation core = Number of correct validations ÷ Total number of validations. Newly registered nodes will be assigned a default neutral initial score.

[0036] The platform ultimately calculates the agent's trust score by weighting the scores across four dimensions: Trust Score = 0.3 × Behavior Score + 0.25 × Computation Score + 0.25 × Data Quality Score + 0.2 × Validation Score. This trust score serves as the basis for task scheduling priority, consensus voting weight, and permission policy configuration. The blockchain auxiliary layer further introduces an improved Practical Byzantine Fault Tolerance (PBFT) algorithm, combining it with the trust score as a consensus node voting weight indicator. A consensus threshold is set at 2 / 3 of the total trust weight to enhance fault tolerance against malicious nodes and improve collaborative stability.

[0037] To achieve trusted auditing of critical actions and integrity protection of off-chain data, the platform constructs a Distributed Evidence Storage System (DES), combining the InterPlanetary File System (IPFS) with a blockchain hybrid architecture. Sensitive data such as model updates, behavior logs, and training parameters are encrypted and uploaded to IPFS in slices. Simultaneously, smart contracts are used to store file hashes, integrity verification values, and metadata digests on the blockchain. The encryption keys employ Shamir's Secret Sharing algorithm for distributed storage and retrieval, ensuring that data is verifiable, traceable, and tamper-proof under authorization.

[0038] The Application Service Layer provides standardized service access and operational status visualization capabilities to external users and third-party systems. The platform deploys a unified Application Programming Interface Gateway (API Gateway) that supports RESTful and gRPC interface protocols, performing authentication, access control, request frequency control, and call logging. The platform builds a Service Orchestrator, which decouples and encapsulates various functional modules—including Differential Privacy Protection (TIPP), Enhanced Modeling (FedAvg+), Collaboration Mechanism (AACA), and Trust Management (BETM)—based on a microservice architecture. Service flows are described using a Directed Acyclic Graph (DAG), enabling task logic orchestration and dynamic module combination. The platform also provides a Monitoring Dashboard, which provides real-time visualization of model training progress, resource utilization, trust score evolution, communication load, API call frequency, and blockchain write status, providing decision-making support and strategy adjustment assistance for platform operations personnel.

[0039] In summary, the three-tier application architecture enables trusted management and evidence storage of agent behavior through a blockchain auxiliary layer, and standardizes and enables open interaction of platform services through an application service layer, supporting a trusted federated collaborative modeling and operation mechanism for the platform in multi-source heterogeneous and cross-domain sensitive intelligence environments.

[0040] The fusion security layer performs privacy protection processing on the multi-source cross-domain intelligence data of the multiple cross-domain intelligence systems accessed based on the TIPP algorithm. The fusion security layer includes a differential privacy module, a homomorphic encryption module, and a zero-knowledge proof module.

[0041] In this embodiment, the fusion security layer performs privacy protection processing on multi-source cross-domain intelligence data from multiple cross-domain intelligence systems based on the TIPP algorithm (Trusted Information Protection Protocol), aiming to achieve secure desensitization of data before it leaves the domain and trusted use within the platform.

[0042] The integrated security layer comprises a differential privacy module, a homomorphic encryption module, and a zero-knowledge proof module. These modules work collaboratively to construct a multi-layered privacy protection chain. The differential privacy module is used to perturb the model parameters involved in the training process, reducing the impact of a single data point on the model output. The homomorphic encryption module encrypts the perturbed model parameters, allowing the encrypted data to still participate in model training and aggregation. The zero-knowledge proof module verifies the correctness of the differential perturbation and encryption operations without revealing the underlying data or encryption keys. Through the integrated operation of these modules, the integrated security layer prevents privacy risks such as model reverse engineering, transmission eavesdropping, and collaborative fraud, ensuring that multi-source data cannot leave the domain. This provides compliant, secure, and verifiable training input for the platform's subsequent federated learning tasks.

[0043] Furthermore, the platform provided in the application embodiments also includes:

[0044] The fusion security layer performs privacy protection processing on the multi-source cross-domain intelligence data from the multiple cross-domain intelligence systems accessed, based on the TIPP algorithm. The differential privacy module is used to perform noise perturbation processing on the training model parameters of the multi-source cross-domain intelligence data. The homomorphic encryption module is used to perform homomorphic encryption on the noise-perturbed training model parameters. The zero-knowledge proof module is used to verify the zero-knowledge proof of the noise perturbation processing and the homomorphic encryption.

[0045] In this embodiment, the fusion security layer uses the TIPP (Trusted Information Protection Protocol) algorithm to perform phased privacy protection processing on multi-source cross-domain intelligence data provided by multiple cross-domain intelligence systems. This ensures that during federated modeling, the compliance requirement that data from each security domain cannot leave its domain is met, while also preventing privacy leaks during computation and interaction. This processing flow is completed collaboratively by three functional modules: a differential privacy module, a homomorphic encryption module, and a zero-knowledge proof module. Each module progressively enhances the security protection strength of the data according to a privacy protection chain logic.

[0046] First, the Differential Privacy Module performs noise perturbation on the local model parameters generated during the training of each cross-domain system. This module employs the Laplace mechanism, which minimizes the dependence of the training output on a single sample by introducing controlled randomness into the model parameters. The perturbation process is performed based on the following function: θ' i =θ i +Lap(Δf / ε t ); where θ i Let θ' represent the i-th original model parameter. i Here, ε represents the perturbation parameter after adding noise, Δf represents the function sensitivity (measures the maximum change in model output caused by neighboring datasets), and ε represents the perturbation parameter. t Let Lap(·) represent the privacy budget for the current round, and let Lap(·) denote the noise sampled from the Laplace distribution. This process makes the impact of any single sample on the overall model behavior imperceptible, thus achieving ε-differential privacy protection.

[0047] Secondly, the homomorphic encryption module performs encrypted encryption on the perturbed model parameters. The platform employs an additive homomorphic encryption algorithm (such as the Paillier encryption scheme), supporting direct addition and weighted averaging operations within the ciphertext space. When participating in global model aggregation, the encrypted model parameters can complete the federated update process without decryption, effectively blocking attack paths that could lead to intermediate value leakage during model aggregation and ensuring that data remains encrypted throughout the training, transmission, and computation stages.

[0048] Finally, the zero-knowledge proof module verifies the correctness of the differential perturbation process and homomorphic encryption operation. Based on a non-interactive zero-knowledge proof protocol, this module allows data holders to provide valid proofs to other nodes on the platform without revealing any original data, keys, or model details, ensuring the authenticity and integrity of their privacy processing. The platform can quickly verify the submitted proof structure through lightweight verification functions, improving collaborative trust without significantly increasing communication overhead.

[0049] Through the above three-stage processing mechanism, the integrated security layer realizes the privacy protection process of perturbation-encryption-verification under the overall coordination of the TIPP algorithm. It constructs a security closed loop of differential privacy protection, encrypted modeling and verifiable collaboration at the training parameter level, providing compliant, high-strength and full-process privacy security support for the platform's federated modeling tasks in multi-source cross-domain intelligence environments.

[0050] Furthermore, in the platform provided in the application embodiment, the fused security layer further includes a collaborative optimization mechanism, which includes: acquiring differential privacy noise processed by the differential privacy module and encryption parameters for homomorphic encryption by the homomorphic encryption module; the collaborative optimization mechanism is used to randomly share the differential privacy noise and the encryption parameters.

[0051] In this embodiment, the fused security layer includes a collaborative optimization mechanism for establishing a randomness-sharing mechanism between the differential privacy module and the homomorphic encryption module for differential privacy noise and encryption parameters, thereby uniformly managing randomness resources in cross-stage privacy operations. The collaborative optimization mechanism includes acquiring the differential privacy noise generated during noise perturbation processing by the differential privacy module and the encryption parameters generated by the homomorphic encryption module during encryption operations, and synchronously invoking and mapping these two types of random factors based on a unified task context.

[0052] The collaborative optimization mechanism introduces a unified randomness index management table, taking metadata such as model parameter dimensions, node identities, and computation rounds from each training round as input to drive the platform's internal pseudo-random number generator to generate an initial random seed. This seed is used to generate differential privacy noise in the differential privacy module and is simultaneously mapped to the encryption random factor required to generate encryption parameters in the homomorphic encryption module, ensuring structural consistency and execution coherence in privacy protection processing between the two stages.

[0053] To ensure operational security and verifiability, the collaborative optimization mechanism also works in conjunction with the blockchain auxiliary layer to hash the generated differential privacy noise and the randomness seed used in the encryption parameters, and records this hash on the chain via a smart contract to form an immutable audit credential. Through this collaborative design, the platform achieves efficient collaboration between the differential privacy module and the homomorphic encryption module in the integrated security layer, improving resource utilization, system consistency, and the credibility of the federated training process for multi-source cross-domain intelligence data in the privacy protection process.

[0054] Furthermore, the platform provided in the application embodiments also includes:

[0055] The differential privacy module is used to perform noise perturbation processing on the training model parameters of the multi-source cross-domain intelligence data. The magnitude of the noise perturbation processing is obtained through a privacy budget adaptive allocation mechanism. The privacy budget adaptive allocation mechanism adaptively outputs based on the weights of the model training rounds and the factors of model performance to obtain the privacy resources for noise perturbation processing.

[0056] In this embodiment, the differential privacy module is used to perform noise perturbation processing on the parameters of the local training model of multi-source cross-domain intelligence data. The perturbation amplitude is dynamically generated by the privacy budget adaptive allocation mechanism built inside the platform to accurately control the trade-off between privacy leakage risk and model accuracy.

[0057] The privacy budget adaptive allocation mechanism first constructs training phase weights (w) based on the current training round and the total number of training rounds. stage The weight is calculated using a sine function, expressed as follows: Among them, w stage This indicates the weight of the current training phase. This factor changes dynamically as the training progresses, with stronger perturbations applied in the early stages to enhance privacy protection, and the perturbation intensity gradually reduced in the later stages to improve model accuracy.

[0058] The platform then analyzes the performance trends of the model in recent rounds (e.g., the last 5 rounds) and calculates an adjustment factor based on the improving or declining trends of performance metrics. When a declining trend in model performance is detected, the platform sets the adjustment factor to 1.2, increasing the current budget to enhance protection. If the performance trend is stable or improving, the adjustment factor is set to 0.8 to reduce noise intensity and maintain stable model performance. If the number of sample rounds is insufficient, the default adjustment factor is 1.0.

[0059] Based on this, the remaining budget for the current training phase is calculated by combining the total privacy budget with the sum of the used budget. The per-round budget allocation is then estimated based on the current round and the remaining rounds (total rounds - round + 1).

[0060] The specific calculation formula is as follows: Where, ε t This indicates the privacy budget for the current training round.

[0061] The privacy budget ε t The noise sampling amplitude in the Laplace perturbation mechanism is passed as an input parameter to the differential privacy module and applied to the perturbation formula of the training model parameter θ as follows: θ' i =θ i +Lap(Δf / ε t ).

[0062] Through the aforementioned adaptive mechanism, privacy resources for noise perturbation processing are flexibly and dynamically obtained based on the training progress and model feedback accuracy, thereby ensuring differential privacy protection while minimizing the impact on model performance.

[0063] The federated learning scheduling layer schedules federated learning modeling tasks based on privacy-preserving multi-source cross-domain intelligence data. The agent management layer allocates resources for federated learning modeling tasks within the federated learning scheduling layer through the AACA adaptive collaboration mechanism, resulting in a shared global federated learning model.

[0064] In this embodiment, the federated learning scheduling layer in the federated trust management platform is responsible for scheduling federated modeling tasks on multi-source cross-domain intelligence data that has undergone privacy protection processing and is output by the fusion security layer. This scheduling process follows a heterogeneous collaborative distributed modeling logic, relying on the enhanced federated averaging algorithm FedAvg+ to dynamically integrate local model results. FedAvg+ introduces extended mechanisms such as node trust weights and adaptive aggregation factors on top of traditional weighted averaging, effectively addressing model bias and convergence instability issues in non-independent identically distributed (Non-IID) scenarios. Through this mechanism, the platform completes parameter interaction and model synthesis while maintaining the locality of data at each participating node, achieving cross-domain collaborative training.

[0065] Resource allocation for federated learning scheduling tasks is controlled by the Agent Management Layer, which introduces the Adaptive Agent Collaboration Architecture (AACA) mechanism as its core decision-making module. The AACA mechanism constructs a multi-agent reinforcement learning model based on the MADDPG multi-agent deep deterministic policy gradient algorithm. It perceives multi-dimensional state information such as the computational resource utilization, data quality, bandwidth status, and task load of participating nodes, and adjusts the role allocation and resource allocation strategies in federated training in real time accordingly.

[0066] Before task scheduling begins, the state of all available agents is evaluated using the AACA mechanism, and corresponding computational and data processing tasks are allocated based on the complexity of the training task and resource requirements. During scheduling, agents generate local model updates through local training. This updated data is protected by differential privacy and homomorphic encryption in the fusion security layer before being uploaded to the scheduling layer for aggregation. The platform calculates the federated learning global model according to the aggregation logic of the FedAvg+ algorithm and broadcasts the model back to all participating agents, achieving synchronous updates of the modeling results.

[0067] Through the collaborative efforts of the federated learning scheduling layer and the agent management layer, a distributed training system is built in a multi-source heterogeneous intelligence environment, ultimately resulting in a federated learning global model for sharing, achieving intelligent information fusion and collaborative modeling without touching the original data.

[0068] Furthermore, the platform provided in the application embodiments also includes:

[0069] The agent management layer allocates resources for the federated learning modeling task within the federated learning scheduling layer through the AACA adaptive cooperation mechanism, including: constructing multiple agents based on the federated learning modeling task; obtaining the multi-dimensional state space of the multiple agents; using the multi-objective reward function of the AACA adaptive cooperation mechanism to evaluate the cooperation of the multiple agents based on the multi-dimensional state space, and outputting multiple evaluation results, wherein the multi-objective reward function includes model accuracy reward, privacy protection strength reward, cooperation quality reward, and resource utilization efficiency reward; and obtaining the cooperation strategy for the federated learning modeling task based on the multiple evaluation results, wherein the cooperation strategy includes cooperative resource allocation.

[0070] In this embodiment, the agent management layer allocates resources for federated learning modeling tasks within the federated learning scheduling layer through the AACA adaptive collaboration mechanism.

[0071] Specifically, for each federated learning modeling task, multiple corresponding agents are first constructed. Each agent represents a node participating in computation, and its behavioral decision-making capabilities within the platform are mapped. Then, a multi-dimensional state space is obtained for these agents. This multi-dimensional state space reflects the agent's capability state and environmental characteristics across four key dimensions: computational capability state, data quality state, network state, and task load state. Specifically, computational capability state includes CPU utilization, GPU availability, memory usage, processing speed, and concurrent processing capability; data quality state includes local dataset size, data integrity, label quality, data distribution diversity, and update frequency; network state includes bandwidth, latency, packet loss rate, connection stability, and geographical location; and task load state includes the number of current tasks, task complexity, resource requirement vector, and task priority.

[0072] Next, based on the multi-objective reward function defined by the AACA adaptive cooperation mechanism, the cooperation of each agent in the multi-dimensional state space is evaluated. The evaluation content covers four optimization objectives, including model accuracy reward, privacy protection strength reward, cooperation quality reward, and resource utilization efficiency reward.

[0073] The accuracy reward reflects the accuracy of the local model trained by the agent. It is calculated using an exponential function, expressed as: accuracy reward = μ·(2 A -1), where A represents the model accuracy, and μ is the weight parameter for the model accuracy reward, which is preset.

[0074] The privacy reward measures the privacy protection capability demonstrated by an agent in performing differential privacy and homomorphic encryption. It is modeled using a logarithmic function, specifically expressed as privacy reward = γ·log(1+P), where P is the privacy strength index, γ is a pre-defined weighting coefficient for the privacy reward, and P can be obtained by jointly evaluating the noise injection level and encryption strength.

[0075] Collaboration reward measures the cooperation effectiveness and response efficiency of an agent during collaborative training with other nodes. It is mapped using the sigmoid function. Here, Q represents the coordination quality score, and δ is a moderating factor for coordination quality rewards. A higher Q value indicates smoother and more effective coordination.

[0076] Resource utilization efficiency reward represents an agent's ability to use allocated resources to perform effective training tasks; its reward form is as follows: Where U is the effective computation, R is the total resource consumed, and ε is the weighting factor for resource utilization efficiency rewards.

[0077] The four sub-reward values ​​are summed to form the total evaluation score for each agent. Then, the non-dominated sorting genetic algorithm NSGA-II is used to construct a Pareto front solution set from the evaluation results, forming multiple candidate cooperative strategies. These strategies are then input into the reinforcement learning model within the AACA mechanism. Simulation experiments and policy updates are conducted using the MADDPG algorithm, ultimately selecting the optimal cooperative strategy for task execution. The cooperative strategy includes a cooperative resource allocation scheme, determining the computational and network resource usage weights and priorities for each agent. It also includes a role allocation mechanism and communication coordination scheme, enabling differentiated configuration of functions and communication frequency scheduling among agents, thereby efficiently completing federated learning modeling tasks in heterogeneous environments.

[0078] Furthermore, the platform provided in the application embodiments also includes:

[0079] The intelligent agent management layer is equipped with a blockchain-enhanced trust management mechanism. This mechanism enhances the trust weights of multiple intelligent agents within the management layer through a trust evaluation model. The evaluation dimensions of the trust evaluation model include historical behavior, computational contribution, data quality, and verification accuracy.

[0080] In this embodiment, the intelligent agent management layer deploys a blockchain-enhanced trust management mechanism to strengthen the trust weight of multiple intelligent agents within the management layer. This mechanism, based on the Ethereum smart contract runtime environment, achieves trusted registration, automatic recording, and traceable management of the identities, behaviors, and interactions of all intelligent agents through intelligent agent registration contracts, behavior recording contracts, trust evaluation contracts, and task scheduling contracts. Specifically, a trust evaluation model (TrustScore) is introduced to quantitatively evaluate the trustworthy behavior of intelligent agents from multiple dimensions, and the evaluation results are used to dynamically enhance their trust weight within the intelligent agent management layer, thus achieving trust weight enhancement for multiple intelligent agents within the management layer.

[0081] The trust assessment model comprises four dimensions: historical behavior, computational contribution, data quality, and verification accuracy. The historical behavior dimension uses a historical behavior decay function to weight the score of task behavior over time, emphasizing the importance of recent behavior. The computational contribution dimension combines indicators such as the number of tasks completed by the agent, cumulative computation time, and resource efficiency, and obtains a score after weighted normalization. The data quality dimension comprehensively evaluates the completeness, accuracy, consistency, and timeliness of the data, forming a composite score reflecting data reliability. The verification accuracy dimension calculates the correctness rate of the agent's participation in historical verification tasks, used to assess the credibility of its output results. The platform ultimately weights and combines the scores of these four dimensions according to preset weights to obtain the agent's overall trust score.

[0082] With the enhanced trust management mechanism of blockchain, trust scores are not only used to guide the task scheduling priority and resource allocation strategy in the federated learning modeling process, but also serve as the basis for voting weight in the consensus process, so as to improve the ability to resist malicious behavior and the stability of multi-agent collaboration, thereby comprehensively enhancing the trust weight of multiple agents in the agent management layer.

[0083] Furthermore, the platform provided in the application embodiments also includes:

[0084] The Federal Trust Management Platform also includes a standardized processing template, which includes a protocol conversion engine and a data processing model; the standardized processing template is used to convert and process multiple cross-domain intelligence systems connected to the Federal Trust Management Platform.

[0085] In this embodiment, the Federal Trusted Management Platform also includes a standardized processing template, which comprises a protocol conversion engine and a data processing model. This template is used to perform unified standard conversion processing on heterogeneous protocols and data formats accessed by multiple cross-domain intelligence systems. The process begins with the protocol conversion engine identifying and adapting communication protocols. Based on a pre-built protocol mapping rule base, the engine automatically parses the communication protocols used by different intelligence systems, supporting multiple network protocol formats including HTTP, HTTPS, WebSocket, MQTT, and gRPC. The engine introduces an intermediate abstract protocol model to complete the mapping conversion between the source protocol and the platform's internal standard protocols, ensuring that the message body structure, request methods, and parameter passing mechanisms between different systems remain semantically consistent, thereby achieving standardized access to message exchange and remote access interfaces.

[0086] After completing the communication protocol conversion, the data processing model begins structuring and semantically standardizing the cross-domain intelligence data. This model combines schema mapping technology with a domain knowledge-driven metadata matching mechanism to automatically identify the field structure, data type, and naming rules of the source data. It then uses a field semantic alignment table to map these data to a unified data schema defined by the platform, achieving standardized field names, unified data formats, and semantic consistency. To ensure the accuracy and completeness of data fusion, the data processing model also introduces rule-based missing value imputation algorithms and outlier detection mechanisms to preprocess input data that is structurally incomplete, has inconsistent data types, or contains interfering values. This includes, but is not limited to, mean imputation, context inference, and data unit normalization.

[0087] Through the synergy of the aforementioned protocol conversion engine and data processing model, standardized processing templates enable protocol adaptation and data semantic fusion for multiple heterogeneous cross-domain intelligence systems during their access to the federated trust management platform. This provides consistent and highly resolvable standard input data support for privacy protection modeling and federated task scheduling, ensuring the stability and reliability of the cross-domain system collaborative training environment.

[0088] In summary, the embodiments of this application have at least the following technical effects:

[0089] The federated trusted management platform described in this application connects to multiple cross-domain intelligence systems. The federated trusted management platform includes a primary core architecture comprising a fusion security layer, a federated learning scheduling layer, and an agent management layer. The fusion security layer performs privacy protection processing on the multi-source cross-domain intelligence data from the multiple connected cross-domain intelligence systems based on the TIPP algorithm. This fusion security layer includes a differential privacy module, a homomorphic encryption module, and a zero-knowledge proof module. The federated learning scheduling layer schedules federated learning modeling tasks according to the privacy-protected multi-source cross-domain intelligence data. The agent management layer allocates resources for the federated learning modeling tasks within the federated learning scheduling layer through an AACA adaptive collaboration mechanism, obtaining a shared global federated learning model. This invention addresses the technical problem of multi-layered privacy leakage risks faced by multi-source cross-domain intelligence in the federated learning process in existing technologies. By introducing the TIPP triple fusion privacy protection mechanism into the fusion security layer and collaboratively integrating differential privacy, homomorphic encryption, and zero-knowledge proof modules, it achieves the technical effect of trusted sharing and joint modeling of cross-domain intelligence while ensuring data privacy and security.

[0090] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0091] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0092] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A federal trust management platform based on cross-domain intelligence sharing, characterized in that, The federated trust management platform is connected to multiple cross-domain intelligence systems. The federated trust management platform includes a primary core architecture, which includes a converged security layer, a federated learning scheduling layer, and an intelligent agent management layer. The fusion security layer performs privacy protection processing on the multi-source cross-domain intelligence data of the multiple cross-domain intelligence systems accessed based on the TIPP algorithm. The fusion security layer includes a differential privacy module, a homomorphic encryption module, and a zero-knowledge proof module. The federated learning scheduling layer schedules federated learning modeling tasks based on privacy-preserving multi-source cross-domain intelligence data. The agent management layer allocates resources for federated learning modeling tasks within the federated learning scheduling layer through the AACA adaptive collaboration mechanism to obtain a shared global federated learning model. The differential privacy module is used to perform noise perturbation processing on the training model parameters of the multi-source cross-domain intelligence data; The homomorphic encryption module is used to homomorphically encrypt the parameters of the training model after noise perturbation, and the zero-knowledge proof module is used to verify the zero-knowledge proof of noise perturbation processing and homomorphic encryption. The fused security layer also includes a collaborative optimization mechanism, which includes: Obtain the differential privacy noise processed by the differential privacy module and the encryption parameters for homomorphic encryption by the homomorphic encryption module; The collaborative optimization mechanism is used to randomly share the differential privacy noise and the encryption parameters; The differential privacy module is used to perform noise perturbation processing on the training model parameters of the multi-source cross-domain intelligence data, and the magnitude of the noise perturbation processing is obtained through a privacy budget adaptive allocation mechanism. The privacy budget adaptive allocation mechanism adaptively outputs based on the weights of the model training rounds and the factors of model performance, thereby obtaining privacy resources that have undergone noise perturbation processing. The agent management layer allocates resources for federated learning modeling tasks within the federated learning scheduling layer through the AACA adaptive collaboration mechanism, including: Construct multiple intelligent agents based on the federated learning modeling task; The multidimensional state space of the multiple agents is obtained, and the multi-objective reward function of the AACA adaptive cooperation mechanism is used to evaluate the cooperation of the multiple agents based on the multidimensional state space, and output multiple evaluation results. The multi-objective reward function includes model accuracy reward, privacy protection strength reward, cooperation quality reward and resource utilization efficiency reward. The collaboration strategy for the federated learning modeling task is obtained based on the multiple evaluation results, and the collaboration strategy includes the allocation of collaborative resources.

2. The platform as described in claim 1, characterized in that, The primary core architecture also includes a trusted computing layer, through which a trusted execution environment is constructed, and the federated learning scheduling layer executes federated learning modeling tasks in the trusted execution environment.

3. The platform as described in claim 1, characterized in that, The federal trust management platform also includes a two-tier infrastructure, which comprises an infrastructure layer, a resource management layer, a data management layer, and a network communication layer.

4. The platform as described in claim 1, characterized in that, The federal trust management platform also includes a three-tier application architecture, which includes a blockchain auxiliary layer and an application service layer.

5. The platform as described in claim 1, characterized in that, The intelligent agent management layer is equipped with a blockchain-enhanced trust management mechanism, which enhances the trust weight of multiple intelligent agents in the intelligent agent management layer through a trust assessment model. The evaluation dimensions of the trust assessment model include historical behavior dimension, computational contribution dimension, data quality dimension, and verification accuracy dimension.

6. The platform as described in claim 1, characterized in that, The federal trust management platform also includes standardized processing templates, which include a protocol conversion engine and a data processing model. The standardized processing template is used to transform and process multiple cross-domain intelligence systems connected to the federal trust management platform.

Images