Delay security path customization scheduling method in a2a isomer non-trustworthy environment

Abstract

This invention discloses a customized scheduling method for latency-secure paths in A2A heterogeneous untrusted environments, relating to the field of artificial intelligence technology. The method includes the following steps: S1, constructing a task-customized joint constraint model of latency, security, and routing; S2, constructing a dynamic security subgraph dependent on the task, designing a trust-gated graph neural network encoder, and learning refined representations of topology and security dependencies from the dynamic network state; S3, constructing a distributed multi-agent network, introducing a DSR security action filtering mechanism to pre-screen infeasible and unsafe actions, constructing a primary and dual multi-agent near-end policy optimization framework, and completing model training and online inference deployment. This invention employs the above-mentioned customized scheduling method for latency-secure paths in A2A heterogeneous untrusted environments, improving the task constraint satisfaction capability, scheduling stability, and secure completion capability in untrusted heterogeneous network environments, and achieving customized latency-security-routing collaborative intelligent task scheduling.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a time-delay-safe path customization scheduling method in A2A heterogeneous and untrusted environments. Background Technology

[0002] In recent years, task scheduling in heterogeneous network environments has become a key research direction for network resource optimization and security assurance. Researchers have conducted extensive research on heuristic optimization, graph neural networks, and multi-agent reinforcement learning, but these approaches still fall short of meeting the collaborative scheduling needs of dynamic, untrusted scenarios. In the field of heuristic and optimization scheduling, the IEEE Internet of Things Journal proposed a multi-objective Harris Hawks optimization fog scheduling method in 2024, a decomposed scheduling method for safety-critical mobile cyber-physical systems was proposed in the same journal in 2022, and a security-driven scheduling model for latency-sensitive tasks in fog networks was proposed in Springer in 2022. While these studies incorporate latency and security factors, they largely rely on heuristic search, lacking global optimality and dynamic adaptability, and their high-complexity modeling cannot be adapted to large-scale heterogeneous, untrusted networks.

[0003] In the field of graph neural network modeling, a heterogeneous graph attention network was proposed at the World Wide Web Conference in 2019, achieving fusion representation of multi-type node and edge information. In 2024, the TrustGuard trust evaluation framework based on GNN was proposed at IEEE Transactions on Dependable and Secure Computing, supporting dynamic trust quantification. These technologies can effectively represent topology and trust relationships, but they only focus on trust evaluation and do not achieve a unified integration of trust modeling with scheduling constraints, action feasibility, and routing constraints, thus failing to directly support scheduling decisions. In the field of multi-agent reinforcement learning scheduling, a flexible job shop scheduling method combining GNN and deep reinforcement learning was proposed at IEEE Transactions on Industrial Informatics in 2022; a large-scale machine learning cluster multi-agent graph reinforcement learning scheduling method was proposed at IEEE Transactions on Network and Service Management in 2022; MAPPO was used for UAV cooperative trajectory optimization at the IEEE WoWMoM conference in 2023; and privacy-preserving multi-agent reinforcement learning was used for edge computing resource allocation at IEEE Transactions on Cognitive Communications and Networking in 2025. Related research further confirms the constraint coupling problem: In 2023, IEEE Transactions on Communications pointed out that scheduling needs to take into account time constraints; in 2024, IEEE Transactions on Mobile Computing showed that bandwidth and routing feasibility significantly affect edge scheduling; latency, security, and routing constraints are coupled with each other, and existing technologies have not provided a unified model for dynamic trust evolution, node service capacity decay, queuing latency, and routing feasibility. Summary of the Invention

[0004] The purpose of this invention is to provide a time-delay-secure path customization scheduling method in A2A heterogeneous untrusted environments, which solves the problem that existing task scheduling schemes in heterogeneous untrusted networks, as mentioned in the background, cannot simultaneously satisfy latency, security, and hard routing constraints under dynamic attacks, time-varying trust, and heterogeneous resource conditions.

[0005] To achieve the above objectives, this invention provides a time-delay-safe path customization scheduling method for A2A heterogeneous untrusted environments, comprising the following steps: S1. To address the heterogeneous untrusted network infrastructure and differentiated task requirements, a directed graph is used to formally model network nodes, communication links, and their inherent attributes. Seven-tuples are used to define the core parameters of heterogeneous tasks, constructing a task-customized joint constraint model for latency, security, and routing. This joint constraint model takes network topology, node trust levels, link transmission characteristics, and task-specific requirements as input to generate and dynamically update a task-specific security subgraph. A trust-aware service queuing model is used to output end-to-end performance metrics and scheduling hard constraint boundaries. S2. Based on the joint constraint model and task-specific security subgraph in S1, limit the range of legal topologies for scheduling; based on the end-to-end performance indicators and scheduling constraints output by S1, design a graph neural network encoder with a trust gating mechanism to learn a refined representation of topology and security dependencies from dynamic network states. S3. Based on the refined representation results obtained in S2, construct a distributed multi-agent network, introduce a DSR safe action filtering mechanism to screen and retain compliant scheduling actions; adopt a centralized training distributed execution paradigm to construct a primary and dual multi-agent proximal policy optimization framework, and perform model training and online inference deployment through a dual-timescale primary-dual update strategy.

[0006] Preferably, the specific steps of S1 are as follows: S11. Model the heterogeneous untrusted infrastructure as a directed graph and construct a heterogeneous untrusted network infrastructure model; define the set of heterogeneous task flows and the single task seven-tuple representation within the scheduling time slot, and construct a customized task model; define the scheduling time slot duration and the set of neighbor nodes for each node, and complete the basic formal definition of the scheduling scenario. S12. For each task, based on the global network state, select the set of nodes and edges that meet the requirements of security, congestion and transmission, generate a safe subgraph model that the task depends on, and set a triggered dynamic update mechanism to provide feasible topological constraints for subsequent scheduling decisions. S13. Quantify the impact of node trust value on effective computing power through trust decay factor, derive node effective computing power and service rate, define three types of decision variables: task admission, execution deployment and routing forwarding, calculate node processing latency and define link transmission latency based on M / G / 1 queuing model, and construct a trust-aware service and queuing model. S14. Define aggregation rules for end-to-end total latency, security cost, and reliability cost; set deployment and admission consistency, node computing capacity, link bandwidth capacity, and end-to-end DSR customized hard constraints; and clarify the performance indicators and constraint boundaries of scheduling decisions. S15. The scheduling problem is formalized into a stochastic optimization problem of maximizing system utility with long-term DSR constraints. Non-negative dual variables are introduced for Lagrangian relaxation, which transforms it into a saddle point problem.

[0007] Preferably, the specific steps of S11 are as follows: S111. Model the heterogeneous untrusted infrastructure as a directed graph and define the core attributes of each node and each communication link respectively. Among them, the core attributes of the node include nominal computing power, real-time queue back pressure and dynamic trust value, and the core attributes of the communication link include link bandwidth, one-way propagation delay and data packet loss probability. S112. For heterogeneous task flows arriving within the scheduling time slot, define a task index set and the corresponding task set, and define a single task as a complete seven-tuple. S113. Define the duration of a single scheduling time slot and construct a corresponding set of neighbor nodes for each node, which will be used for subsequent routing decisions and topology feature extraction.

[0008] Preferably, the specific steps of S12 are as follows: S121. For each task, based on the globally observable state at the corresponding moment, construct a set of safe nodes that meet the customized safety and congestion requirements of the task. S122. Based on the set of secure nodes, construct a set of secure edges that meets the requirements of task transmission; S123. Based on the set of safe nodes and the set of safe edges, generate a task-specific safety subgraph to provide a conservative and feasible topology for subsequent scheduling decisions and eliminate interference from low-trust, high-congestion nodes and unreliable links. S124. For dynamic updates of the security subgraph, a minimum update interval and a triggered update mechanism are set. When the change in the trust value of a node in the network or the back pressure of the node queue exceeds the threshold, the reconstruction of the security subgraph is immediately triggered to ensure the consistency between topological constraints and the real-time state of the network.

[0009] Preferably, the specific steps of S13 are as follows: S131. Define the trust decay factor to quantify the impact of node trust value on effective computing power. S132. Based on the trust decay factor, the effective computing power and effective service rate of the computing node; S133. Define three types of core decision variables for task scheduling. The core decision variables include task admission decision, task execution deployment decision, and task routing and forwarding decision. S134. Based on the queuing model, calculate the average processing latency of the nodes.

[0010] Preferably, the specific steps of S14 are as follows: S141. Based on node processing delay and link transmission delay, the end-to-end total delay of the task is obtained by aggregation; S142. The end-to-end security cost and end-to-end reliability cost of the aggregation task are used to quantify the cumulative security risk and cumulative transmission failure risk of the entire task path, respectively. S143. Define the hard constraints that scheduling decisions must meet, including deployment and admission consistency constraints, node computing capacity constraints, link bandwidth capacity constraints, and end-to-end DSR customization constraints.

[0011] Preferably, the specific steps of S15 are as follows: S151. The scheduling problem is formalized as a constrained long-term system utility maximization problem, and the objective function is defined. S152. Transform the long-term DSR constraint into a discount accumulation constraint function, define the time delay constraint function, the security constraint function and the reliability constraint function respectively, and transform the original problem into a stochastic optimization problem with long-term constraints. S153. Introduce a nonnegative dual variable vector, perform Lagrangian relaxation on the long-term DSR constraint, construct the Lagrangian function, and transform the stochastic optimization problem into a saddle point solution problem.

[0012] Preferably, the specific steps of S2 are as follows: S21. By defining the seven-tuple form of Dec-POMDP, the local observation space and the global state space, the jointly encoded local action space and the constraint-aware reward function, a decentralized partially observable Markov decision process with task conditionalization is constructed, providing a standardized Markov decision framework for distributed scheduling decisions. S22. By embedding and fusing task-conditional nodes, focusing attention aggregation in trust-gated graphs, updating temporally dependent states, and constraining representation stability, a trust-gated graph encoder is designed to achieve a fusion representation of topology, trust state, and task requirements. S23. By constructing a hierarchical set of actionable actions, constructing an additive mask matrix and filtering the log probability of actions, performing secondary verification of action legality and caching mechanism, and prioritizing actionable actions, a DSR safe action filtering mechanism is designed to block infeasible and unsafe actions before policy sampling. S24. Representation quality assessment and feature optimization are performed through representation optimization based on contrastive learning, representation consistency constraints, and feature dimension adaptation.

[0013] Preferably, the specific steps of S3 are as follows: S31. By designing a distributed actuator network, a centralized evaluator network, iteratively updating dual variables, optimizing training stability, and constructing a distributed multi-agent network, a primitive dual MAPPO optimization framework based on a centralized training and distributed execution paradigm is constructed. S32. Design a dual-timescale original-dual update strategy, including fast-timescale original strategy update, full-timescale dual variable update and timescale adaptive adjustment, to achieve decoupling and synergy between strategy optimization and constraint control. S33. Based on spectral normalization and trust gating, stochastic convex optimization, performance difference lemma and safe subgraph filtering and action masking mechanism, configure gradient propagation constraints, dual variable update strategy, policy improvement rules and constraint verification logic for algorithm training to ensure encoder gradient stability, dual variable convergence, monotonic policy performance improvement and scheduling actions always satisfy hard constraints. S34. Through end-to-end model training, online inference deployment, performance monitoring and updates, and generalization robustness enhancement processes, the scheduling framework achieves stable convergence, real-time decision-making, continuous optimization, and reliable operation.

[0014] Preferably, the specific steps of S34 are as follows: S341. Design the end-to-end process of the model, including training initialization, trajectory sampling, policy and value network update, dual variable update and iteration termination judgment, to achieve stable convergence of the framework. S342, Online inference deployment process design, including real-time environmental state perception, construction of task-specific security subgraphs, feature encoding and action filtering and multi-agent distributed scheduling decision generation, to achieve real-time scheduling decision; S343, Model performance monitoring and online update mechanism, including real-time performance monitoring, distributed offset detection and incremental online update; S344, Model generalization and robustness enhancement mechanisms, including multi-scenario data hybrid training, adversarial training enhancement, and incremental fine-tuning mechanisms.

[0015] Therefore, the present invention employs the aforementioned customized scheduling method for delay-safe paths in heterogeneous and untrusted A2A environments, which has the following beneficial effects: (1) The present invention can improve the ability to satisfy task constraints, schedule stability and secure completion in untrusted heterogeneous network environments.

[0016] (2) This invention effectively addresses the pain points of multi-constraint coupling and insufficient security in heterogeneous untrusted networks, and realizes customized latency-security-routing collaborative intelligent task scheduling.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the time-delay-safe path customization scheduling method for A2A heterogeneous untrusted environments according to the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example Please see Figure 1 This invention provides an A2A (Agent-to-Agent) customized scheduling method for latency-secure paths in heterogeneous and untrusted environments. In this embodiment, A2A can be understood as an information interaction and collaborative decision-making mechanism among multiple scheduling agents in a heterogeneous and untrusted network. Specifically, edge nodes, cloud nodes, and routing gateways can all be abstracted as scheduling agents. Each agent interacts based on local state, task requirements, actionable sets, and trust information to achieve collaborative scheduling of task admission, execution deployment, and path routing. To align with the patent title "A2A Customized Scheduling Method for Latency-Secure Paths in Heterogeneous and Untrusted Environments," it is uniformly interpreted as follows: A2A (Agent-to-Agent) refers to the information interaction and collaborative decision-making process among multiple scheduling agents in a heterogeneous and untrusted network, centered around task state, node resources, trust state, and actionable space.

[0022] This method combines topology- and trust-aware graph neural network representations, and forms a cooperative scheduling scheme based on task-related security subgraph construction, trust-aware graph representation learning, action filtering, and primal-dual multi-agent policy optimization. First, a heterogeneous untrusted network model and a task model are constructed, and a task-related security subgraph is generated based on task requirements to filter out low-trust nodes, severely congested nodes, and infeasible links. Then, a trust-aware graph representation model is constructed, jointly encoding node local states, task requirements, neighbor trust information, and temporal evolution information to obtain a risk-aware state representation for scheduling decisions. Next, a security action filtering mechanism is constructed to mask actions that do not satisfy topology constraints and security subgraph constraints, generating an action space that satisfies hard constraints. Finally, the primal-dual multi-agent near-end policy optimization method is used to train and update the scheduling policy, outputting task acceptance, execution deployment, and secure routing results. This method includes the following steps: S1. For heterogeneous untrusted network infrastructure and differentiated task requirements, construct a task-customized joint constraint model of latency, security and routing to formalize the system modeling and scheduling problem of trust-computation coupling.

[0023] S11. Construct a heterogeneous untrusted network infrastructure model and a customized task model, and complete the basic formal definition of the scheduling scenario.

[0024] S111. Model heterogeneous untrusted infrastructure as a directed graph. ,in, It is a collection of all computing nodes in the network, covering three types of heterogeneous nodes: cloud servers, edge nodes, and routing gateways; This is a set of directed communication links between nodes. For each node... Define its core attributes, including nominal computing power. (Unit: CPU cycles / s), Real-time queue backpressure Dynamic trust value For each communication link Define its core attributes, including link bandwidth. One-way propagation delay Data packet loss probability .

[0025] S112, Regarding scheduling time slots t Heterogeneous task flows arriving within the region, with the task index set defined as follows: The corresponding task set is and single task The complete definition is a 7-tuple, which refers to seven basic attributes that provide a complete formal description of a single task to be scheduled. Specifically: ; The seven elements are as follows: This is the source node for the task. This represents the total computational load of the task. The amount of input data transmitted for the task. This represents the maximum end-to-end latency requirement for the task. This represents the minimum trust threshold requirement for the task nodes. This represents the minimum end-to-end transmission success rate requirement for the task. These seven elements collectively characterize the system utility weights after the task is completed. They represent the task's origin, resource requirements, data transmission requirements, latency constraints, security constraints, routing reliability constraints, and scheduling benefits. To simplify the formula, a time slot index is used. Only the task identifier is marked. The inherent attributes of the task remain fixed after the task is completed, and no additional time slot index is marked.

[0026] S113. Define the duration of a single scheduling time slot as follows: and for each node Corresponding to the neighbor node set This is the basic definition used for subsequent routing decisions and topology feature extraction.

[0027] S12. Construct a safe subgraph model for task dependencies to provide upfront feasible topological constraints for scheduling decisions.

[0028] S121, For each task Based on time t Within the globally observable state, construct a set of safe nodes that meets the customized safety and congestion requirements of the task, defined as: ; in, The congestion screening threshold for nodes is dynamically adapted based on the latency requirements of the task. The more stringent the latency requirements, the more conservative the threshold setting.

[0029] S122. Based on the set of secure nodes, construct a set of secure edges that meets the requirements of task transmission, defined as: ; in, Used to ensure the complete transmission of task data that can be completed within a unit time slot. This is used to ensure that the success rate of single-link transmission meets the end-to-end reliability requirements of the task.

[0030] S123. Generate a task-specific security subgraph based on the security node set and security edge set. This provides a conservative and feasible topology for subsequent scheduling decisions, eliminating interference from low-trust, high-congestion nodes and unreliable links at the source.

[0031] S124. For dynamic updates of the security subgraph, a minimum update interval and a triggered update mechanism are set. When the change in the trust value of a node in the network exceeds 20%, or the back pressure of the node queue exceeds the threshold, the reconstruction of the security subgraph is immediately triggered to ensure the consistency between topology constraints and the real-time state of the network.

[0032] S13. Construct a trust-aware service and queuing model to characterize the non-linear coupling relationship between security verification and computing service capabilities.

[0033] S131. Define a trust decay factor to address the computational resource consumption caused by continuous node security verification. This is used to quantify the impact of node trust values ​​on effective computing power, and the expression is: ; in, As the basic safety attenuation coefficient, , is a trust sensitivity parameter. The lower the node trust value, the more computing resources are used for security verification, and the more significant the decline in effective computing power.

[0034] S132. Based on the trust decay factor, the effective computing power and effective service rate of a computing node are expressed as follows: ; ； in, To enhance the real-time and effective computing capabilities of nodes, The effective processing rate of the node's tasks. This represents the maximum task processing rate under full node load.

[0035] S133. Define the three core decision variables for task scheduling, including task admission decision. Task execution and deployment decisions Task routing and forwarding decision ;in, This indicates that the task has been approved. This indicates that the task is in the node. implement, Indicates that the task is carried out through the link. S134, based on The queuing model calculates the average processing latency of nodes using the following expression: ; in, For real-time node utilization, Let the square of the coefficient of variation of the node service time be denoted by ; and the transmission delay expression for a single link be defined as follows: .

[0036] S14. Define end-to-end performance metric aggregation and scheduling constraints.

[0037] S141. Based on node processing delay and link transmission delay, the end-to-end total delay of the task is obtained by aggregation, expressed as: ; in, It is a binary indicator factor for node participation in a task, used to characterize whether a node participates in the forwarding or execution of a task.

[0038] S142. The end-to-end security cost and end-to-end reliability cost of the aggregation task are expressed as follows: ; ; Among them, end-to-end security costs Used to quantify the cumulative security risks and end-to-end reliability costs throughout the entire task path. Used to quantify the cumulative transmission failure risk throughout the entire task path.

[0039] S143. Define the hard constraints that scheduling decisions must satisfy, including: (1) Deployment and admission consistency constraints: ; Ensure that each admitted task has exactly one execution node.

[0040] (2) Node computation capacity constraints: ; Ensure that the computing load on a node does not exceed its effective computing capacity.

[0041] (3) Link bandwidth capacity constraints: ; Ensure that the transmission load of the link does not exceed its bandwidth limit.

[0042] (4) End-to-end DSR (Delay-Security-Routing) customized constraints: ; Ensure that all requirements for latency, security, and reliability of the mission are met.

[0043] In this embodiment, DSR represents three core constraint dimensions for task-customization requirements: Delay: The total end-to-end delay constraint experienced by a task from its creation at the source node to its completion. Security: Trust value of nodes in the task execution path, probability of security services, or cumulative security risk constraints; Routing: Constraints on the reachability of a task's transmission path in the network, link reliability, and end-to-end transmission success rate.

[0044] Therefore, DSR customized scheduling refers to jointly completing task acceptance, execution deployment, and secure routing decisions based on the latency, security, and path reliability requirements of different tasks. This embodiment has explicitly stated that the system must simultaneously meet "hard constraints on latency, security, and routing," and uses a security action filtering mechanism to pre-screen infeasible and unsafe actions.

[0045] S15. Formalization of the scheduling problem and Lagrange dual relaxation.

[0046] S151. The scheduling problem is formalized as a constrained long-term system utility maximization problem, with the objective function defined as: ; in, For global scheduling strategy, As a discount factor, For policy-based The mathematical expectation is to maximize the long-term cumulative system utility.

[0047] S152. Transform the long-term DSR constraint into a discount accumulation constraint function, and define the time delay constraint function respectively. Safety constraint functions Reliability constraint function This transforms the original problem into a stochastic optimization problem with long-term constraints.

[0048] S153, Introducing a nonnegative dual variable vector Lagrangian relaxation is applied to the long-term DSR constraints to construct the Lagrangian function: ; The stochastic optimization problem is transformed into a saddle point problem, which maximizes policy utility while minimizing dual penalty, providing a mathematical foundation for subsequent primal-dual reinforcement learning solutions.

[0049] S2. To achieve safe and compliant scheduling decision space constraints, a dynamic security subgraph of task dependencies is constructed, and a trust-gated graph neural network encoder is designed to learn a refined representation of topology and security dependencies from the dynamic network state.

[0050] S21. Construct a task-conditional decentralized partially observable Markov decision process (Dec-POMDP) ​​to provide a standardized Markov decision framework for distributed scheduling decisions.

[0051] S211. Define the seven-tuple form of Dec-POMDP as follows: ,in, For the set of decision nodes, For the global state space, For the local observation space of the node, For the local action space of the node, Let be the state transition probability function. For the reward function, This is the discount factor.

[0052] S212. Define the local observation space and the global state space. Make decisions for each task node. Local observation is defined as ,in This is the local state vector of the node. This is a local topology graph of the nodes. Customize the DSR configuration file for the task; the global state space is the union of all node local observations, the global adjacency matrix and the global task set, and is only open to the evaluator during the centralized training phase.

[0053] S213. Define the local action space of the joint encoding. For each decision... Define local scheduling actions ,in Indicates that the task refuses the action. Indicates the action to be performed locally by the task. This indicates that the task is forwarded to the neighboring node. The actions are combined to express the three types of decisions—task admission, deployment, and routing—through compact action coding.

[0054] S214. Define the constraint-aware reward function. Based on fixed dual variables. Define the local reward of a node as a single-slot realization of a Lagrange objective, expressed as: ; in, For nodes The set of tasks to be processed The value represents the violation amount of a single-slot DSR constraint; the global reward is the sum of the local rewards of all nodes, i.e. .

[0055] S22, Design Trust Gating The graph encoder enables a fusion representation of topology, trust state, and task requirements.

[0056] S221, Task-Conditioned Node Embedding and Fusion. For each active decision pair... Through a learnable multilayer perceptron By fusing the local state of nodes with the task DSR configuration file, a task-conditional node embedding is generated, expressed as: ; in, As a vector concatenation operator, it enables the same physical node to generate differentiated potential representations under different DSR task requirements through task-conditional fusion, thus adapting to customized scheduling needs.

[0057] S222, Trust-Gated Graph Attention Aggregation. To suppress unreliable message passing by low-trust neighbors while capturing structural dependencies in the local topology, a trust-gated aggregation coefficient is defined, expressed as: ; in, For learnable parameter vectors, It is the sigmoid activation function. The real-time trust value of the neighboring nodes; the passivated structural embedding calculation is performed based on the aggregation coefficient, and the expression is: ; in, , As a trainable weight matrix, a trust gating mechanism is used to attenuate the information of low-trust neighbors during the feature aggregation stage, thereby avoiding the interference of unreliable information on the representation quality.

[0058] S223. Temporal Dependency State Update. To capture the temporal dependencies of trust fluctuations and congestion relationships in the network, a Gated Recurrent Unit (GRU) is introduced to complete the iterative update of the temporal state. The expression is: ; in, This is the hidden state from the previous moment. This provides the node potential representation for the encoder's final output, offering reliable state input for subsequent action filtering and policy optimization.

[0059] S224. Representation Stability Constraint. A normalization constraint is imposed on the encoder's weight matrix to restrict the Lipschitz constant of gradient propagation. This ensures that the encoder's output representation remains stable under extreme scenarios such as dynamic attacks and abrupt changes in node states, avoiding representation shift and gradient explosion problems caused by low-trust nodes.

[0060] S23. Design a DSR safety action filtering mechanism to block infeasible and unsafe actions before policy sampling.

[0061] S231, Hierarchical construction of action sets. (For the current node) Tasks to be processed It employs a three-layer filtering logic of "topological legitimacy - security compliance - resource feasibility" to construct a set of actionable actions. The expression is: ; in, This is a set of topologically valid actions, ensuring that task rejection actions only take effect on the source node. For the set of compliant actions in a security subgraph, only locally executed and forwarded actions that conform to the constraints of the security subgraph are retained. The set of available actions for resources retains only the actions that meet the task requirements with the remaining computing / bandwidth resources.

[0062] S232, Additive Mask Matrix Construction and Action Log-Odds Filtering. This involves filtering the raw action log-odds output by the encoder. Construct an additive mask matrix, expressed as: ; The filtered action log odds are obtained based on the mask matrix: ; By using additive masking, the sampling probability of unsafe actions is made zero, thus avoiding unsafe actions and illegal decision generation during the strategy exploration phase from the root.

[0063] S233. Secondary Legality Verification and Caching Mechanism for Actions. A secondary legality verification is performed on the filtered feasible actions to ensure that the actions simultaneously meet both the real-time resource constraints of the node and the end-to-end constraints of the task. Simultaneously, time-slot-level caching is applied to the set of feasible actions for the same task and the same node. When the network state does not change significantly, the cached results are directly reused, avoiding redundant calculations and improving online decision-making efficiency.

[0064] S234. Prioritization of Actions. Within the set of actionable actions, they are dynamically prioritized according to the rule of "safety priority > latency priority > resource efficiency priority." This provides the policy network with a cost-effective set of action candidates, guiding the policy to converge quickly to the optimal feasible solution and improving the constraint satisfaction speed.

[0065] S24. Characterization quality assessment and feature optimization.

[0066] S241. Representation optimization based on contrastive learning. A contrastive learning strategy of sampling positive and negative sample pairs is adopted. Node representations within the same safe subgraph are used as positive samples, and low-trust node representations outside the safe subgraph are used as negative samples, thereby optimizing the feature discrimination ability of the encoder.

[0067] S242. Representation Consistency Constraint. For different node representations of the same task within the same time slot, a consistency loss function is set to ensure that the encoder maintains consistency in its representation of the same task requirements and similar node states, thereby improving the generalization ability of the policy.

[0068] S243. Feature Dimension Adaptation. For network topologies of different sizes, the hidden layer dimension and graph aggregation layer number of the encoder are adaptively adjusted to ensure efficient feature extraction and inference speed even in large-scale network scenarios.

[0069] S3. To improve the scheduling constraint satisfaction capability and decision stability in dynamic untrusted environments, a distributed multi-agent network is constructed. A DSR safe action filtering mechanism is introduced to pre-screen infeasible and unsafe actions. The original-dual multi-agent proximal policy optimization framework DSR-SafeG-MAPPO is constructed to complete model training and online inference deployment, and realize long-cycle multi-constraint collaborative scheduling decision-making.

[0070] S31. Construct the original-dual MAPPO optimization framework based on the centralized training and distributed execution (CTDE) paradigm.

[0071] S311, Distributed Actuator Network Design. Local actuators on all nodes share policy parameters. Based on the node representations output by the encoder and the filtered action / log odds, compliant scheduling actions are generated. The executor adopts a two-layer fully connected network structure, outputting the log odds of the corresponding action space. The policy parameters are updated by pruning the agent objective function, and the objective function expression is: ; in, For time slots t The set of all active task-node decision pairs within the system. The ratio of the probability of actions under the new strategy to that under the old strategy. For PPO trimming parameters, It is a generalized advantage estimate of dual perception, ensuring that the policy is optimized only within the actionable domain.

[0072] S312, Centralized Evaluator Network Design. The centralized evaluator is based on global state. The system estimates the constraint-aware state-value function to provide a global value assessment for policy updates. The evaluator employs a three-layer fully connected network structure, taking the global state vector as input and outputting the cumulative reward estimate for the corresponding state. Parameter updates are achieved by minimizing the mean squared error. The loss function expression is as follows: ; in, For the evaluator network parameters, The cumulative return target value for a single time slot is obtained through the order difference method.

[0073] S313. Iterative Update of Dual Variables. Based on the DSR constraint violation vectors aggregated in each iteration batch, the projection / gradient update of the dual variables is completed, expressed as: ; in, For the iteration round index, To update the step size for the dual variable, For the first DSR constraint violation vectors aggregated in round-iteration batches. The non-negative orthogonal projection operation ensures that the dual variable remains non-negative. Through adaptive dual variable updates, the penalty for persistently violated constraints is increased, while the penalty weight for satisfied constraints is reduced, thus achieving dynamic control of long-term constraints.

[0074] S314. Training stability optimization. Batch normalization is performed on the generalized dominance estimate to eliminate the magnitude difference in dominance values ​​between different batches; at the same time, the gradients of the executor and evaluator are pruned to limit the maximum range of the gradients, prevent gradient explosion during training, and improve the convergence stability of the framework.

[0075] S315. Construct a distributed multi-agent network to implement a coordination and scheduling strategy.

[0076] S3151. Multi-agent network topology construction. Edge nodes, cloud nodes, and routing gateways in a heterogeneous untrusted network are mapped as independent scheduling agents to form a distributed heterogeneous multi-agent network. The network topology is fully aligned with the physical network topology, and the communication links between agents are consistent with the physical communication links.

[0077] S3152. Roles and Functions of Intelligent Agents. The multi-agent network is divided into three categories: executing agents, routing agents, and collaborative decision-making agents. Executing agents are responsible for local task execution and resource allocation; routing agents are responsible for task forwarding and link scheduling; and collaborative decision-making agents are responsible for global constraint perception and issuing multi-agent collaborative instructions. These three types of agents achieve information exchange and collaborative decision-making through the multi-agent network.

[0078] S3153, Multi-agent Communication and Interaction Mechanism. A trust-gated agent communication protocol is designed, allowing only high-trust agents within a secure subgraph to interact. Interaction content includes node status, task requirements, and action sets. Local information sharing is achieved through a broadcast mechanism, and precise decision-making collaboration is accomplished through point-to-point communication, preventing low-trust agents from interfering with collaborative decision-making.

[0079] S3154. Multi-agent collaborative decision-making logic. Based on the CTDE structure of a multi-agent network, the execution agent and the routing agent independently generate local compliant actions, while the collaborative decision-making agent globally summarizes action information. Through primal-dual optimization, global utility is maximized and constraints are satisfied, achieving a scheduling effect of "distributed execution, centralized collaboration, and global optimization".

[0080] S32. Design a dual-timescale original-dual update strategy to achieve decoupling and synergy between strategy optimization and constraint control.

[0081] S321. Fast timescale original policy update. In each iteration's single-slot decision, based on fixed dual variables, each agent in the multi-agent network synchronously completes the rapid parameter update of the executor and evaluator, adapting to the real-time dynamic changes in network state, task flow, and trust value, ensuring the scheduling policy's rapid response capability to environmental changes.

[0082] S322. Slow Time-Scale Dual Variable Update. After each complete trajectory iteration, the dual variable is iteratively updated based on the constraint violation statistics of the entire trajectory. By using a time scale slower than the policy update, frequent fluctuations in the dual variable are avoided from interfering with multi-agent training, thus ensuring the stability of the training process.

[0083] S323. Adaptive adjustment of time scale. Based on the changing trend of constraint violation rate, the update step size and update frequency of dual variable are adaptively adjusted; when the constraint violation rate is continuously higher than the threshold, the update speed of dual variable is accelerated to quickly tighten constraints; when the constraint violation rate is stably lower than the threshold, the update speed of dual variable is slowed down to prioritize the optimization of system utility.

[0084] S33. Design the theoretical guarantee mechanism of the model to verify the convergence and security of the framework.

[0085] S331, Proof of Encoder Gradient Stability. Through spectral normalization constraints and trust gating mechanisms, it is proved that the encoder gradient propagation satisfies Lipschitz continuity, ensuring that the encoder gradient will not explode or vanish under dynamic attack scenarios, and that the output remains stable.

[0086] S332. Proof of Convergence of Dual Variables. Based on stochastic convex optimization theory, it is proved that the iterative update process of the dual variable satisfies the convergence condition of projective gradient descent, and the dual variable can converge to the optimal dual solution, thus strictly satisfying the long-term constraint.

[0087] S333, Proof of the Conservative Original Policy Improvement Boundary. Based on the performance difference lemma, it is proved that the policy update process can achieve a conservative improvement of the Lagrange objective, ensuring that the policy optimization process will not experience performance collapse and guaranteeing the monotonically non-decreasing convergence of the training process.

[0088] S334, Proof of Hard Constraint Satisfaction. Based on the safety subgraph pre-filtering and action masking mechanism, it is proved that all actions sampled by the policy satisfy the hard constraints of topology, safety, and resources, theoretically ensuring that no safety violations or constraint violations will occur during the scheduling process.

[0089] S34. Design of the end-to-end training and online inference deployment process for the model.

[0090] S341. Design of end-to-end training process for the model to achieve stable convergence of the framework.

[0091] S3411, Training Initialization. Initialize the multi-agent network parameters, actuator network, and evaluator network. SafeG- Encoder The encoder parameters are initialized with a zero vector as the dual variable, and training hyperparameters are set, including learning rate, number of iterations, trajectory length, pruning parameter, regularization coefficient, etc.

[0092] S3412, Trajectory Sampling. In each iteration, the environmental state is reset, and the multi-agent network executes the following sequentially according to the scheduling time slots: environmental state perception, safety subgraph construction, node feature encoding, action filtering, scheduling action execution, and reward and constraint violation signal collection, to complete the sampling of the complete trajectory and store the sampled data in the experience replay pool.

[0093] S3413. Policy and Value Network Update. Based on the sampled trajectory data, the generalized advantage estimate and reward target value are calculated. Within the multi-agent network, each agent simultaneously performs multiple rounds of gradient updates to the actuator network and the evaluator network to optimize the original policy.

[0094] S3414. Dual Variable Update. Based on the constraint violation statistics of the entire trajectory, complete the projection update of the dual variable and adjust the penalty weights of each constraint dimension.

[0095] S3415, Iteration Termination Judgment. Repeat steps S3412 to S3414 until the cumulative reward and constraint violation rate of the model converge, and output the multi-agent scheduling policy and dual variables after training.

[0096] S342. Online inference deployment process design to achieve real-time scheduling decisions.

[0097] S3421. Real-time environmental status perception. The multi-agent network collects real-time status information of each agent node, such as computing resources, queue backpressure, trust value, and link bandwidth, latency, and packet loss rate, and simultaneously collects the DSR customization requirements of arriving tasks.

[0098] S3422. Construction of Task-Specific Safety Subgraph. For each arriving task, a task-specific safety subgraph is dynamically constructed based on its real-time DSR requirements and network status, thus completing the pre-constraints on the multi-agent decision space.

[0099] S3423, Feature Encoding and Action Filtering. This is accomplished through training. ScaeG-Encoder The encoder generates task-conditional representations for each agent node, filters possible actions based on the safety subgraph and resource state, and generates a masked action space.

[0100] S3424, Multi-agent Distributed Scheduling Decision Generation. Within the multi-agent network, each agent outputs the optimal scheduling action based on node representations and the filtered action space. They collaboratively complete joint decisions regarding task admission, execution deployment, and routing forwarding, achieving end-to-end customized DSR constraint satisfaction.

[0101] S343, Model performance monitoring and online update mechanism.

[0102] S3431, Real-time Performance Monitoring. During online deployment, the multi-agent network monitors key performance indicators in real time, including task completion time and security, end-to-end latency, security violation rate, and system throughput.

[0103] S3432. Distribution Shift Detection. Based on online collected runtime data, detect the degree of shift between the network state, task flow distribution, and training scenario. When the distribution shift exceeds a threshold, online fine-tuning of the model is triggered.

[0104] S3433, Incremental Online Update. Based on newly collected scenario data, the parameters of the multi-agent network are incrementally fine-tuned. The model parameters are updated without interrupting online services to adapt to long-term changes in the network environment.

[0105] S344, Model generalization and robustness enhancement mechanism.

[0106] S3441. Multi-scenario data hybrid training. During training, scenario data of different network sizes, attack intensities, and task loads are mixed to improve the generalization ability of multi-agent networks in heterogeneous scenarios.

[0107] S3442, Adversarial Training Enhancement. Adversarial interference such as sudden changes in node trust values, link interruptions, and sudden task flows are injected during training to improve the robustness of multi-agent networks in extreme dynamic scenarios.

[0108] S3443, Incremental Fine-tuning Mechanism. For new network scenarios and task types, an incremental fine-tuning strategy is designed, updating only the top-level parameters of the encoder and the output layer parameters of the executor. This allows for rapid adaptation to new scenario requirements while retaining the basic multi-agent scheduling capabilities.

[0109] Therefore, the present invention adopts the above-mentioned customized scheduling method for latency-secure paths in heterogeneous and untrusted environments, which can improve the task constraint satisfaction, scheduling stability and secure completion capabilities in untrusted heterogeneous network environments. It can effectively solve the pain points of multi-constraint coupling and insufficient security assurance in heterogeneous and untrusted networks, and realize customized latency-security-routing collaborative intelligent task scheduling.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. 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 still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A time-delay-safe path customization scheduling method for A2A heterogeneous untrusted environments, characterized in that, Includes the following steps: S1. For heterogeneous untrusted network infrastructure and differentiated task requirements, a directed graph is used to formally model network nodes, communication links and their inherent attributes, and a seven-tuple is used to define the core parameters of heterogeneous tasks to construct a task-customized joint constraint model of latency, security and routing. The joint constraint model takes the network topology state, node trust level, link transmission characteristics and task-specific requirements as input to generate and dynamically update the task-specific security subgraph. Using a trust-aware service queuing model, output end-to-end performance metrics and scheduling hard constraint boundaries; S11. Model the heterogeneous untrusted infrastructure as a directed graph and construct a heterogeneous untrusted network infrastructure model; define the set of heterogeneous task flows and the single task seven-tuple representation within the scheduling time slot and construct a customized task model; set the scheduling time slot duration and the neighbor node set of each node to complete the basic formal definition of the scheduling scenario. S12. For each task, based on the global network state, select the set of nodes and edges that meet the requirements of security, congestion and transmission, generate a safe subgraph model that the task depends on, and set a triggered dynamic update mechanism to provide feasible topological constraints for subsequent scheduling decisions. S13. Quantify the impact of node trust value on effective computing power through trust decay factor, derive node effective computing power and service rate, define three types of decision variables: task admission, execution deployment and routing forwarding, calculate node processing latency and define link transmission latency based on M / G / 1 queuing model, and construct a trust-aware service and queuing model. S14. Define aggregation rules for end-to-end total latency, security cost, and reliability cost; set deployment and admission consistency, node computing capacity, link bandwidth capacity, and end-to-end DSR customized hard constraints; and clarify the performance indicators and constraint boundaries of scheduling decisions. S15. The scheduling problem is formalized into a stochastic optimization problem of maximizing system utility with long-term DSR constraints. Non-negative dual variables are introduced for Lagrange relaxation, which transforms the problem into a saddle point problem. S2. Based on the joint constraint model and task-specific security subgraph in S1, limit the range of legal topologies for scheduling; based on the end-to-end performance indicators and scheduling constraints output by S1, design a graph neural network encoder with a trust gating mechanism to learn a refined representation of topology and security dependencies from dynamic network states. S3. Based on the refined representation results obtained in S2, construct a distributed multi-agent network, introduce a DSR safety action filtering mechanism, and screen and retain compliant scheduling actions. We adopt a centralized training and distributed execution paradigm to construct a primitive dual multi-agent proximal policy optimization framework. We then use a dual time-scale primitive dual update strategy to perform model training and online inference deployment.

2. The time-delay-safe path customized scheduling method in an A2A heterogeneous untrusted environment according to claim 1, characterized in that, The specific steps of S11 are as follows: S111. Model the heterogeneous untrusted infrastructure as a directed graph and define the core attributes of each node and each communication link respectively. Among them, the core attributes of the node include nominal computing power, real-time queue back pressure and dynamic trust value, and the core attributes of the communication link include link bandwidth, one-way propagation delay and data packet loss probability. S112. For heterogeneous task flows arriving within the scheduling time slot, define a task index set and the corresponding task set, and define a single task as a complete seven-tuple. S113. Define the duration of a single scheduling time slot and construct a corresponding set of neighbor nodes for each node, which will be used for subsequent routing decisions and topology feature extraction.

3. The time-delay-safe path customized scheduling method in an A2A heterogeneous untrusted environment according to claim 1, characterized in that, The specific steps of S12 are as follows: S121. For each task, based on the globally observable state at the corresponding moment, construct a set of safe nodes that meet the customized safety and congestion requirements of the task. S122. Based on the set of secure nodes, construct a set of secure edges that meets the requirements of task transmission; S123. Based on the set of safe nodes and the set of safe edges, generate a task-specific safety subgraph to provide a conservative and feasible topology for subsequent scheduling decisions and eliminate interference from low-trust, high-congestion nodes and unreliable links. S124. For dynamic updates of the security subgraph, a minimum update interval and a triggered update mechanism are set. When the change in the trust value of a node in the network or the back pressure of the node queue exceeds the threshold, the reconstruction of the security subgraph is immediately triggered to ensure the consistency between topological constraints and the real-time state of the network.

4. The time-delay-safe path customization scheduling method for A2A heterogeneous untrusted environments according to claim 1, characterized in that, The specific steps of S13 are as follows: S131. Define the trust decay factor to quantify the impact of node trust value on effective computing power. S132. Based on the trust decay factor, the effective computing power and effective service rate of the computing node; S133. Define three types of core decision variables for task scheduling. The core decision variables include task admission decision, task execution deployment decision, and task routing and forwarding decision. S134. Based on the queuing model, calculate the average processing latency of the nodes.

5. The time-delay-safe path customized scheduling method in an A2A heterogeneous untrusted environment according to claim 1, characterized in that, The specific steps of S14 are as follows: S141. Based on node processing delay and link transmission delay, the end-to-end total delay of the task is obtained by aggregation; S142. The end-to-end security cost and end-to-end reliability cost of the aggregation task are used to quantify the cumulative security risk and cumulative transmission failure risk of the entire task path, respectively. S143. Define the hard constraints that scheduling decisions must meet, including deployment and admission consistency constraints, node computing capacity constraints, link bandwidth capacity constraints, and end-to-end DSR customization constraints.

6. The time-delay-safe path customization scheduling method for A2A heterogeneous untrusted environments according to claim 1, characterized in that, The specific steps of S15 are as follows: S151. The scheduling problem is formalized as a constrained long-term system utility maximization problem, and the objective function is defined. S152. Transform the long-term DSR constraint into a discount accumulation constraint function, define the time delay constraint function, the security constraint function and the reliability constraint function respectively, and transform the original problem into a stochastic optimization problem with long-term constraints. S153. Introduce a nonnegative dual variable vector, perform Lagrangian relaxation on the long-term DSR constraint, construct the Lagrangian function, and transform the stochastic optimization problem into a saddle point solution problem.

7. The time-delay-safe path customization scheduling method for A2A heterogeneous untrusted environments according to claim 1, characterized in that, The specific steps of S2 are as follows: S21. By defining the seven-tuple form of Dec-POMDP, the local observation space and the global state space, the jointly encoded local action space and the constraint-aware reward function, a decentralized partially observable Markov decision process with task conditionalization is constructed, providing a standardized Markov decision framework for distributed scheduling decisions. S22. By embedding and fusing task-conditional nodes, focusing attention aggregation in trust-gated graphs, updating temporally dependent states, and constraining representation stability, a trust-gated graph encoder is designed to achieve a fusion representation of topology, trust state, and task requirements. S23. By constructing a hierarchical set of actionable actions, constructing an additive mask matrix and filtering the log probability of actions, performing secondary verification of action legality and caching mechanism, and prioritizing actionable actions, a DSR safe action filtering mechanism is designed to block infeasible and unsafe actions before policy sampling. S24. Representation quality assessment and feature optimization are performed through representation optimization based on contrastive learning, representation consistency constraints, and feature dimension adaptation.

8. The time-delay-safe path customized scheduling method in an A2A heterogeneous untrusted environment according to claim 1, characterized in that, The specific steps of S3 are as follows: S31. By designing a distributed actuator network, a centralized evaluator network, iteratively updating dual variables, optimizing training stability, and constructing a distributed multi-agent network, a primitive dual MAPPO optimization framework based on a centralized training and distributed execution paradigm is constructed. S32. Design a dual-timescale original-dual update strategy, including fast-timescale original strategy update, full-timescale dual variable update and timescale adaptive adjustment, to achieve decoupling and synergy between strategy optimization and constraint control. S33. Based on spectral normalization and trust gating, stochastic convex optimization, performance difference lemma and safe subgraph filtering and action masking mechanism, configure gradient propagation constraints, dual variable update strategy, policy improvement rules and constraint verification logic for algorithm training to ensure encoder gradient stability, dual variable convergence, monotonic policy performance improvement and scheduling actions always satisfy hard constraints. S34. Through end-to-end model training, online inference deployment, performance monitoring and updates, and generalization robustness enhancement processes, the scheduling framework achieves stable convergence, real-time decision-making, continuous optimization, and reliable operation.

9. The time-delay-safe path customized scheduling method in an A2A heterogeneous untrusted environment according to claim 8, characterized in that, The specific steps of S34 are as follows: S341. Design the end-to-end process of the model, including training initialization, trajectory sampling, policy and value network update, dual variable update and iteration termination judgment, to achieve stable convergence of the framework. S342, Online inference deployment process design, including real-time environmental state perception, construction of task-specific security subgraphs, feature encoding and action filtering and multi-agent distributed scheduling decision generation, to achieve real-time scheduling decision; S343, Model performance monitoring and online update mechanism, including real-time performance monitoring, distributed offset detection and incremental online update; S344, Model generalization and robustness enhancement mechanisms, including multi-scenario data hybrid training, adversarial training enhancement, and incremental fine-tuning mechanisms.

Images