An infrastructure security protection method for power cloud edge-end collaborative architecture

Abstract

The application relates to the technical field of power system information security, and specifically discloses an infrastructure security protection method for a power cloud edge-end collaborative architecture, which first generates a multi-dimensional decentralized identifier fused with a traceability relationship and model parameter hash for a dynamic derived intelligent agent based on an elliptic curve encryption algorithm, combines a power grid operation environment and a historical credit feedback to issue a capability certificate bound with an authorization, and realizes fine-grained authorization isolation; intercepts cross-domain traffic through a protocol adaptation middleware and completes privacy protection authentication by using zero-knowledge proof, adapts a heterogeneous link to establish a controlled session; performs baseline deviation analysis and credit integral operation on session behavior through a global state synchronizer, feeds back credit logs and blocking instructions to the certificate issuing and traffic filtering links in real time to form a dynamic closed loop, simultaneously introduces a derived topological neighborhood deviation aggregation and nonlinear decay mechanism, and realizes early perception and adaptive fusing of a common-source intelligent agent collaborative attack.

Description

Technical Field

[0001] This application relates to the field of power system information security technology, and more specifically, to an infrastructure security protection method for a power cloud-edge-device collaborative architecture. Background Technology

[0002] With the deepening of the digital transformation of the new power system, cloud computing, edge computing, and artificial intelligence technologies have been widely deployed in core businesses such as load forecasting, dispatch optimization, and equipment operation and maintenance, building a multi-level AI service system across cloud, edge, and terminal. A large number of AI agents with autonomous decision-making capabilities are dynamically derived and collaboratively execute tasks across heterogeneous nodes. However, the identity and access management mechanisms of existing power information systems still follow the traditional model based on static API keys or fixed role tokens. This model is essentially designed for human users or fixed programs, and its trust anchor only includes simple identity identifiers, without covering multi-dimensional attribute information such as the capability boundaries of the agent, the integrity of the underlying model parameters, and hierarchical derivation and tracing relationships.

[0003] When a cloud-based global coordinating agent dynamically derives child agents from various substation edge nodes to address local power grid anomalies, traditional mechanisms, due to their coarse-grained authorization, often result in child agents inheriting all permissions from their parent. If these edge agents are compromised or hijacked, attackers can escalate privileges and even gain cross-domain access to core cloud data. Furthermore, the lack of a unified middleware that strongly binds identity and real-time session state in heterogeneous cross-protocol networks means that when the cloud detects abnormal operations by a child agent, the system cannot synchronously revoke all derived sessions across multiple communication protocols within milliseconds. This creates a risk of interface abuse and data leakage during the security revocation window. Moreover, when multiple sibling agents derived from the same contaminated parent source are distributed across different substation nodes, existing solutions only perform isolated reputation assessments on each agent, completely ignoring the trust contagion effect under the shared-source derivation topology. This makes it difficult to detect and block coordinated attacks in the early stages of lateral spread.

[0004] Therefore, an optimized infrastructure security protection solution for a power cloud-edge-device collaborative architecture is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an infrastructure security protection method for a power cloud-edge-device collaborative architecture, comprising: Step 1: Based on the elliptic curve cryptography algorithm, perform multi-dimensional cryptographic identity initialization and identifier generation on the agent's derived attribute data to obtain the agent's decentralized identifier and agent digital archive metadata. Step 2: Based on the regional power grid operation environment feature matrix and session reputation log, dynamically quantify the metadata of the agent's digital profile to obtain the quantification result, and bind the quantification result with the agent's decentralized identifier for permission signing to obtain the agent's capability certificate and certificate verification public key data. Step 3: After intercepting cross-domain access data packets and completing blacklist filtering in conjunction with access blocking instructions, the zero-knowledge proof computing circuit is used to perform privacy protection authentication on the agent's capability credentials and the agent's decentralized identifier to obtain the zero-knowledge authorization payload and cross-domain protocol feature identifier. Step 4: Perform cryptographic reverse verification on the zero-knowledge authorization payload based on the credential verification public key data, and complete heterogeneous protocol adaptation and session establishment based on cross-domain protocol feature identifiers to obtain a session authorization mapping table and a session state snapshot. Step 5: Perform behavioral baseline deviation analysis and reputation score estimation on the session authorization mapping table and session state snapshot to obtain session reputation logs and access blocking instructions.

[0006] Compared with existing technologies, this application proposes an infrastructure security protection method for a power cloud-edge-device collaborative architecture. First, based on elliptic curve cryptography, it generates a multi-dimensional decentralized identifier and digital profile for each dynamically derived agent, integrating parent tracing relationships, model parameter hashes, and lifecycle timestamps, making identity no longer a static label but a verifiable cryptographic profile. Then, it dynamically quantifies the agent's capabilities and issues verifiable credentials with bound permissions, combining the regional power grid operating environment and historical reputation feedback, achieving fine-grained authorization isolation. In the cross-domain access phase, it intercepts traffic through protocol adaptation middleware and uses zero-knowledge proof circuits to complete privacy-protected authentication, allowing agents to prove their access legitimacy without exposing original credentials. Simultaneously, it extracts protocol feature identifiers to adapt to heterogeneous communication links and establish controlled sessions. Finally, it performs baseline deviation analysis and reputation score calculation on session behavior through a global session state synchronizer, and feeds back reputation logs and blocking instructions in real time to the credential issuance and traffic filtering stages, forming a dynamic protection closed loop. It also introduces derived topology neighborhood deviation aggregation and nonlinear decay mechanisms to achieve early detection and adaptive circuit breaking of collaborative attacks by co-source agents. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1This is a flowchart of an infrastructure security protection method for a power cloud-edge-device collaborative architecture according to an embodiment of this application; Figure 2 This is a data flow diagram illustrating an infrastructure security protection method for a power cloud-edge-device collaborative architecture according to an embodiment of this application. Figure 3 This is a flowchart illustrating an infrastructure security protection method for a power cloud-edge-device collaborative architecture based on an elliptic curve cryptography algorithm, which initializes and generates multi-dimensional cryptographic identity of agent-derived attribute data to obtain decentralized agent identifiers and agent digital archive metadata. Figure 4 This is a flowchart illustrating a method for infrastructure security protection of a power cloud-edge-device collaborative architecture based on a regional power grid operating environment feature matrix and session reputation logs. The flowchart describes how the metadata of an agent's digital archive is dynamically quantified to obtain a quantification result, and how the quantification result is bound to the agent's decentralized identifier for permission signing to obtain the agent's capability certificate and certificate verification public key data. Figure 5 This is a flowchart illustrating a method for infrastructure security protection of a power cloud-edge-device collaborative architecture according to an embodiment of this application. After intercepting cross-domain access data packets and completing blacklist filtering with access blocking instructions, a zero-knowledge proof computing circuit is used to perform privacy protection authentication on the agent's capability credentials and the agent's decentralized identifier to obtain the zero-knowledge authorized payload and cross-domain protocol feature identifier. Figure 6 This is a flowchart illustrating a method for infrastructure security protection of a power cloud-edge-device collaborative architecture, based on an embodiment of this application, which performs behavioral baseline deviation analysis and reputation score estimation on a session authorization mapping table and session state snapshot to obtain session reputation logs and access blocking instructions. Figure 7 This is a flowchart illustrating a method for infrastructure security protection of a power cloud-edge-device collaborative architecture, based on an embodiment of this application. It describes a method for calculating historical reputation score decay based on the quantified value of behavioral baseline deviation to update the reputation score, and using a step indicator function and a circuit breaker threshold for over-limit comparison to determine the reputation decay score and high-risk behavior risk identifier. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] In existing power cloud-edge-device collaborative architectures, static identity authentication mechanisms cannot adapt to the dynamic derivation characteristics of AI agents, resulting in coarse-grained cross-domain access authorization and the inability to achieve millisecond-level global trust revocation. Furthermore, isolated reputation assessment ignores the trust contagion effect under co-originating topologies, making it difficult to detect collaborative attacks in their early lateral spread stages. Therefore, this application proposes an infrastructure security protection method for power cloud-edge-device collaborative architectures. This method generates decentralized identifiers and digital archive metadata for each dynamically derived agent based on elliptic curve cryptography, integrating parent tracing relationships and model parameter hashes, thus upgrading identity from static labels to verifiable cryptographic archives. It dynamically quantifies the agent's capabilities and issues permission-bound capability credentials based on regional power grid operating environment characteristics and historical reputation feedback, achieving fine-grained authorization isolation that adapts to environmental changes. Cross-domain traffic is intercepted through protocol-adaptive execution middleware, and privacy-preserving authentication is completed using zero-knowledge proof computation circuits, allowing agents to prove access legitimacy without exposing original credentials. After session establishment, baseline deviation analysis and reputation score decay calculations are performed on session behavior through a global state synchronizer, and the generated reputation logs and blocking instructions are fed back to the credential issuance and traffic filtering stages in real time, forming a dynamic protection closed loop. Furthermore, it introduces derived topology neighborhood deviation aggregation and an adaptive circuit breaker threshold mechanism based on topology density, enabling the system to perceive the infection risk between co-source agents and achieve precise blocking in the early stages of collaborative attacks.

[0014] Figure 1This is a flowchart of an infrastructure security protection method for a power cloud-edge-device collaborative architecture according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating an infrastructure security protection method for a power cloud-edge-device collaborative architecture according to an embodiment of this application. Figure 1 and Figure 2 As shown, an infrastructure security protection method for a power cloud-edge-device collaborative architecture according to an embodiment of this application includes: S1, performing multi-dimensional cryptographic identity initialization and identifier generation on agent-derived attribute data based on elliptic curve cryptography to obtain an agent decentralized identifier and agent digital profile metadata; S2, performing dynamic capability quantization on agent digital profile metadata based on the regional power grid operation environment feature matrix and session reputation logs to obtain quantization results, and binding the quantization results with the agent decentralized identifier for permission signature to obtain agent capability credentials and credential verification public key data; S3, blocking... After intercepting cross-domain access data packets and combining them with access blocking instructions to complete blacklist filtering, a zero-knowledge proof computing circuit is used to perform privacy-preserving authentication on the agent's capability credentials and decentralized agent identifier to obtain the zero-knowledge authorization payload and cross-domain protocol feature identifier; S4, based on the credential verification public key data, cryptographic reverse verification is performed on the zero-knowledge authorization payload, and heterogeneous protocol adaptation and session establishment are completed based on the cross-domain protocol feature identifier to obtain the session authorization mapping table and session state snapshot; S5, behavioral baseline deviation analysis and reputation score estimation are performed on the session authorization mapping table and session state snapshot to obtain the session reputation log and access blocking instructions.

[0015] Specifically, in step S1, based on the elliptic curve cryptography algorithm, multi-dimensional cryptographic identity initialization and identifier generation are performed on the agent's derived attribute data to obtain the agent's decentralized identifier and agent digital file metadata. It should be noted that, given that in the power cloud-edge-device collaborative architecture, the cloud-based global coordinating agent dynamically derives multiple edge sub-agents and distributes them to various substation nodes when performing scheduling or fault diagnosis tasks, traditional static identity mechanisms only contain simple identifiers (IDs) and do not cover multi-dimensional attributes such as the agent's parental origin relationship, model parameter integrity, and lifecycle, resulting in the inability to fully verify and track the identity of derived agents. Therefore, the technical solution of this application further uses the elliptic curve cryptography algorithm to perform multi-dimensional cryptographic identity initialization and identifier generation on the agent's derived attribute data, which includes the parent agent source identifier, derived task instruction code, model parameter hash value, and runtime lifecycle timestamp, to obtain the agent's decentralized identifier and agent digital file metadata. Through the above processing, the identity of an intelligent agent can be upgraded from a static label to a verifiable cryptographic file that integrates hierarchical traceability and model integrity information, providing a reliable foundation of trust for subsequent dynamic capability authorization and cross-domain authentication.

[0016] Figure 3 This document describes a flowchart illustrating an infrastructure security protection method for a power cloud-edge-device collaborative architecture, based on an elliptic curve cryptography algorithm. The flowchart describes the multi-dimensional cryptographic identity initialization and identifier generation of agent-derived attribute data to obtain a decentralized agent identifier and agent digital profile metadata. (See attached flowchart.) Figure 3 As shown, step S1 includes: S11, performing XOR perturbation and digest calculation on the parent source identifier and task instruction code in the agent derived attribute dataset to obtain a digest result, and fusing the digest result with the model parameter hash value and lifecycle timestamp in a tree structure to obtain a basic attribute tree structure; S12, performing asymmetric key pair generation and public key secure hash encoding conversion on the basic attribute tree structure to obtain the agent public key and agent decentralized identifier; S13, performing byte-level concatenation and splicing of the basic attribute tree structure and the agent public key, and using the agent decentralized identifier as the primary key index for standardized document encapsulation to obtain agent digital archive metadata.

[0017] In step S11, the parent source identifier and task instruction code in the agent's derived attribute dataset are XORed and digested to obtain a digest result. The digest result is then fused with the model parameter hash value and lifecycle timestamp in a tree structure to obtain the basic attribute tree structure. It should be noted that since the parent agent source identifier and derived task instruction code respectively carry the agent's lineage traceability information and business semantic information, directly concatenating the two makes it easy for attackers to reverse engineer the original identifier structure through pattern matching. Therefore, unpredictable perturbations need to be introduced at the bit level to enhance anti-reverse engineering capabilities. Based on this, the technical solution of this application first performs XOR perturbation and digest calculation on the parent source identifier and task instruction code in the agent's derived attribute dataset to obtain a digest result, and then fuses the digest result with the model parameter hash value and lifecycle timestamp in a tree structure to obtain the basic attribute tree structure. Through the above processing, discrete multidimensional attribute elements can be aggregated into a unified data structure with hierarchical verification capabilities, providing a high-entropy and tamper-proof input entropy source for the subsequent generation of asymmetric key pairs.

[0018] More specifically, in a concrete example of this application, the parent agent source identifier and derived task instruction code are first extracted from the agent derived attribute dataset. These are then aligned to equal-length bytes and subjected to a bitwise XOR operation, causing the two original data segments to perturb each other at the bit level, eliminating the statistical regularity of a single field. Subsequently, the output of the XOR operation is input into the SHA-256 secure hash algorithm for irreversible digest compression calculation, generating a fixed-length 256-bit digest result, which is the source feature hash value representing the derived lineage. Next, the model parameter hash value and runtime timestamp are extracted from the agent derived attribute dataset. Using the above digest result as the root node, and the model parameter hash value and runtime timestamp as the left and right leaf nodes respectively, hash concatenation and re-digestion operations are performed layer by layer upwards according to the Merkle tree construction rules, ultimately forming a basic attribute tree structure containing three attribute dimensions and possessing layer-by-layer integrity verification capabilities. When the data of any leaf node in this tree structure is tampered with, its root hash value will change, thus providing a tamper-proof structured input for subsequent key derivation.

[0019] In step S12, the basic attribute tree structure undergoes asymmetric key pair generation and public key secure hash encoding to obtain the agent's public key and decentralized identifier. It should be noted that since the basic attribute tree structure is only a static data organization form and does not yet possess cryptographic identity verification capabilities, it needs to be transformed into an asymmetric key system that can be used for signature verification and global addressing. Based on this, the technical solution of this application further performs asymmetric key pair generation and public key secure hash encoding on the basic attribute tree structure to obtain the agent's public key and decentralized identifier. Through the above processing, a unique and unforgeable cryptographic identity anchor can be established for each derived agent, enabling it to be independently verified and addressed in cross-domain interactions.

[0020] More specifically, in a concrete example of this application, the root hash value of the basic attribute tree structure is first input as a high-entropy seed into the key derivation function, and a modulo operation is performed on the order of the elliptic curve finite field to generate a private key scalar that conforms to the curve parameter specification. The calculation process is as follows:

[0021] in, This represents the generated private key of the intelligent agent. This represents the key derivation function. This represents the root hash value of the basic attribute tree structure. This represents the prime order of the selected elliptic curve group. Subsequently, the generated private key is multiplied by a predefined base point on the elliptic curve to obtain the corresponding agent's public key. The calculation process is as follows:

[0022] in, Represents the public key of the intelligent agent. This represents the base point of the elliptic curve. Finally, a second secure hash operation is performed on the agent's public key to hide the characteristics of the real public key, and the hash result is converted to Base58 encoding format to generate a globally unique and cross-domain resolvable decentralized identifier for the agent. The calculation process is as follows:

[0023] in, This represents a decentralized identifier for an intelligent agent. This represents the Base58 encoding conversion function. This indicates a secure hash algorithm. This identifier serves both as a global addressing index in cross-domain communication and, due to its derivation from the attribute tree structure, naturally binds to the agent's lineage and model integrity information.

[0024] In step S13, the basic attribute tree structure and the agent's public key are concatenated byte-wise, and standardized document encapsulation is performed using the agent's decentralized identifier as the primary key index to obtain the agent's digital profile metadata. It should be noted that since the basic attribute tree structure and the agent's public key still exist as separate data objects at this point, lacking a unified document encapsulation and indexing mechanism, they cannot be efficiently retrieved and fully invoked in subsequent credential issuance and cross-domain authentication processes. Based on this, the technical solution of this application further concatenates the basic attribute tree structure and the agent's public key byte-wise, and performs standardized document encapsulation using the agent's decentralized identifier as the primary key index to obtain the agent's digital profile metadata. Through the above processing, all identity elements of the agent can be integrated into a parsable, searchable, and structurally complete digital profile, providing a standardized data input source for subsequent dynamic capability quantification.

[0025] More specifically, in a concrete example of this application, the basic attribute tree structure is first serialized into a byte stream, and the agent's public key is also converted into a fixed-length byte sequence. Then, a concatenated concatenation operation is performed, with the tree structure byte stream first and the public key byte stream second, forming a continuous composite byte sequence. Next, this composite byte sequence is used as the main content of the document, with the agent's decentralized identifier as the unique primary key index field. It is standardized and encapsulated according to the W3C Distributed Identifier Document Specification. The document's verification method field points to the agent's public key, the server endpoint field records the agent's communication address information, and the metadata field includes parameters such as creation timestamp and lifecycle validity period. The final generated agent digital profile metadata is stored in a local trusted storage area using the agent's decentralized identifier as the key. Any subsequent node receiving this identifier can parse it to obtain its complete attribute information and public key verification materials.

[0026] Specifically, in step S2, based on the regional power grid operating environment feature matrix and session reputation logs, the metadata of the agent's digital profile is dynamically quantified to obtain the quantification result. The quantification result is then bound to the agent's decentralized identifier for permission signing to obtain the agent's capability certificate and certificate verification public key data. It should be noted that, given that traditional access control mechanisms often inherit all permissions from the parent agent during dynamic agent derivation, the authorization granularity is coarse and disconnected from the real-time power grid operating status. Once an edge agent is hijacked, it can achieve unauthorized cross-domain access through excessive authorization. Therefore, the technical solution of this application further uses the regional power grid operating environment feature matrix and session reputation logs to dynamically quantify the metadata of the agent's digital profile to obtain the quantification result. The quantification result is then bound to the agent's decentralized identifier for permission signing to obtain the agent's capability certificate and certificate verification public key data. Through the above processing, the access permission boundaries can be dynamically determined based on the current power grid load status and the agent's historical behavior reputation, achieving fine-grained and environment-adaptive capability authorization isolation.

[0027] Figure 4 This document describes a method for infrastructure security protection based on a power cloud-edge-device collaborative architecture, according to an embodiment of this application. It involves dynamically quantifying the metadata of an agent's digital profile to obtain a quantification result, and then binding the quantification result with the agent's decentralized identifier through permission signing to obtain the agent's capability certificate and certificate verification public key data. Figure 4As shown, step S2 includes: S21, performing retrieval matching and security threshold gating filtering on the session reputation log based on the agent's decentralized identifier, and performing attribute dimensionality reduction and task feature extraction from the agent's digital archive metadata to obtain the expected task feature vector; S22, performing high-order cross-feature fusion and adaptive weight dimensionality reduction on the expected task feature vector and the regional power grid operation environment feature matrix to obtain the dynamic capability limit parameter set; S23, based on the authoritative node's private key, performing concatenated binding and digital signature encapsulation on the dynamic capability limit parameter set and the agent's decentralized identifier to obtain the agent capability certificate and certificate verification public key data.

[0028] In step S21, the session reputation logs are searched and matched based on the agent's decentralized identifier, and security threshold gating filtering is applied. Attribute dimensionality reduction and task feature extraction are then performed from the agent's digital profile metadata to obtain the expected task feature vector. It should be noted that since malicious agents marked as having low reputation should not be eligible for capability credential issuance, a reputation gating mechanism is needed for pre-screening before feature extraction. Based on this, the technical solution of this application further performs retrieval and matching on the session reputation logs based on the agent's decentralized identifier, and security threshold gating filtering is applied. Attribute dimensionality reduction and task feature extraction are then performed from the agent's digital profile metadata to obtain the expected task feature vector. Through the above processing, agents that do not meet the reputation requirements can be blocked at the entry point of the credential issuance chain, while low-dimensional feature representations suitable for subsequent environmental constraint calculations are extracted for the selected agents.

[0029] More specifically, in a concrete example of this application, the decentralized identifier of the agent is first used as the primary key for retrieval. Index matching is performed in the session reputation log to extract the agent's comprehensive reputation score from the previous evaluation period. Then, a step indicator function is introduced as a security gating mechanism. The extracted reputation score is numerically compared with a preset security threshold. When the reputation score is lower than the security threshold, the indicator function outputs 0, directly setting subsequent features to zero and terminating the credential issuance process. When the reputation score is greater than or equal to the security threshold, the indicator function outputs 1, allowing entry into the feature extraction stage. After gating, the complete node information of the basic attribute tree structure is parsed from the agent's digital profile metadata. Principal component analysis is performed on the high-dimensional attributes, such as derived task instruction codes and model parameter hash values, to reduce dimensionality. The first few principal components that characterize the agent's core business capabilities are retained and combined into a fixed-dimensional task expectation feature vector. This vector numerically describes the resource requirements and business permission range of the scheduling or diagnostic tasks that the agent is expected to execute.

[0030] In step S22, a high-order cross-feature fusion and adaptive weight dimensionality reduction are performed on the task expectation feature vector and the regional power grid operating environment feature matrix to obtain a dynamic capability limit parameter set. It should be noted that since the actual capability authorization of an agent depends not only on its own expected task requirements but also on the real-time operating status of the current regional power grid, the authorization quota should be tightened during periods of high load or high alarm. Therefore, it is necessary to cross-fuse the two types of information to dynamically determine the capability boundary. Based on this, the technical solution of this application further performs high-order cross-feature fusion and adaptive weight dimensionality reduction on the task expectation feature vector and the regional power grid operating environment feature matrix to obtain a dynamic capability limit parameter set. Through the above processing, the authorization quota can be adaptively adjusted according to the power grid situation, avoiding security risks caused by excessively broad authorization during high-risk periods of the power grid.

[0031] More specifically, in a concrete example of this application, the expected feature vector of the task is first subjected to a Kronecker product operation with the feature matrix of the regional power grid operation environment to generate a high-order cross-feature tensor containing all possible combinations. Each element in this tensor corresponds to the product interaction of a certain dimension of the agent's task requirement feature and a certain dimension of the power grid's environmental state index, thereby capturing the nonlinear coupling relationship between the two. Subsequently, this high-order cross-feature tensor is input into a fully connected mapping layer with a nonlinear activation function, multiplied by a pre-trained environment adaptive weight matrix, and a bias term is added to perform dimensionality reduction and normalization processing. The calculation process is as follows:

[0032] in, This represents the set of dynamic capability limit parameters for the output. This indicates that the non-linear activation function is used to map the output to the effective quota range. This represents the pre-trained environment-adaptive weight matrix. This represents the expected feature vector of the task. This represents the Kronecker product operation. This represents the characteristic matrix of the regional power grid operating environment. This represents the bias term. After the above calculations, each dimension component in the output tensor is converted into a specific interface access permission level and resource quota value, which are combined to form a dynamic capability limit parameter set. This parameter set includes quantitative constraints such as the maximum range of API interfaces that the agent is allowed to call under the current power grid operating conditions and the maximum request frequency per unit time.

[0033] Specifically, the training process of the environment-adaptive weight matrix aims to minimize historical quota allocation errors. This is achieved by using verified safe and reasonable quota allocation records from past power infrastructure operation cycles as supervisory labels. The weight parameters are continuously adjusted through backpropagation to minimize the mean square error between the mapped output quota limit and the historical safe quota, given the same expected task features and grid environment characteristics as input. The nonlinear activation function is the Sigmoid function, whose output value is strictly constrained to a closed interval between 0 and 1. This naturally normalizes the components of each dimension in the dynamic capacity limit parameter set into proportional coefficients, which can then be multiplied by the corresponding physical upper limit of the resource dimension to convert into the actual interface call frequency limit or memory quota bytes. The update cycle and triggering conditions of the weight matrix follow a dual mechanism: firstly, periodic offline retraining is performed according to a fixed time window to adapt to the seasonal drift of grid load patterns; secondly, instant incremental updates are triggered when the alarm level indicators in the regional grid operation environment characteristic matrix exceed a preset change range for multiple consecutive sampling periods, ensuring that the weight parameters can promptly track sudden changes in the grid operation status.

[0034] In step S23, based on the authoritative node's private key, the dynamic capability limit parameter set and the agent's decentralized identifier are concatenated, bound, and digitally signed to obtain the agent's capability certificate and certificate verification public key data. It should be noted that since transmitting the dynamic capability limit parameter set in plaintext poses a risk of man-in-the-middle tampering with the quota value, it is necessary to bind it inseparably with the agent's identity through the authoritative node's digital signature to ensure the integrity and unforgeability of the certificate during cross-domain transfer. Based on this, the technical solution of this application further concatenates and binds the dynamic capability limit parameter set and the agent's decentralized identifier based on the authoritative node's private key, encapsulating them with a digital signature to obtain the agent's capability certificate and certificate verification public key data. Through the above processing, a tamper-proof certificate that locks the identity identifier and permission quota together can be generated, and the corresponding verification public key can be exported for use in subsequent authentication processes.

[0035] More specifically, in a concrete example of this application, the decentralized identifier of the agent and the dynamic capability limit parameter set are first subjected to a strict concatenation operation at the byte sequence level to form a continuous core payload data block to be signed. The concatenation order is fixed as identifier first, limit parameter set second, to ensure the consistency of data structure during issuance and verification. Subsequently, the private key held by the cloud global coordination center as the authoritative node is invoked to perform a SHA-256 secure hash operation on the core payload data block to generate a message digest. Then, an elliptic curve digital signature algorithm is performed on the digest result to generate an unforgeable authorization digital signature string. Next, the core payload data block and authorization digital signature string are encapsulated according to the standard structure of verifiable credentials to generate an agent capability credential, which records both the issuance timestamp and validity period fields. Finally, the public key material uniquely mathematically paired with the authoritative node's private key is extracted and exported synchronously, and encapsulated into independent credential verification public key data. This public key data will be loaded by the policy decision point of the target node in subsequent steps for reverse verification of the agent capability credential.

[0036] Specifically, in step S3, after intercepting cross-domain access data packets and completing blacklist filtering with access blocking instructions, a zero-knowledge proof computing circuit is used to perform privacy-preserving authentication on the agent's capability credentials and decentralized agent identifier to obtain the zero-knowledge authorization payload and cross-domain protocol feature identifier. It should be noted that, given that cross-domain interactions in the power cloud edge architecture involve multiple heterogeneous protocols such as HTTP and MQTT, directly transmitting the capability credentials in plaintext during authentication would expose sensitive information such as agent model parameters and permission quotas. Furthermore, malicious agents that have already triggered circuit breakers may continue to initiate requests; therefore, interception is necessary before authentication. Based on this, the technical solution of this application further intercepts cross-domain access data packets and, after completing blacklist filtering with access blocking instructions, uses a zero-knowledge proof computing circuit to perform privacy-preserving authentication on the agent's capability credentials and decentralized agent identifier to obtain the zero-knowledge authorization payload and cross-domain protocol feature identifier. Through the above processing, the mathematical proof of access legitimacy can be effectively completed without disclosing the original content of the credentials, and the underlying communication protocol features can be extracted simultaneously to support subsequent heterogeneous link adaptation.

[0037] Figure 5 This document describes a method for infrastructure security protection in a power cloud-edge-device collaborative architecture, based on an embodiment of this application. After intercepting cross-domain access data packets and completing blacklist filtering using access blocking instructions, a zero-knowledge proof computing circuit is used to perform privacy-preserving authentication on the agent's capability credentials and decentralized agent identifier to obtain the zero-knowledge authorized payload and cross-domain protocol feature identifier. (See flowchart for example.) Figure 5As shown, step S3 includes: S31, intercepting and parsing the cross-domain access data packets using the protocol adaptation execution middleware, and filtering them using blacklist Boolean matching based on the access blocking instruction to obtain compliant cross-domain access data packets; S32, performing feature encoding conversion of the underlying communication protocol header parameters and path parsing of the destination interface Uniform Resource Identifier on the compliant cross-domain access data packets to obtain the cross-domain protocol feature identifier and the destination interface addressing path; S33, loading the corresponding zero-knowledge proof computation circuit based on the destination interface addressing path, and using the agent capability certificate and the agent decentralized identifier as secret witness input to generate a privacy-preserving proof to obtain the zero-knowledge authorization payload.

[0038] In step S31, the protocol adaptation execution middleware intercepts and parses the source identifier of cross-domain access packets, and performs blacklist Boolean matching filtering based on the access blocking command to obtain compliant cross-domain access packets. It should be noted that since the access blocking command generated in step five records the identifier information of intelligent agents confirmed to be malicious, if these agents are not intercepted before the authentication calculation, the subsequent zero-knowledge proof circuit will consume computational resources for known malicious requests. Therefore, the technical solution of this application further intercepts and parses the source identifier of cross-domain access packets using the protocol adaptation execution middleware, and performs blacklist Boolean matching filtering based on the access blocking command to obtain compliant cross-domain access packets. Through the above processing, malicious traffic can be quickly eliminated at the forefront of the authentication chain with low computational overhead, allowing only compliant requests to enter subsequent processing stages.

[0039] More specifically, in a concrete example of this application, a protocol adaptation execution middleware is first deployed at the network ingress of the edge node. This middleware performs packet-level interception on all cross-domain access packets destined for core infrastructure using a bypass hooking method, temporarily storing the packets in a memory buffer. Then, deep parsing of the intercepted packets' headers is performed to extract the agent identifier field from the transport and application layer protocol headers. Next, the extracted source identifier field is input into a blacklist hash set constructed based on access blocking instructions to perform a Boolean matching operation to determine if the identifier exists in the blacklist set. If a match is found, the packet is determined to originate from a malicious agent that has already been blocked, and the packet is directly discarded from the memory buffer with corresponding resources released; if no match is found, it is determined to be a compliant request, the packet is released from the buffer and marked as a compliant cross-domain access packet, and passed to the subsequent protocol feature parsing and zero-knowledge authentication stages.

[0040] In step S32, the compliant cross-domain access data packet undergoes feature encoding conversion of the underlying communication protocol header parameters and path parsing of the destination interface's Uniform Resource Identifier to obtain the cross-domain protocol feature identifier and the destination interface addressing path. It should be noted that since the compliant cross-domain access data packet simultaneously carries both underlying communication protocol type information and destination interface address information—the former determining the protocol conversion path during subsequent session establishment, and the latter determining which zero-knowledge proof circuit to load—they need to be extracted separately to serve different downstream processing stages. Based on this, the technical solution of this application further performs feature encoding conversion of the underlying communication protocol header parameters and path parsing of the destination interface's Uniform Resource Identifier on the compliant cross-domain access data packet to obtain the cross-domain protocol feature identifier and the destination interface addressing path. Through the above processing, the protocol features and access targets in heterogeneous messages can be output in a structured manner, providing accurate indexes for subsequent zero-knowledge circuit loading and heterogeneous protocol adaptation.

[0041] More specifically, in a concrete example of this application, the underlying communication protocol header parameters are first extracted from the transport and application layer message headers of the compliant cross-domain access data packet. These parameters include information such as the protocol type identifier field, protocol version number, and transmission encoding method. Subsequently, a feature encoding conversion operation is performed on the extracted protocol header parameters, mapping the original header fields of different protocol types to a fixed-length encoded vector in a unified format, generating a cross-domain protocol feature identifier. This identifier, in the form of a structured string, represents the specific communication protocol type and its version parameters used in the current request. Simultaneously, a regular expression engine is used to parse the application layer payload area in the compliant cross-domain access data packet, extracting the Uniform Resource Identifier (URI) field of the destination interface. This field is then split level by level according to the path separator, preserving the complete interface path hierarchy, and generating the destination interface addressing path. This addressing path precisely points to the specific microservice interface node requested by the agent, and will be used as an index key to load the zero-knowledge proof computation circuit corresponding to the security policy of this interface from the trusted hardware library in the next step.

[0042] In step S33, the corresponding zero-knowledge proof computation circuit is loaded based on the destination interface addressing path. The agent's capability credential and decentralized agent identifier are used as secret witness inputs to generate a privacy-preserving proof, thus obtaining the zero-knowledge authorization payload. It should be noted that since the access permission constraint logic differs for different destination interfaces, a matching verification circuit needs to be loaded according to the specific addressing path. Furthermore, the model parameter hash and permission quota plaintext in the credential must not be exposed during the authentication process. Based on this, the technical solution of this application further loads the corresponding zero-knowledge proof computation circuit based on the destination interface addressing path, using the agent's capability credential and decentralized agent identifier as secret witness inputs to generate a privacy-preserving proof, thus obtaining the zero-knowledge authorization payload. Through the above processing, a mathematical proof that can be verified by a third party can be generated without disclosing the agent's original identity attributes and permission quota.

[0043] More specifically, in a concrete example of this application, firstly, using the destination interface addressing path as the index key, a zero-knowledge proof computation circuit pre-bound to the security policy of the interface is retrieved and loaded from the local trusted hardware library. This circuit defines the set of permission conditions required to access the interface in the form of an R1CS constraint relation matrix, with different interface paths corresponding to constraint circuits of different complexities. Subsequently, the agent's decentralized identifier and agent capability credentials are input into the circuit as the witness vector of the prover. The decentralized identifier is used to prove the legitimacy of the requester's identity, and the capability credentials are used to prove that the permission quota held by the agent meets the access threshold requirements of the destination interface. Inside the circuit, constraint satisfaction checks and encryption operations are performed on the witness vector based on a multinomial commitment mechanism to generate a zero-knowledge authorization payload. The calculation process is as follows:

[0044] in, This represents the generated zero-knowledge authorization payload. This represents the zero-knowledge proof generating function. This represents a zero-knowledge proof computation circuit loaded based on a destination interface addressing path. Represents the set of secret witness vectors. This represents a decentralized identifier for an intelligent agent. This payload represents the agent's capability credentials. It possesses completeness, reliability, and zero-knowledge, meaning that a legitimate agent can always generate a valid proof, any attacker without legitimate credentials cannot computationally forge a valid proof, and the verifier cannot extract any secret witness information from the payload other than the access legitimacy determination.

[0045] Specifically, in step S4, the zero-knowledge authorization payload is cryptographically reverse-engineered based on the credential verification public key data, and heterogeneous protocol adaptation and session establishment are completed based on cross-domain protocol feature identifiers to obtain a session authorization mapping table and a session state snapshot. It should be noted that since the zero-knowledge authorization payload is only a mathematical proof generated unilaterally by the prover, the target node needs to use the authoritative issuer's public key to reverse-engineer it to confirm its legitimacy. After successful verification, the node also needs to establish a link and record the session state according to the requested protocol type for subsequent behavior monitoring. Based on this, the technical solution of this application further performs cryptographically reverse-engineering on the zero-knowledge authorization payload based on the credential verification public key data, and completes heterogeneous protocol adaptation and session establishment based on cross-domain protocol feature identifiers to obtain a session authorization mapping table and a session state snapshot. Through the above processing, the final adjudication of the access request's legitimacy can be completed and a controlled session channel can be established. Simultaneously, the resource baseline at the moment of session establishment is collected to provide a reference for subsequent deviation analysis.

[0046] More specifically, in a concrete example of this application, the credential verification public key data is first used as the decryption parameter of the verifier. This data, along with the zero-knowledge authorization payload, is then input into an elliptic curve-based bilinear pairwise verification function to perform a cryptographic reverse verification. The calculation process is as follows:

[0047] in, Indicates the Boolean test result. This represents the zero-knowledge proof verification function. Indicates zero-knowledge authorization payload, This indicates the public key data being verified. When the pairing equation is true, it outputs a truth value of 1 indicating valid authorization; when the equation is false, it outputs a false value of 0 indicating invalid authorization.

[0048] Subsequently, the Boolean result is used as a logical gating multiplier. When the result is 0, all subsequent operations are interrupted and the connection request is rejected. When the result is 1, the protocol adaptation engine is activated. After activation, the cross-domain protocol feature identifier is used as an index to perform route matching operations in the pre-set heterogeneous protocol adaptation matrix to find the target-side protocol conversion rules and physical link allocation parameters corresponding to the current request protocol type. Based on the matching result, the gateway interface is called to establish the physical link. A globally unique temporary session identifier is allocated through a pseudo-random number generator. Simultaneously, the source IP, destination IP, source port, destination port, and transport layer protocol number are allocated in the network stack to form the network 5-tuple parameters.

[0049] Finally, the temporary session identifier and network quintuple parameters are structurally bound in trusted execution memory and written into a memory hash mapping structure in key-value pairs to generate a session authorization mapping table. This table records the complete network connection parameters and corresponding agent identity information of the currently active session. At the precise timestamp of the physical connection establishment, a bypass monitoring probe is triggered to perform slice-style data collection on the instantaneous CPU utilization, memory allocation rate, and request queue depth of the current infrastructure. The collection results are encapsulated into a multidimensional tensor to generate a session state snapshot, which will serve as the initial reference benchmark for subsequent behavioral baseline deviation analysis.

[0050] Specifically, in step S5, behavioral baseline deviation analysis and reputation score estimation are performed on the session authorization mapping table and session state snapshot to obtain session reputation logs and access blocking instructions. It should be noted that, given that the actual operational behavior of an agent after session establishment may deviate from the expected range declared by its credentials, and that existing solutions only perform isolated evaluations of each agent while ignoring the trust contagion effect under the co-source derived topology, it is necessary to continuously monitor the behavioral characteristics during the session and feed the results back to the front-end defense to form a dynamic protection closed loop. Based on this, the technical solution of this application further performs behavioral baseline deviation analysis and reputation score estimation on the session authorization mapping table and session state snapshot to obtain session reputation logs and access blocking instructions. Through the above processing, abnormal behavior of agents can be perceived in real time, and protection decisions can be fed back to the credential issuance and traffic filtering stages through reputation decay and circuit breaker mechanisms, forming a dynamic security protection closed loop covering the entire lifecycle.

[0051] Figure 6 This document presents a flowchart illustrating a method for infrastructure security protection in a power cloud-edge-device collaborative architecture, based on an embodiment of this application. The method involves performing behavioral baseline deviation analysis and reputation score estimation on a session authorization mapping table and session state snapshots to obtain session reputation logs and access blocking instructions. Figure 6 As shown, step S5 includes: S51, after receiving the session authorization mapping table and session state snapshot through the global session state synchronizer and writing them into the distributed collaborative ledger, performing multi-dimensional weighted deviation distance estimation on the session state snapshot relative to the preset normal operation baseline distribution parameters to obtain the behavioral baseline deviation quantification value; S52, performing historical reputation score decay calculation on the behavioral baseline deviation quantification value to update the reputation score, and using the step indicator function and the circuit breaker threshold to perform over-limit comparison judgment to obtain the reputation decay score value and high-risk behavior risk identifier; S53, persistently encapsulating the reputation decay score value in the log, and using the high-risk behavior risk identifier as the circuit breaker trigger, extracting malicious connection parameters from the session authorization mapping table in reverse and generating blocking messages to obtain the session reputation log and access blocking instruction.

[0052] In step S51, after receiving the session authorization mapping table and session state snapshot through the global session state synchronizer and writing them into the distributed collaborative ledger, a multi-dimensional weighted deviation distance estimation is performed on the session state snapshot relative to the preset normal operation baseline distribution parameters to obtain a behavioral baseline deviation quantification value. It should be noted that since the session authorization mapping table and session state snapshot are scattered across various edge nodes, they need to be aggregated into a global view to achieve unified state awareness across nodes, and the deviation of the current session behavior is quantified based on the normal operation baseline to provide a numerical basis for subsequent reputation decay calculations. Based on this, the technical solution of this application further obtains a behavioral baseline deviation quantification value by receiving the session authorization mapping table and session state snapshot through the global session state synchronizer and writing them into the distributed collaborative ledger, and then performing a multi-dimensional weighted deviation distance estimation on the session state snapshot relative to the preset normal operation baseline distribution parameters. Through the above processing, the scattered local session states can be aggregated into a globally consistent view, and the degree of deviation of the current session behavior from the normal baseline can be numerically characterized.

[0053] More specifically, in a concrete example of this application, the global session state synchronizer first pulls real-time session authorization mapping tables and session state snapshots from each edge node, and synchronizes them to the distributed collaborative ledger in an append-only manner to ensure state consistency and data immutability among multiple nodes. Subsequently, it retrieves baseline distribution parameters pre-trained based on historical normal operation data of the power infrastructure. These parameters contain the mean vector and covariance matrix of resource indicators for each dimension under normal conditions. Next, it performs Mahalanobis distance calculation on the session state snapshot relative to the baseline distribution parameters to overcome the influence of dimensional differences and feature correlations among multi-dimensional indicators such as CPU utilization, memory allocation rate, and request queue depth. The calculation process is as follows:

[0054] in, This represents the quantified value of the behavioral baseline deviation. Represents the current session state snapshot vector. This represents the mean vector of the preset normal operating baseline. This represents the inverse of the covariance matrix of the normal baseline distribution. A larger Mahalanobis distance value indicates that the current session behavior deviates further from the normal baseline, and the higher the probability of anomalies.

[0055] In step S52, the quantified behavioral baseline deviation value is subjected to historical reputation score decay calculation to update the reputation score. An over-limit comparison is then performed using a step indicator function and a circuit breaker threshold to obtain the reputation decay score and a high-risk behavior risk indicator. It should be noted that since the quantified behavioral baseline deviation value is only an instantaneous distance scalar, it cannot be directly used for access control decisions. It needs to be converted into a reputation score with a time-cumulative effect and compared with a defined circuit breaker threshold to output an executable blocking decision. Based on this, the technical solution of this application further performs historical reputation score decay calculation on the quantified behavioral baseline deviation value to update the reputation score, and uses a step indicator function and a circuit breaker threshold to perform over-limit comparison to obtain the reputation decay score and a high-risk behavior risk indicator. Through the above processing, continuous deviation observations can be converted into discrete risk judgment signals, providing a clear triggering basis for the generation of subsequent blocking instructions.

[0056] Figure 7 This document describes a method for infrastructure security protection in a power cloud-edge-device collaborative architecture, based on an embodiment of this application. It involves performing historical reputation score decay calculations on quantified behavioral baseline deviation values ​​to update reputation scores, and using a step indicator function and a circuit breaker threshold for over-limit comparison to determine the reputation decay score and high-risk behavior risk identifier. Figure 7 As shown, step S52 includes: S521, constructing derived topological adjacency relationships based on the parent source identifier field of each agent in the session authorization mapping table, and performing homogeneous neighborhood deviation aggregation on the behavior baseline deviation quantization value to obtain the topological contagion deviation aggregation value; S522, performing weighted linear fusion of individual local deviation and neighborhood contagion signal on the behavior baseline deviation quantization value and the topological contagion deviation aggregation value through a preset topological coupling coefficient to obtain the coupling deviation fusion value; S523, performing nonlinear reputation decay calculation on the coupling deviation fusion value to obtain the reputation decay integral value, and performing density inverse scaling and over-limit comparison judgment on the baseline circuit breaker threshold based on the number of active homogeneous agents in the session authorization mapping table to obtain the high-risk behavior risk identifier.

[0057] In step S521, a derived topology adjacency relationship is constructed based on the parent source identifier field of each agent in the session authorization mapping table. The behavioral baseline deviation quantification value is then aggregated using the same-source neighborhood deviation to obtain the topology contagion deviation aggregation value. It should be noted that existing reputation score decay calculations only perform isolated evaluations based on the behavioral baseline deviation quantification value of the currently evaluated agent, ignoring the common-source derived topology relationships between agents in the power cloud edge architecture. When an attacker poisons the model of a parent agent, all sibling agents derived from the same contaminated source may carry malicious parameters, and existing solutions are unaware of this common-source trust contagion effect spreading laterally along the derived topology. Therefore, the technical solution of this application further constructs a derived topology adjacency relationship based on the parent source identifier field of each agent in the session authorization mapping table, and aggregates the behavioral baseline deviation quantification value using the same-source neighborhood deviation to obtain the topology contagion deviation aggregation value. Through the above processing, a topological contagion risk association perception channel is effectively established between homogeneous derived agents, enabling the capture of latent signs of cooperative attacks through neighborhood aggregation signals in the early stages before a single edge agent independently triggers an alarm.

[0058] More specifically, in a concrete example of this application, the parent source identifier field encoded within the decentralized identifier of the agent bound to each active session is first parsed from the session authorization mapping table. An agent-derived topology adjacency matrix is ​​then constructed based on the same parent-child relationship. A matrix element of 1 indicates that two agents share the same parent-derived source and are thus topological sibling nodes, while an element of 0 indicates no direct association. In power dispatching and maintenance scenarios, this adjacency matrix comprehensively depicts the common-source trust coupling topology among all edge sub-agents distributed by the same cloud-based dispatching agent to various substations.

[0059] After establishing topological associations, the deviation values ​​of the current agent within the behavioral baseline deviation quantification are used as the center. An attention-weighted neighborhood aggregation mechanism is then used to weight and aggregate the deviation values ​​of all its topological neighboring agents. The attention weights are obtained by normalizing the deviation values ​​of neighboring agents using softmax, ensuring that sibling nodes with larger deviations contribute more to the current node's spread. The calculation process for the attention weights is as follows:

[0060] in, This represents the attention contagion weight coefficient of the topological sibling agent j to the currently evaluated agent i. This represents the quantized value of the baseline deviation of the behavior of the sibling agent j. This represents the first-order neighborhood set of the current agent i in the derived topology graph, i.e., the set of indices of all sibling agents that share the same parent source identifier. This indicates that the natural exponential function is used to map bias values ​​to positive probability weights. Subsequently, based on the aforementioned attention weights, a neighborhood-weighted aggregation operation is performed:

[0061] in, The output topological contagion bias aggregate value represents the collective risk signal perceived by the current agent from its derived family.

[0062] Through this aggregation operation, even if the deviation of a certain substation edge agent is still within a controllable range, if multiple of its sibling nodes simultaneously exhibit abnormal deviations, the amplification effect of the aggregation signal will enable it to receive an early warning of the topology-level infection risk. For example, when a cloud-based scheduling agent derives edge sub-agents from substations A, B, and C to handle regional load anomalies, if an attacker launches a model poisoning attack on the parent agent, the sub-agents of substations A and B may exhibit moderate deviations but have not independently triggered the circuit breaker threshold. At this time, the sub-agent of substation C can perceive the accumulation of abnormal signals from its sibling nodes through the aforementioned attention-weighted aggregation. Its topology infection deviation aggregation value will be amplified, thus providing an early warning at an early stage before the agent itself exhibits clear abnormal behavior. This advances the attack detection time from multiple rounds of cumulative detection in the traditional isolated evaluation mode to the same evaluation cycle in which the first batch of sibling nodes exhibits anomalies.

[0063] In step S522, a weighted linear fusion of individual local deviation and neighborhood contagion signal is performed on the behavioral baseline deviation quantification value and the topology contagion deviation aggregation value using a preset topology coupling coefficient to obtain a coupling deviation fusion value. It should be noted that relying solely on the agent's own behavioral baseline deviation quantification value or solely on the neighborhood topology contagion deviation aggregation value for a single-source, one-sided judgment has limitations. The former lacks a global topology perspective, while the latter may introduce misleading signals due to occasional fluctuations in individual sibling nodes. Therefore, it is necessary to quantitatively fuse individual local anomaly observations with global contagion signals to obtain a comprehensive risk criterion. Based on this, the technical solution of this application further uses a preset topology coupling coefficient to perform a weighted linear fusion of individual local deviation and neighborhood contagion signal on the behavioral baseline deviation quantification value and the topology contagion deviation aggregation value to obtain a coupling deviation fusion value. Through the above processing, the abnormal signals of sibling agents from the same source, dispersed across multiple substation edge nodes, are effectively aggregated into the individual evaluation dimension, enabling isolated edge nodes that originally only had a local perspective to acquire global topology-aware risk assessment capabilities.

[0064] More specifically, in a particular example of this application, a topological coupling coefficient is introduced to perform a weighted linear fusion of the behavioral baseline deviation quantization value and the topological contagion deviation aggregation value. The calculation process is as follows:

[0065] in, This represents the output coupling deviation fusion value. The topology coupling coefficient ranges from [0,1] and is used to control the impact of topology contagion signals on individual assessments. It is configured by the power system safety administrator based on business sensitivity. This represents the quantified value of the baseline deviation of the behavior of the agent currently being evaluated. This represents the aggregated topology contagion bias value obtained in the previous step. Through this fusion operation, the individual assessment dimension is injected with collective risk perception signals from the topological neighborhood, expanding the risk assessment perspective of edge nodes from local to the global derived family level. When the topology coupling coefficient is large, the contribution of the neighborhood contagion signal to the fusion result increases, making it suitable for core business scenarios such as power dispatching that are highly sensitive to coordinated attacks; when the topology coupling coefficient is small, the individual bias dominates, making it suitable for general operation and maintenance scenarios with low tolerance for false alarms.

[0066] For example, when the cloud-based scheduling agent derives edge sub-agents from substations A, B, and C respectively, and the topology coupling coefficient is configured to 0.4, if the baseline deviation quantization value of the sub-agent's own behavior is 0.3, and its topology contagion deviation aggregation value reaches 0.7 due to the accumulation of abnormal signals from the sibling nodes of substations A and B, then the fusion calculation result is 0.46, which is about 53% higher than 0.3 in the isolated evaluation mode. This fusion value will trigger a steeper integral decline in the subsequent nonlinear reputation decay stage, thus generating an effective early warning in the early stage of the collaborative attack spread.

[0067] In step S523, a nonlinear reputation decay calculation is performed on the coupling deviation fusion value to obtain a reputation decay integral value. Then, based on the number of active agents from the same source in the session authorization map table, a density inverse scaling and over-limit comparison are performed on the baseline circuit breaker threshold to obtain a high-risk behavior risk indicator. It should be noted that, given that simple linear subtraction cannot mathematically characterize the physical reality that topology contagion risk amplifies exponentially with association density, when multiple sibling nodes simultaneously exhibit moderate deviations, the penalty of linear decay on a single node is not strengthened by the existence of topological association, thus creating a security blind spot for a sluggish response to coordinated attacks. Furthermore, the fixed circuit breaker threshold cannot adaptively adjust with changes in the size of the same-source derived family, resulting in insufficient early warning sensitivity when the cloud-based scheduling agent dispatches sub-agents to a large number of substations at once. Based on this, the technical solution of this application further performs a nonlinear reputation decay calculation on the coupling deviation fusion value to obtain a reputation decay integral value, and performs a density inverse scaling and over-limit comparison on the baseline circuit breaker threshold based on the number of active agents from the same source in the session authorization map table to obtain a high-risk behavior risk indicator. Through the above processing, it is effectively ensured that the reputation score is almost unaffected under normal business fluctuations, thus avoiding frequent service interruptions caused by false alarms. At the same time, when topology contagion signals are superimposed, the reputation collapse rate accelerates exponentially, and the circuit breaker sensitivity of each individual is automatically tightened in large-scale homogeneous distribution scenarios.

[0068] More specifically, in a particular example of this application, the linear subtraction mode is abandoned, and a negative exponential decay function is used to update the credit score nonlinearly. The calculation process is as follows:

[0069] in, This represents the reputation decay score after the current time step update. This represents the historical credit baseline value of the previous assessment period. This indicates that the exponential decay sensitivity factor is used to control the steepness of reputation decline. This represents the negative exponential decay operator. This represents the coupling deviation fusion value obtained from the previous step. When the coupling deviation fusion value is small, the output of the negative exponential function approaches 1, meaning that the reputation hardly decays, corresponding to a scenario where normal fluctuations only cause slight penalties. However, when the coupling deviation fusion value increases sharply due to the superposition effect of topology contagion signals, the output of the negative exponential function rapidly approaches 0, causing the reputation score to plummet.

[0070] Subsequently, the total number of active sibling agents belonging to the same derived family of the current agent is counted from the session authorization mapping table and used as the topology density factor. This factor is then input into the adaptive threshold adjustment function to dynamically scale the fixed baseline threshold. The calculation process is as follows:

[0071] in, This represents the adaptive dynamic circuit breaker threshold after topology density adjustment. This indicates the system's preset fixed reference fuse threshold. The density scaling sensitivity coefficient is used to control how sensitive the increase in the number of siblings is to threshold compression. The topology density factor represents the total number of active sibling agents derived from the current agent. When the number of sibling agents approaches zero, the denominator approaches the threshold to maintain a baseline level. However, as the number of sibling agents continues to increase, the denominator increases accordingly, causing the threshold to be continuously suppressed. This reciprocal decay characteristic ensures the rigorous mathematical implementation of the defense strategy that the wider the impact, the stricter the tolerance: when the power dispatch cloud uniformly dispatches agents to a large number of substations, it automatically tightens the safety tolerance of each edge agent.

[0072] Finally, the coupling bias fusion value is compared with the adaptive dynamic circuit breaker threshold to output a high-risk behavior risk indicator. The calculation process is as follows:

[0073] in, This indicates a risk marker for high-risk behavior. This indicates that the step indicator function outputs 1 when the condition is true, triggering subsequent circuit breaker blocking logic; and outputs 0 when the condition is false, maintaining normal monitoring. This represents the coupling deviation fusion value obtained from the previous step. This indicates the adaptive dynamic circuit breaker threshold calculated in this step.

[0074] For example, when the cloud-based scheduling agent dispatches sub-agents to 20 substations at once, with a density scaling sensitivity coefficient configured to 0.1 and a baseline circuit breaker threshold of 0.5, the adaptive dynamic circuit breaker threshold is compressed to approximately 0.167, which is about 67% tighter than the fixed threshold of 0.5. If the coupling deviation fusion value of the sub-agent in substation C is 0.2, under the fixed threshold mode, this value will not trigger an alarm if it does not exceed 0.5. However, under the adaptive threshold mode, this value exceeds 0.167, immediately outputting a high-risk behavior flag and triggering circuit breaker blocking. This ensures that the trusted AI service for power systems has a level of protection granularity that matches the threat's reach when facing large-scale coordinated poisoning attacks.

[0075] In step S53, the reputation decay score is persistently encapsulated in a log, and malicious connection parameters are extracted from the session authorization mapping table and a blocking message is generated using a high-risk behavior risk identifier as the circuit breaker trigger to obtain the session reputation log and access blocking instruction. It should be noted that since the reputation decay score needs to be retrieved and called in the credential issuance stage of step two during subsequent evaluation cycles, and the connection parameters of the malicious agent need to be immediately converted into an executable blocking message and sent to the traffic filtering stage of step three after the high-risk behavior risk identifier is triggered, the entire dynamic protection feedback chain is closed. Based on this, the technical solution of this application further persistently encapsulates the reputation decay score in a log, and uses a high-risk behavior risk identifier as the circuit breaker trigger to extract malicious connection parameters from the session authorization mapping table and generate a blocking message to obtain the session reputation log and access blocking instruction. Through the above processing, the analysis results of behavior monitoring can be transformed into structured instructions that can be directly consumed by the front-end defense, forming a dynamic protection closed loop covering both credential issuance and traffic interception stages.

[0076] More specifically, in a concrete example of this application, the reputation decay score, along with the corresponding agent's decentralized identifier, the current evaluation timestamp, and the identifier of its derived topology family, is first structurally encapsulated and persistently stored as a session reputation log in an append-only manner. This log will be retrieved by the reputation gating step in the next evaluation cycle using the identifier as the primary key, serving as the core criterion for deciding whether to continue issuing capability credentials to the agent. Subsequently, a high-risk behavior risk identifier is used as a circuit breaker trigger switch. When the identifier value is 0, it indicates that the current session is within the security fluctuation range, and the access blocking instruction is set to an empty set without performing any blocking operation. When the identifier value is 1, the emergency circuit breaker processing logic is triggered. The temporary session identifier and network 5-tuple parameters corresponding to the malicious agent are reverse-parsed from the session authorization mapping table, and its source IP address, source port number, and agent decentralized identifier, etc., addressing features, are extracted. These features are combined with the physical disconnection instruction at the message level to generate an access blocking instruction. The instruction will be immediately sent to the blacklist set of the protocol adaptation execution middleware in step three, so that any cross-domain access requests subsequently initiated by the malicious agent will be directly discarded during the message interception stage, without consuming the computing resources of the zero-knowledge proof circuit.

[0077] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for infrastructure security protection in a power cloud-edge-device collaborative architecture, characterized in that, include: Step 1: Based on the elliptic curve cryptography algorithm, perform multi-dimensional cryptographic identity initialization and identifier generation on the agent derived attribute dataset to obtain the agent's decentralized identifier and agent digital archive metadata. Step 2: Based on the regional power grid operation environment feature matrix and session reputation log, dynamically quantify the metadata of the agent's digital profile to obtain the quantification result, and bind the quantification result with the agent's decentralized identifier for permission signing to obtain the agent's capability certificate and certificate verification public key data. Step 3: After intercepting cross-domain access data packets and completing blacklist filtering in conjunction with access blocking instructions, the zero-knowledge proof computing circuit is used to perform privacy protection authentication on the agent's capability credentials and the agent's decentralized identifier to obtain the zero-knowledge authorization payload and cross-domain protocol feature identifier. Step 4: Perform cryptographic reverse verification on the zero-knowledge authorization payload based on the credential verification public key data, and complete heterogeneous protocol adaptation and session establishment based on cross-domain protocol feature identifiers to obtain a session authorization mapping table and a session state snapshot. Step 5: Perform behavioral baseline deviation analysis and reputation score estimation on the session authorization mapping table and session state snapshot to obtain session reputation logs and access blocking instructions.

2. The infrastructure security protection method for a power cloud-edge-device collaborative architecture according to claim 1, characterized in that, The agent-derived attribute dataset includes the parent agent source identifier, derived task instruction code, model parameter hash value, and runtime lifecycle timestamp.

3. The infrastructure security protection method for a power cloud-edge-device collaborative architecture according to claim 2, characterized in that, Step one includes: XOR perturbation and digest calculation are performed on the parent source identifier and task instruction code in the agent derived attribute dataset to obtain the digest result. The digest result is then fused with the model parameter hash value and life cycle timestamp in a tree structure to obtain the basic attribute tree structure. Asymmetric key pair generation and public key secure hash encoding conversion are performed on the basic attribute tree structure to obtain the agent's public key and agent's decentralized identifier; The basic attribute tree structure is concatenated byte-level with the agent's public key, and the agent's decentralized identifier is used as the primary key index for standardized document encapsulation to obtain the agent's digital archive metadata.

4. The infrastructure security protection method for a power cloud-edge-device collaborative architecture according to claim 1, characterized in that, Step two includes: Based on the decentralized identifier of the agent, the session reputation log is retrieved, matched and filtered with security thresholds, and attribute dimensionality reduction and task feature extraction are performed from the metadata of the agent's digital archive to obtain the expected feature vector of the task. High-order cross-feature fusion and adaptive weight dimensionality reduction are performed on the expected feature vector of the task and the feature matrix of the regional power grid operation environment to obtain a dynamic capacity limit parameter set; Based on the authoritative node's private key, the dynamic capability limit parameter set and the agent's decentralized identifier are concatenated, bound, and digitally signed to obtain the agent's capability credentials and credential verification public key data.

5. The infrastructure security protection method for a power cloud-edge-device collaborative architecture according to claim 1, characterized in that, Step three includes: The protocol adaptation execution middleware intercepts and parses the source identifier of cross-domain access packets, and performs blacklist boolean matching filtering based on access blocking instructions to obtain compliant cross-domain access packets. The cross-domain access data packets are subjected to feature encoding conversion of the underlying communication protocol header parameters and path parsing of the destination interface Uniform Resource Identifier to obtain the cross-domain protocol feature identifier and the destination interface addressing path. Based on the destination interface addressing path, the corresponding zero-knowledge proof computation circuit is loaded, and the agent's capability certificate and the agent's decentralized identifier are used as secret witness inputs to generate a privacy-preserving proof to obtain the zero-knowledge authorization payload.

6. The infrastructure security protection method for a power cloud-edge-device collaborative architecture according to claim 1, characterized in that, Step four includes: Using the credential verification public key data as the decryption parameter of the verifier, the zero-knowledge authorization payload is subjected to bilinear pairwise cryptographic inverse verification based on elliptic curves, and the true or false judgment is output to obtain the verification Boolean result. The Boolean test result is used as a logical gating multiplier to activate the protocol adaptation engine. Based on the cross-domain protocol feature identifier, heterogeneous protocol conversion route matching and physical link connection allocation are performed to obtain temporary session identifiers and network quintuple parameters. The temporary session identifier is structured and bound to the network 5-tuple parameter and written into the memory hash map. At the time of connection establishment, system resource state slices are collected to obtain the session authorization mapping table and session state snapshot.

7. The infrastructure security protection method for a power cloud-edge-device collaborative architecture according to claim 1, characterized in that, Step five includes: After receiving the session authorization mapping table and session state snapshot by the global session state synchronizer and writing them into the distributed collaborative ledger, the session state snapshot is estimated by multidimensional weighted deviation distance relative to the preset normal operation baseline distribution parameters to obtain the behavioral baseline deviation quantification value. The historical credit score decay calculation is performed on the quantitative value of the behavioral baseline deviation to update the credit score, and the step indicator function is used to compare and determine the over-limit judgment with the circuit breaker threshold to obtain the credit decay score and high-risk behavior risk indicator. The reputation decay score is persistently encapsulated in the log, and the malicious connection parameters are extracted from the session authorization mapping table and the blocking message is generated using the high-risk behavior risk identifier as the circuit breaker trigger to obtain the session reputation log and access blocking instruction.

8. The infrastructure security protection method for power cloud-edge-device collaborative architecture according to claim 7, characterized in that, The historical credit score decay calculation is performed on the quantified behavioral baseline deviation value to update the credit score, and an over-limit comparison is performed using a step indicator function and a circuit breaker threshold to determine the credit decay score and high-risk behavior risk indicator, including: Based on the parent source identifier field of each agent in the session authorization mapping table, a derived topological adjacency relationship is constructed, and the same-source neighborhood deviation is aggregated to obtain the topological contagion deviation aggregation value. By using a preset topological coupling coefficient, the behavioral baseline deviation quantization value and the topological contagion deviation aggregation value are weighted linearly fused with individual local deviation and neighborhood contagion signal to obtain the coupling deviation fusion value. The coupling deviation fusion value is subjected to nonlinear reputation decay calculation to obtain reputation decay integral value, and the baseline circuit breaker threshold is determined by density inverse scaling and over-limit comparison based on the number of active agents of the same source in the session authorization mapping table to obtain high-risk behavior risk identification.

Images