Communication risk identification methods, devices, electronic equipment and storage media

By constructing intent description information and intent directed graphs, the problem of poor accuracy in identifying risks in traditional intelligent agent communication is solved, and efficient identification and accurate assessment of complex attacks in multi-agent collaborative systems are achieved.

CN123074750APending Publication Date: 2026-11-10BEIJING BAIDU NETCOM SCI & TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611047007.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Traditional methods for identifying communication risks in intelligent agent systems are inaccurate, and attackers can easily bypass the identification process by changing the wording, making it difficult to effectively identify communication risks in multi-agent collaborative systems.

Method used

By acquiring communication messages from the target agent, intent description information is constructed, intent evolution paths are determined, and risk assessment is performed based on the intent evolution paths. Lightweight and large language models are used for intent classification and structured extraction, a directed intent graph is constructed, and risk assessment results are generated, supporting single-agent and cross-agent recognition.

Benefits of technology

It improves the accuracy of communication risk identification, can identify complex attack behaviors in multi-agent collaborative systems, reduces computational overhead, and enables dynamic risk assessment and precise security control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074750A_ABST
    Figure CN123074750A_ABST
Patent Text Reader

Abstract

This disclosure provides a method, apparatus, electronic device, and storage medium for identifying communication risks, relating to artificial intelligence fields such as natural language processing, intelligent agents, large-scale models, information security, and knowledge graphs. The method may include: acquiring a communication message sent by a target intelligent agent, where the target intelligent agent is any intelligent agent in a multi-agent collaborative system; acquiring intent description information corresponding to the communication message; determining the intent evolution path of the target intelligent agent within a recently predetermined time period based on the intent description information; determining a risk assessment result of the communication message based on the intent evolution path; and determining whether the communication message is a risky message based on the risk assessment result. Applying the scheme described in this disclosure can improve the accuracy of the identification results, etc.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of natural language processing, intelligent agents, large models, information security, and knowledge graphs, and especially to communication risk identification methods, devices, electronic devices, and storage media. Background Technology

[0002] With the rapid adoption of multi-agent collaborative systems in enterprise applications, communication security between agents has become a critical bottleneck restricting the reliable deployment of such systems. In business scenarios such as financial transactions, enterprise operations, and intelligent maintenance, multiple agents with autonomous decision-making capabilities can complete complex tasks through message exchange. To improve communication security, risk identification of communication messages sent by agents is necessary. Traditional identification methods are mostly based on keyword filtering, such as using a predefined sensitive word library to perform real-time matching and filtering of communication messages between agents. However, the accuracy of this method is usually poor; attackers can easily bypass the identification by simply changing the wording. Summary of the Invention

[0003] This disclosure provides methods, apparatus, electronic devices, and storage media for identifying communication risks.

[0004] A method for identifying communication risks, comprising:

[0005] Acquire communication messages sent by a target intelligent agent, wherein the target intelligent agent is any intelligent agent in a multi-agent cooperative system;

[0006] Obtain the intent description information corresponding to the communication message, and determine the intent evolution path of the target agent within the most recent predetermined time period based on the intent description information;

[0007] The risk assessment result of the communication message is determined based on the intent evolution path, and whether the communication message is a risk message is determined based on the risk assessment result.

[0008] A communication risk identification device includes: a message acquisition module, an intent processing module, and a risk identification module;

[0009] The message acquisition module is used to acquire communication messages sent by the target intelligent agent, which is any intelligent agent in the multi-agent cooperative system.

[0010] The intent processing module is used to obtain intent description information corresponding to the communication message, and determine the intent evolution path of the target agent within the most recent predetermined time period based on the intent description information.

[0011] The risk identification module is used to determine the risk assessment result of the communication message based on the intent evolution path, and to determine whether the communication message is a risk message based on the risk assessment result.

[0012] An electronic device, comprising:

[0013] At least one processor; and

[0014] A memory communicatively connected to the at least one processor; wherein,

[0015] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described above.

[0016] A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the methods described above.

[0017] A computer program product includes a computer program / instructions that, when executed by a processor, implement the method described above.

[0018] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0019] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0020] Figure 1 This is a flowchart of an embodiment of the communication risk identification method described in this disclosure;

[0021] Figure 2 This is a flowchart of an embodiment of the method for generating intent description information as described in this disclosure;

[0022] Figure 3 A flowchart illustrating an embodiment of the method for determining an intention evolution path as described in this disclosure;

[0023] Figure 4 This is a flowchart of a first embodiment of the method for generating risk scores as described in this disclosure;

[0024] Figure 5 This is a flowchart of a second embodiment of the method for generating risk scores as described in this disclosure;

[0025] Figure 6 This is a flowchart illustrating an embodiment of the method for determining whether a communication message is a risky message as described in this disclosure;

[0026] Figure 7 This is a schematic diagram of the composition structure of the first embodiment 700 of the communication risk identification device described in this disclosure;

[0027] Figure 8 This is a schematic diagram of the composition structure of the second embodiment 800 of the communication risk identification device described in this disclosure;

[0028] Figure 9 A schematic block diagram of an electronic device 900 that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0029] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0030] Furthermore, it should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0031] Figure 1 This is a flowchart illustrating an embodiment of the communication risk identification method described in this disclosure. Figure 1 As shown, the specific implementation methods are as follows.

[0032] In step 101, the communication message sent by the target agent is obtained. The target agent is any agent in the multi-agent cooperative system.

[0033] In step 102, the intent description information corresponding to the communication message is obtained, and the intent evolution path of the target agent within the most recent predetermined time period is determined based on the intent description information.

[0034] In step 103, the risk assessment result of the communication message is determined based on the intention evolution path, and whether the communication message is a risk message is determined based on the risk assessment result.

[0035] By adopting the scheme described in the above method embodiments, the intent description information corresponding to the communication message to be processed can be obtained, and the temporal intent evolution path of the target intelligent agent can be constructed based on the intent description information. Then, risk identification can be completed based on the intent evolution path, that is, it can be determined whether the communication message is a risk message. In this way, the communication message is abstracted from a specific textual expression to a semantic representation at the intent level, avoiding the problem that attackers can bypass the identification by simply changing the wording in the traditional method, and improving the accuracy of the identification results.

[0036] The communication messages and intent descriptions in the embodiments described in this disclosure are not targeted at any specific user and are not intended to reflect the personal information of any specific user. The collection, storage, use, processing, transmission, provision, and disclosure of any type of information, such as user personal information, involved in the technical solutions of this disclosure comply with relevant laws and regulations and do not violate public order and good morals.

[0037] For the acquired communication messages, their corresponding intent description information can be obtained. In some embodiments of this disclosure, the communication messages can first be classified by intent to obtain intent classification results, and then the intent description information can be determined based on the communication messages and intent classification results.

[0038] By classifying communication messages into intent categories, and then combining the communication messages themselves with the intent classification results, the intent description information can be determined. This enables hierarchical intent extraction from coarse to fine granular levels. On the one hand, intent classification can provide clear guidance for subsequent element extraction. On the other hand, it can ensure the integrity and accuracy of intent description information by combining the original communication messages, thereby providing a reliable semantic foundation for subsequent intent evolution path construction, risk identification, and other tasks.

[0039] In some embodiments of this disclosure, before obtaining a communication message, an intent ontology tree can be generated. The intent ontology tree may include a strategic intent layer (L1), a business intent layer (L2), and an execution action layer (L3). The strategic intent layer may include at least two strategic intent nodes, each strategic intent node may correspond to a strategic intent category, each strategic intent node may belong to at least one business intent node in the business intent layer, and each business intent node may belong to at least one execution action node in the execution action layer. Accordingly, the communication message is classified into intents, and the way to obtain the intent classification result may include: determining the strategic intent category to which the communication message belongs from each strategic intent category through a lightweight classification model, and determining the strategic intent category to which it belongs as the desired intent classification result.

[0040] For example, a three-layer intent ontology tree can be pre-constructed based on the intent classification system defined by domain experts. The intent ontology tree can be composed of a node set and an edge set. The node set V = VL1∪VL2∪VL3, where VL1 represents the nodes in the strategic intent layer, VL2 represents the nodes in the business intent layer, and VL3 represents the nodes in the execution action layer. The edge set E represents the parent-child hierarchical relationship.

[0041] For example, an intent ontology tree fragment in a financial scenario might look like this:

[0042] L1: Financial

[0043] ├── L2: Payment

[0044] │ ├── L3: Application Programming Interface (API) Transfer

[0045] │ ├── L3: Enterprise Resource Planning (ERP) Payment

[0046] │ └── L3: Bank Transfer

[0047] ├── L2: Refund

[0048] │ ├── L3: Return via the original route

[0049] │ └── L3: Manual Refund

[0050] └── L2: Settlement

[0051] ├── L3: Supplier Settlement

[0052] └── L3: Salary Payment

[0053] Compared to child nodes under different parent nodes, child nodes under the same parent node usually have higher semantic similarity.

[0054] By adopting the above processing method, a three-layer hierarchical intent ontology tree including strategic intent layer, business intent layer and execution action layer can be generated, which can organize the originally flat and discrete intent categories into a tree structure with parent-child relationship, thereby providing a unified semantic space and standardized reference system for subsequent processing.

[0055] Accordingly, a lightweight classification model can be used to determine the strategic intent category to which the communication message belongs from each strategic intent category, and the strategic intent category to which the communication message belongs can be determined as the desired intent classification result.

[0056] Lightweight classification models can be lightweight pre-trained models, such as MiniLM (Miniature Language Model) and ModernBERT (Modern Bidirectional Encoder Representations from Transformers).

[0057] The input of a lightweight classification model can be a communication message, and the output can be the probability distribution corresponding to each strategic intent category. The strategic intent category corresponding to the highest probability value can be determined as the strategic intent category to which the communication message belongs.

[0058] For example, suppose there are four strategic intent categories: Financial, Data, Resource, and Identity, and suppose the communication message is "Please help me find the contact information and contract amount of all customers in East China". Then the probability distribution of the model output can be: {Financial: 0.05, Data: 0.92, Resource: 0.02, Identity: 0.01}. Accordingly, Data can be identified as the strategic intent category to which the communication message belongs.

[0059] By using the above processing method and employing a lightweight classification model, the computational power consumption and response latency of inference can be significantly reduced. Furthermore, by leveraging the semantic understanding capabilities of the lightweight classification model, the accuracy of the obtained intent classification results can be improved.

[0060] Based on the communication message and intent classification results, the intent description information corresponding to the communication message can be determined. In some embodiments of this disclosure, a large language model can be used to extract structured intent elements from the communication message based on the intent classification results, and the extraction results can be determined as the required intent description information.

[0061] The communication messages and intent classification results can be input into the large language model to obtain the structured intent element extraction results. The extraction results may include intent element triples.<goal,resource,target> In this context, "goal" represents the target action, "resource" represents the resources involved (i.e., the system resources involved), and "target" represents the object of action. Besides the intent element triplet, the extraction results may further include other information, such as intent attributes. Intent attributes may include data scope, destination, format, and method.

[0062] For example, if the communication message is "Please help me find the contact information and contract amounts of all customers in East China", then the extraction results could be:

[0063] {

[0064] "goal": "Read"

[0065] "resource": "Customer Relationship Management (CRM) database"

[0066] "target": "Customers in East China"

[0067] "scope": "Full amount (contact information + contract amount)"

[0068] "format": null

[0069] }

[0070] For example, if the communication message is "Export comma-separated (CSV) file containing all fields", then the extraction result could be:

[0071] {

[0072] "goal": "Export"

[0073] "resource": "query result set"

[0074] "target": "current data"

[0075] "scope": "All fields"

[0076] "format": "CSV"

[0077] }

[0078] By utilizing large language models to extract structured intent elements from communication messages, natural language communication messages can be transformed into structured intent elements, thereby converting unstructured text information into structured data that can be processed and analyzed later. At the same time, large language models have powerful semantic understanding and context awareness capabilities, which can accurately capture key information such as target actions, resources involved, and objects of action in communication messages, thereby improving the accuracy and comprehensiveness of the extracted results.

[0079] Based on the above introduction, Figure 2 This is a flowchart illustrating an embodiment of the method for generating intent description information as described in this disclosure. Figure 2As shown, the specific implementation methods are as follows.

[0080] In step 201, based on the intent ontology tree, a lightweight classification model is used to determine the strategic intent category to which the communication message belongs from each strategic intent category.

[0081] In step 202, the strategic intent category is determined as the intent classification result of the communication message.

[0082] In step 203, based on the intent classification results, a large language model is used to extract structured intent elements from the communication message.

[0083] In step 204, the extraction result is determined as the intent description information of the communication message.

[0084] After obtaining the intent description information, the intent evolution path of the target agent within the most recent predetermined time period can be further determined based on the intent description information.

[0085] In some embodiments of this disclosure, the intent path corresponding to the communication message can be determined first based on the intent description information, intent classification results, and intent ontology tree. Then, new path nodes can be generated based on the intent path, and the intent directed graph corresponding to the target agent can be updated based on the new path nodes. The intent directed graph includes path nodes and directed edges between nodes, and the intent evolution path can be determined based on the intent directed graph.

[0086] This allows communication messages to be mapped to standardized intent paths in the intent ontology tree, and the directed intent graph of the target agent can be updated in real time based on the generated path nodes. This ensures that the intent evolution history of each agent is continuously maintained in the form of a structured directed graph. Based on the directed intent graph, the system can trace the trajectory of the agent's intent changes in chronological order, thus providing a complete temporal data foundation for subsequent risk identification. At the same time, an incremental graph update strategy can be adopted, which only requires adding new path nodes and corresponding directed edges each time, without having to rebuild the entire graph, thereby reducing computational overhead and improving update efficiency.

[0087] In some embodiments of this disclosure, the method for determining the intent path corresponding to the communication message may include: determining the strategic intent node corresponding to the intent classification result as a first-level target node; determining the second-level target node that matches the intent description information from each second-level candidate node, wherein the second-level candidate node is each business intent node under the first-level target node; determining the third-level target node that matches the intent description information from each third-level candidate node, wherein the third-level candidate node is each execution action node under the second-level target node; and determining the path composed of the first-level target node, the second-level target node, and the third-level target node as the intent path.

[0088] Through the above processing, precise positioning can be achieved layer by layer from coarse-grained to fine-grained. Each layer matches within the range of child nodes of the target node in the previous layer, which not only narrows the matching range and reduces computational overhead, but also ensures the hierarchical consistency and semantic coherence of the path, avoiding semantic deviations caused by cross-layer jump mapping.

[0089] Based on the above introduction, Figure 3 This is a flowchart illustrating an embodiment of the method for determining the intended evolution path as described in this disclosure. Figure 3 As shown, the specific implementation methods are as follows.

[0090] In step 301, the strategic intent node corresponding to the intent classification result is determined as a first-level target node.

[0091] For example, if the intent classification result is "Data", then the strategic intent node corresponding to "Data" in the intent ontology tree can be determined as a target node at the first level.

[0092] In step 302, a second-level target node that matches the intent description information is determined from each second-level candidate node. The second-level candidate nodes are the business intent nodes under the first-level target node.

[0093] Assuming the second-level candidate nodes (corresponding node names) include: Read, Export, Aggregate, Share, and Transfer, then the cosine similarity between the embedding vector of the goal in the intent description information and the embedding vector of the node name of each second-level candidate node can be calculated. The second-level candidate node with the highest cosine similarity value can be determined as the second-level target node, such as the second-level target node being "Read".

[0094] In step 303, a three-layer target node that matches the intent description information is determined from each of the three-layer candidate nodes. The three-layer candidate nodes are the execution action nodes under the two-layer target nodes.

[0095] Assuming the three candidate nodes include database query, file read, and API retrieval, then by combining the goal and resource in the intent description information, the three target nodes can be determined. For example, if resource = CRM database, then based on semantic relevance, the three target nodes can be determined as "database query".

[0096] In step 304, the path consisting of the first-level target node, the second-level target node, and the third-level target node is determined as the intention path.

[0097] The intent path can be in the form of "L1→L2→L3", such as: Data→Read→database query.

[0098] In step 305, new path nodes are generated based on the intent path, and the intent directed graph corresponding to the target agent is updated based on the new path nodes.

[0099] The path nodes corresponding to the same agent can be organized into an intent directed graph according to the time sequence. Each path node can include: intent path, timestamp, and intent attribute (from intent description information). The directed edges between path nodes represent intent transition relationships.

[0100] For example, if the target agent sends four communication messages between 14:01 and 14:08, the corresponding directed graph of intent can be:

[0101] Node1 (14:01, Data→Read→Database Query, scope=full)

[0102] ↓ Transition (Time difference = 2 minutes)

[0103] Node2 (14:03, Data→Export→CSV export, scope=all fields)

[0104] ↓ Transition (Time difference = 2 minutes)

[0105] Node3 (14:05, Data→Package→Compressed Packaging, method=Compressed Archive Format (zip))

[0106] ↓ Transition (Time difference = 3 minutes)

[0107] Node4 (14:08, Data→Transfer→External Upload, dest=external cloud drive).

[0108] In step 306, the intention evolution path of the target agent within the most recent predetermined time period is determined based on the intention directed graph.

[0109] The specific value of the predetermined duration can be determined according to actual needs.

[0110] Since the current time is constantly changing, the time period corresponding to the most recent scheduled duration is also constantly changing, which is equivalent to maintaining a sliding time window. The current time can be used as the end point of the sliding time window, and the scheduled duration shifted backward from the current time can be used as the starting point of the sliding time window. The path nodes and directed edges between these path nodes within the intention directed graph can be extracted. The extracted path nodes and directed edges can then be combined in chronological order to obtain the intention evolution path of the target agent within the most recent scheduled duration.

[0111] Based on the intention evolution path, the risk assessment results of the communication message can also be determined.

[0112] In some embodiments of this disclosure, in response to determining that single agent identification is required, the risk assessment result can be directly determined based on the intent evolution path. In response to determining that cross-agent identification is required, the intent evolution path of the associated agent that is related to the target agent can be obtained within the most recent predetermined time period, and a joint intent chain can be generated based on the intent evolution paths of the target agent and the associated agent, and then the risk assessment result can be determined based on the joint intent chain.

[0113] There are no restrictions on how to determine whether single-agent recognition or cross-agent recognition is needed; for example, it can be determined according to preset rules.

[0114] In other words, the solution described in this disclosure supports both single-agent recognition and cross-agent recognition modes, thereby meeting the needs of different scenarios. In the single-agent recognition mode, it can independently judge the risk of individual operations based on its own intent evolution path. In the cross-agent recognition mode, it can build a joint intent chain among multiple agents, thereby effectively capturing the data flow behavior of multi-party collaboration, identifying collusion-based hidden risks, improving the risk identification dimension, and further enhancing the accuracy and comprehensiveness of the identification results.

[0115] In some embodiments of this disclosure, the method of directly determining the risk assessment result based on the intent evolution path may include: obtaining the matching degree between the intent evolution path and each abnormal path pattern in the pre-generated abnormal path pattern library, determining the weighting factor corresponding to the intent evolution path, determining the risk score based on the matching degree with the largest value and the weighting factor, and determining the risk score as the risk assessment result.

[0116] Accordingly, Figure 4 This is a flowchart of a first embodiment of the method for generating risk scores as described in this disclosure. Figure 4 As shown, the specific implementation methods are as follows.

[0117] In step 401, the matching degree between the intended evolution path and each abnormal path pattern in the pre-generated abnormal path pattern library is obtained.

[0118] The abnormal path pattern library can include multiple abnormal path patterns. For example, an abnormal path pattern is: Read / Full Data → Export / Full Fields → Package → Transfer / External Upload. This pattern belongs to the abnormal path pattern of data leakage.

[0119] There are no restrictions on how to obtain the matching degree between the intent evolution path and each abnormal path pattern. For example, the path nodes in the intent evolution path can be sequentially aligned with the path nodes in the abnormal path pattern to find the longest node sequence that is consistent in time order between the two, and the ratio of the number of nodes in the longest node sequence to the number of nodes in the abnormal path pattern can be obtained. This ratio is determined as the matching degree, and the value can be between 0 and 1.

[0120] In step 402, the weighting factor corresponding to the intended evolution path is determined.

[0121] For example, the intent attributes corresponding to each path node in the intent evolution path can be obtained separately. The intent attributes may include data range and target address, etc. Accordingly, the weighting factor corresponding to the target address and the weighting factor corresponding to the data range can be obtained separately.

[0122] For example, if the target address of any path node is an external target, then the weighting factor corresponding to the target address can be determined to a predetermined larger value, such as 1.5; otherwise, it can be determined to a predetermined smaller value, such as 0.8. If the data range of any path node is the entire data range, then the weighting factor corresponding to the data range can be determined to a predetermined larger value, such as 1.3; otherwise, it can be determined to a predetermined smaller value, such as 1.0.

[0123] In step 403, the risk score is determined based on the maximum matching degree and the weighting factor, and the risk score is determined as the risk assessment result of the communication message.

[0124] The maximum matching degree refers to the maximum value among the matching degrees corresponding to each abnormal path pattern.

[0125] In addition, the risk score can be:

[0126] Risk = Match * W1 * W2; (1)

[0127] Where Risk represents the risk score, Match represents the maximum matching degree, W1 represents the weighting factor corresponding to the data range, and W2 represents the weighting factor corresponding to the target address.

[0128] By employing the above processing method, and comparing the intent evolution path with various abnormal path patterns, it is possible to effectively identify intent evolution paths similar to known attack patterns. At the same time, by introducing a weighting factor, high-risk attributes such as full data or external targets can be reasonably reflected in the risk score, thereby effectively improving the accuracy of the obtained risk score.

[0129] For cross-agent recognition patterns, the intention evolution paths of related agents that are associated with the target agent can be obtained first within the most recent predetermined time period. Then, a joint intention chain can be generated based on the intention evolution paths of the target agent and related agents, and the risk assessment result can be determined based on the joint intention chain.

[0130] For example, by traversing the agent communication graph, associated agents that have a relationship with the target agent can be identified. The agent communication graph can be a structured topology graph used to describe the communication relationships between agents in a multi-agent collaboration scenario, and it can be updated at any time according to the actual situation. In addition, the agent communication graph can also include the trust score corresponding to each agent.

[0131] Based on the intent evolution paths of the target agent and associated agents, a joint intent chain can be generated. For example, for each intent evolution path, the nodes of each path can be aligned according to the interaction sequence, and associated behaviors such as mutual transmission and reading / writing can be matched and linked together to form a complete temporal sequence, that is, a joint intent chain covering the collaborative behaviors of multiple agents.

[0132] Furthermore, the risk assessment result can be determined based on the joint intent chain. In some embodiments of this disclosure, the matching degree between the joint intent chain and each abnormal path pattern in the pre-generated abnormal path pattern library can be obtained, and a cooperative agent group can be formed using the target agent and associated agents. The predetermined evaluation parameters corresponding to the cooperative agent group can be determined, and then the risk score can be determined based on the matching degree with the largest value and the predetermined evaluation parameters. The risk score can be determined as the risk assessment result of the communication message.

[0133] Accordingly, Figure 5 This is a flowchart of a second embodiment of the method for generating risk scores as described in this disclosure. Figure 5 As shown, the specific implementation methods are as follows.

[0134] In step 501, the matching degree between the joint intent chain and each abnormal path pattern in the pre-generated abnormal path pattern library is obtained.

[0135] The matching degree can be obtained in a similar way to step 401.

[0136] In step 502, a cooperative group of agents is formed by using the target agent and associated agents, and the predetermined evaluation parameters corresponding to the cooperative group of agents are determined.

[0137] For example, assuming the target agent is agent a, and the associated agents are agents b and c, then agents a, b, and c can be used to form an agent cooperation group, and the predetermined evaluation parameters corresponding to the agent cooperation group can be determined.

[0138] The specific parameters included in the predetermined evaluation parameters can be determined according to actual needs. For example, they may include collaboration anomaly degree and trust decay factor. Specifically, the number of times each agent in the agent collaboration group has historically performed the same or similar collaborations can be queried, and the collaboration anomaly degree can be determined based on the number of times; the fewer the number of times, the higher the collaboration anomaly degree. Additionally, the trust score of each agent in the agent collaboration group can be obtained, and the average of each trust score can be calculated. The trust decay factor can then be obtained by subtracting the average from 1.

[0139] In step 503, a risk score is determined based on the maximum matching degree and predetermined evaluation parameters, and the risk score is determined as the risk assessment result of the communication message.

[0140] The maximum matching degree refers to the maximum value among the matching degrees corresponding to each abnormal path pattern.

[0141] In addition, the risk score can be:

[0142] Risk'=β1*Match'+β2*CoopAnomaly+β3*TrustDecay; (2)

[0143] Where Risk' represents the risk score, Match' represents the maximum matching degree, CoopAnomaly represents the collaboration anomaly degree, TrustDecay represents the trust decay factor, and β1, β2 and β3 are normalized weight coefficients, and β1+β2+β3=1.

[0144] Assuming Match' is 0.82, CoopAnomaly is 0.85, TrustDecay is 0.70, and β1, β2, and β3 are 0.4, 0.35, and 0.25 respectively, then the risk score is 0.4*0.82 + 0.35*0.85 + 0.25*0.70 = 0.91.

[0145] Unlike single-point intrusion detection in traditional network security, communication risks in multi-agent collaborative systems exhibit three unique characteristics: diversity of intent expression (e.g., the same strategic goal can be achieved through different expressions), temporal combination of attacks (e.g., a single communication message is harmless, but a specific sequence constitutes an attack chain), and cross-agent distribution of execution (e.g., multiple agents each perform seemingly harmless operations, but collaborate to achieve malicious goals).

[0146] The scheme described in this disclosure can normalize communication messages with different expressions into a unified intention semantic space through a hierarchical intention ontology tree, thereby overcoming the recognition difficulties caused by the diversity of intention expressions. By constructing the intention evolution path of the agent within the most recent predetermined time and performing temporal pattern matching, the problem that a single communication message may be harmless but a specific sequence may constitute an attack chain is effectively overcome. Furthermore, by aligning the temporal evolution paths of the target agent and associated agents to construct a joint intention chain, and by generating a risk score by comprehensively considering matching degree, collaboration anomaly degree, and trust decay factor, it is possible to effectively capture distributed covert attack behaviors such as data leakage and split transfer completed by multi-agent collaboration, avoiding the problem of not being able to identify cross-step collaborative attack chains in traditional methods, and significantly improving the recognition accuracy and recognition coverage of combined attacks in multi-agent collaborative systems.

[0147] Based on the obtained risk score, it can be determined whether the communication message is a risky message.

[0148] In some embodiments of this disclosure, a risk score may be compared with a first threshold. In response to determining that the risk score is greater than or equal to the first threshold, a communication message may be determined as a risk message. In response to determining that the risk score is less than the first threshold, a communication message may be determined as a non-risk message.

[0149] By comparing risk scores with a first threshold, communication messages whose risk scores reach or exceed the first threshold are identified as risky messages. This provides a clear and quantifiable decision boundary for determining the risk level of communication messages, enabling automated risk classification. Furthermore, the size of the first threshold can be flexibly adjusted to adapt to the differentiated needs of different business scenarios regarding security sensitivity and false alarm tolerance. For example, when applied to scenarios with high security requirements, such as financial transactions and enterprise operations, the first threshold can be appropriately lowered to improve the recall rate of potential attacks. When applied to scenarios with higher tolerance, the first threshold can be appropriately increased to reduce false alarm interference, thereby achieving a dynamic balance between detection rate and false alarm rate, and improving the flexibility and scenario adaptability of risk identification.

[0150] In some embodiments of this disclosure, in response to determining that a communication message is a risky message and determining that the risk score is greater than or equal to a second threshold, the communication message can be blocked. If the second threshold is greater than a first threshold, in response to determining that a communication message is a risky message and determining that the risk score is less than the second threshold, alarm information for the communication message can be generated. In response to determining that a communication message is a non-risky message, the communication message can be allowed to pass.

[0151] The specific values ​​of the first and second thresholds can be determined according to actual needs. For example, the first threshold can be 0.5 and the second threshold can be 0.8. In addition, the first threshold can also be called the alarm threshold, and the second threshold can also be called the blocking threshold.

[0152] By setting a second threshold higher than the first threshold, risky messages can be further subdivided into two levels: blocking and alarm. Differentiated processing strategies can then be implemented, enabling precise hierarchical security control. Communication messages with extremely high risk scores that reach the blocking threshold can be directly blocked, promptly cutting off clearly identified high-risk attack links and effectively preventing serious consequences such as data leakage. For risky messages with risk scores between the alarm and blocking thresholds, an alarm can be used to notify the administrator for manual review. This avoids interference with normal business operations caused by excessive blocking while retaining a security fallback mechanism for manual intervention and judgment. Meanwhile, communication messages determined to be non-risky can be allowed to pass, ensuring zero-blockage flow of normal business messages.

[0153] Based on the above introduction, Figure 6 This is a flowchart illustrating an embodiment of the method for determining whether a communication message is a risky message as described in this disclosure. Figure 6 As shown, the specific implementation methods are as follows.

[0154] In step 601, it is determined whether the risk score is less than the first threshold. If so, step 602 is executed; otherwise, step 603 is executed.

[0155] In step 602, the communication message is allowed, and then the process ends.

[0156] In step 603, it is determined whether the risk score is greater than or equal to the second threshold. If so, step 604 is executed; otherwise, step 605 is executed.

[0157] In step 604, communication messages are blocked, and then the process ends.

[0158] In step 605, an alarm message for the communication message is generated and sent to the administrator, after which the process ends.

[0159] The alarm message can also carry risk path information, such as "Read / Full (14:01) → Export / Full Fields (14:03) → Package (14:05) → Transfer / External Upload (14:08), the corresponding mode is data leakage", so that the administrator can intuitively trace the attack chain, quickly judge the root cause of the risk and carry out disposal work.

[0160] Furthermore, after blocking communication messages, the trust score of the corresponding intelligent agent (such as the target intelligent agent, or the target intelligent agent and associated intelligent agents) in the agent communication graph can be updated synchronously. For example, 0.9 * 0.7 = 0.63, where 0.9 represents the trust score before the update, 0.63 represents the trust score after the update, and 0.7 represents the trust decay coefficient. Accordingly, when the corresponding intelligent agent subsequently interacts, a lower trust baseline will be introduced into the risk score calculation, improving the sensitivity to identify similar malicious collaborative behaviors and forming a long-term dynamic risk control constraint.

[0161] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this disclosure. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this disclosure. Furthermore, for parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0162] The above is an introduction to the method embodiments. The following describes the solution described in this disclosure further through device embodiments.

[0163] Figure 7 This is a schematic diagram of the structural composition of the first embodiment 700 of the communication risk identification device described in this disclosure. Figure 7 As shown, it includes: a message acquisition module 701, an intent processing module 702, and a risk identification module 703.

[0164] The message acquisition module 701 is used to acquire communication messages sent by the target intelligent agent, which can be any intelligent agent in the multi-agent cooperative system.

[0165] The intent processing module 702 is used to obtain the intent description information corresponding to the communication message and determine the intent evolution path of the target intelligent agent within the most recent predetermined time period based on the intent description information.

[0166] The risk identification module 703 is used to determine the risk assessment result of the communication message based on the intent evolution path, and to determine whether the communication message is a risk message based on the risk assessment result.

[0167] In some embodiments of this disclosure, the intention processing module 702 may obtain the intention description information corresponding to the communication message by: classifying the communication message into intentions, obtaining the intention classification result, and determining the intention description information based on the communication message and the intention classification result.

[0168] Figure 8This is a schematic diagram of the structural composition of the second embodiment 800 of the communication risk identification device described in this disclosure. Figure 8 As shown, it includes: message acquisition module 701, intent processing module 702, risk identification module 703, and preprocessing module 704.

[0169] The preprocessing module 704 can be used to generate an intent ontology tree, which may include a strategic intent layer, a business intent layer, and an execution action layer. The strategic intent layer may include at least two strategic intent nodes, each of which may correspond to a strategic intent category. Each strategic intent node may have at least one business intent node in the business intent layer, and each business intent node may have at least one execution action node in the execution action layer.

[0170] Accordingly, the intent processing module 702 performs intent classification on the communication message, and the intent classification result may be obtained by: determining the strategic intent category to which the communication message belongs from each strategic intent category through a lightweight classification model, and determining the strategic intent category to which the communication message belongs as the desired intent classification result.

[0171] In some embodiments of this disclosure, the intention processing module 702 may determine the intention description information based on the communication message and the intention classification result by: extracting structured intention elements from the communication message using a large language model based on the intention classification result, and determining the extraction result as the intention description information.

[0172] The intention processing module 702 determines the intention evolution path of the target agent within the most recent predetermined time period based on the intention description information in the following ways: determining the intention path corresponding to the communication message based on the intention description information, intention classification results and intention ontology tree, generating new path nodes based on the intention path, and updating the intention directed graph corresponding to the target agent based on the new path nodes. The intention directed graph includes path nodes and directed edges between nodes. The intention evolution path is determined based on the intention directed graph.

[0173] The intent processing module 702 determines the intent path corresponding to the communication message based on the intent description information, intent classification results, and intent ontology tree in the following ways: determining the strategic intent node corresponding to the intent classification results as a first-level target node; determining the second-level target node that matches the intent description information from each second-level candidate node, where the second-level candidate node is each business intent node under the first-level target node; determining the third-level target node that matches the intent description information from each third-level candidate node, where the third-level candidate node is each execution action node under the second-level target node; and determining the path composed of the first-level target node, the second-level target node, and the third-level target node as the intent path.

[0174] In some embodiments of this disclosure, the risk identification module 703 may determine the risk assessment result of the communication message based on the intent evolution path in the following ways: in response to determining that single agent identification is required, directly determine the risk assessment result based on the intent evolution path; in response to determining that cross-agent identification is required, obtain the intent evolution path of the associated agent that is related to the target agent within the most recent predetermined time period, generate a joint intent chain based on the intent evolution paths of the target agent and the associated agent, and determine the risk assessment result based on the joint intent chain.

[0175] In some embodiments of this disclosure, the risk identification module 703 may determine the risk assessment result directly based on the intent evolution path by: obtaining the matching degree between the intent evolution path and each abnormal path pattern in the pre-generated abnormal path pattern library, determining the weighting factor corresponding to the intent evolution path, determining the risk score based on the matching degree with the largest value and the weighting factor, and determining the risk score as the risk assessment result.

[0176] In some embodiments of this disclosure, the risk identification module 703 may determine the risk assessment result based on the joint intent chain by: obtaining the matching degree between the joint intent chain and each abnormal path pattern in the pre-generated abnormal path pattern library; forming an agent cooperation group using the target agent and associated agents; determining the predetermined evaluation parameters corresponding to the agent cooperation group; determining the risk score based on the matching degree with the largest value and the predetermined evaluation parameters; and determining the risk score as the risk assessment result.

[0177] In some embodiments of this disclosure, the risk assessment result may include a risk score. The risk identification module 703 may determine whether a communication message is a risk message based on the risk assessment result by: comparing the risk score with a first threshold; determining the communication message as a risk message in response to determining that the risk score is greater than or equal to the first threshold; and determining the communication message as a non-risk message in response to determining that the risk score is less than the first threshold.

[0178] In some embodiments of this disclosure, the risk identification module 703 may block the communication message in response to determining that the communication message is a risk message and that the risk score is greater than or equal to a second threshold. If the second threshold is greater than a first threshold, it may generate an alarm message for the communication message in response to determining that the communication message is a risk message and that the risk score is less than the second threshold. In response to determining that the communication message is a non-risk message, it may allow the communication message to pass.

[0179] The specific workflow of each of the above device embodiments can be found in the relevant descriptions in the foregoing method embodiments, and will not be repeated here.

[0180] Furthermore, the solutions described in this disclosure are applicable to a variety of scenarios and have broad applicability. For example, they can be applied to security monitoring scenarios of enterprise multi-agent collaboration platforms, enabling real-time monitoring of communication messages between agents and identifying data leakage attack chains, unauthorized privilege escalation, and unauthorized sharing of sensitive information in cross-departmental agent collaboration. Another example is their application to risk control scenarios in financial transaction multi-agent systems, enabling the identification of abnormal fund transfer chains involving multiple agents and split transfer attacks that bypass single transaction limits. Yet another example is their application to security scenarios of intelligent operation and maintenance agent clusters, enabling the identification of covert attack sequences after agents are hijacked, such as "scanning network topology → collecting credentials → lateral movement → installing backdoors." Finally, they can be applied to compliance monitoring scenarios of multi-agent customer service systems, enabling the identification of unauthorized transfer patterns of user privacy data involved in communication between agents, such as "querying user personal information → exporting user profiles → transmitting to unauthorized agents."

[0181] The solutions described in this disclosure can be applied to the field of artificial intelligence, particularly in areas such as natural language processing, intelligent agents, large-scale models, information security, and knowledge graphs. Artificial intelligence is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It involves both hardware and software technologies. Artificial intelligence hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. Artificial intelligence software technologies mainly include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0182] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0183] Figure 9 A schematic block diagram of an electronic device 900 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0184] like Figure 9As shown, the electronic device 900 includes a computing unit 901, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. The RAM 903 may also store various programs and data required for the operation of the electronic device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0185] Multiple components in electronic device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of displays, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows electronic device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0186] The computing unit 901 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as those described in this disclosure. For example, in some embodiments, the methods described in this disclosure can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the methods described in this disclosure can be performed. Alternatively, in other embodiments, the computing unit 901 may be configured to perform the methods described herein by any other suitable means (e.g., by means of firmware).

[0187] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0188] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0189] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0190] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a cathode ray tube (CRT) or liquid crystal display (LCD) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0191] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0192] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0193] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0194] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for identifying communication risks, comprising: Acquire communication messages sent by a target intelligent agent, wherein the target intelligent agent is any intelligent agent in a multi-agent cooperative system; Obtain the intent description information corresponding to the communication message, and determine the intent evolution path of the target agent within the most recent predetermined time period based on the intent description information; The risk assessment result of the communication message is determined based on the intent evolution path, and whether the communication message is a risk message is determined based on the risk assessment result.

2. The method according to claim 1, wherein, The step of obtaining the intent description information corresponding to the communication message includes: The communication messages are classified according to their intents to obtain the intent classification results; The intent description information is determined based on the communication message and the intent classification result.

3. The method according to claim 2, further comprising: Before acquiring the communication messages sent by the target intelligent agent, an intent ontology tree is generated; The intent ontology tree includes a strategic intent layer, a business intent layer, and an execution action layer. The strategic intent layer includes at least two strategic intent nodes, each strategic intent node corresponds to a strategic intent category, each strategic intent node belongs to at least one business intent node in the business intent layer, and each business intent node belongs to at least one execution action node in the execution action layer. The step of classifying the communication message to obtain the intent classification result includes: determining the strategic intent category to which the communication message belongs from each strategic intent category using a lightweight classification model, and determining the strategic intent category to which the communication message belongs as the intent classification result.

4. The method according to claim 2, wherein, The step of determining the intent description information based on the communication message and the intent classification result includes: Based on the intent classification results, a large language model is used to extract structured intent elements from the communication message, and the extraction results are determined as the intent description information.

5. The method according to claim 3, wherein, Determining the intent evolution path of the target agent within the most recent predetermined time period based on the intent description information includes: Based on the intent description information, the intent classification result, and the intent ontology tree, the intent path corresponding to the communication message is determined; New path nodes are generated based on the intent path, and the intent directed graph corresponding to the target agent is updated based on the new path nodes. The intent directed graph includes path nodes and directed edges between nodes. The intention evolution path is determined based on the intention directed graph.

6. The method according to claim 5, wherein, Determining the intent path corresponding to the communication message based on the intent description information, the intent classification result, and the intent ontology tree includes: The strategic intent node corresponding to the intent classification result is determined as a first-level target node; From each of the second-level candidate nodes, determine the second-level target node that matches the intent description information. The second-level candidate nodes are the business intent nodes under the first-level target node. From each of the three-layer candidate nodes, a three-layer target node that matches the intent description information is determined, wherein the three-layer candidate nodes are the execution action nodes under the two-layer target node; The path consisting of the first-level target node, the second-level target node, and the third-level target node is determined as the intent path.

7. The method according to claim 1, wherein, The risk assessment result of the communication message determined based on the intent evolution path includes: In response to the determination that single agent recognition is required, the risk assessment result is determined directly based on the intent evolution path; In response to the determination that cross-agent identification is required, the intention evolution path of the associated agent that is related to the target agent is obtained within the most recent predetermined time period. A joint intention chain is generated based on the intention evolution paths of the target agent and the associated agent. The risk assessment result is determined based on the joint intention chain.

8. The method according to claim 7, wherein, The method of directly determining the risk assessment result based on the intent evolution path includes: The matching degree between the intent evolution path and each abnormal path pattern in the pre-generated abnormal path pattern library is obtained respectively; Determine the weighting factor corresponding to the intended evolution path; The risk score is determined based on the maximum matching degree and the weighting factor, and the risk score is determined as the risk assessment result.

9. The method according to claim 7, wherein, The determination of the risk assessment result based on the joint intent chain includes: The matching degree between the joint intent chain and each abnormal path pattern in the pre-generated abnormal path pattern library is obtained respectively; By using the target agent and the associated agents to form an agent cooperation group, the predetermined evaluation parameters corresponding to the agent cooperation group are determined; The risk score is determined based on the maximum matching degree and the predetermined evaluation parameters, and the risk score is determined as the risk evaluation result.

10. The method according to claim 1, wherein, The risk assessment results include: risk score; Determining whether a communication message is a risky message based on the risk assessment result includes: The risk score is compared with a first threshold. In response to determining that the risk score is greater than or equal to the first threshold, the communication message is identified as the risk message; In response to determining that the risk score is less than the first threshold, the communication message is identified as a non-risk message.

11. The method of claim 10, further comprising: In response to determining that the communication message is the risk message and determining that the risk score is greater than or equal to a second threshold, the communication message is blocked, wherein the second threshold is greater than the first threshold; In response to determining that the communication message is the risk message and that the risk score is less than the second threshold, an alarm message is generated for the communication message; In response to determining that the communication message is a non-risk message, the communication message is allowed to proceed.

12. A communication risk identification device, comprising: The module includes a message acquisition module, an intent processing module, and a risk identification module. The message acquisition module is used to acquire communication messages sent by the target intelligent agent, which is any intelligent agent in the multi-agent cooperative system. The intent processing module is used to obtain intent description information corresponding to the communication message, and determine the intent evolution path of the target agent within the most recent predetermined time period based on the intent description information. The risk identification module is used to determine the risk assessment result of the communication message based on the intent evolution path, and to determine whether the communication message is a risk message based on the risk assessment result.

13. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-11.

14. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-11.

15. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the method of any one of claims 1-11.