A method, electronic device and storage medium for detecting agent-driven proactive interactive social robots.

Abstract

This invention belongs to the field of cyberspace security and social platform governance technology. Specifically, it provides an agent-driven active interactive social robot detection method, electronic device, and storage medium. The method includes acquiring the raw data of the account to be detected and forming a structured data packet; constructing a preliminary screening and evaluation model and outputting the detection task status based on the structured data packet; generating an initial detection status; obtaining interaction records; extracting evidence from the interaction records to obtain the current round of evidence vectors; subsequently, sequentially updating the initial risk score and uncertainty based on the evidence vectors, and outputting the detection result when the stopping condition is met. This invention introduces an active interactive evidence-gathering mechanism of the detection agent, automatically selecting and initiating probe interactions to obtain incremental evidence on the basis of passive preliminary screening, thereby improving the ability to identify highly realistic social robots and outputting an auditable report containing key evidence and interaction links.

Description

Technical Field

[0001] This invention belongs to the field of cyberspace security and social platform governance technology, and relates to a method for detecting intelligent agent-driven proactive interactive social robots, electronic devices, and storage media. Background Technology

[0002] Social media platforms contain a large number of accounts that are automatically or semi-automatically controlled by programs (often referred to as "social bots / automated accounts / fake accounts"). These accounts cause information pollution, degrade user experience, and increase platform governance costs by mass-posting content, artificially inflating interactions, manipulating public opinion, or implementing marketing conversions. In recent years, with the enhancement of large-scale model generation capabilities, bot accounts can generate more natural text, engage in multi-round interactions, and simulate "personalized expression," making detection methods that rely solely on static content or single behavioral characteristics more prone to failure. Platform risk control faces the following typical difficulties:

[0003] (1) Diverse and rapidly iterating camouflage strategies: The robot can anthropomorphize at the text level and perform rhythm randomization or collaborative arrangement at the behavior level, resulting in continuous drift of the detection feature distribution.

[0004] (2) Limited Data Sources: Privacy compliance and interface limitations make it difficult for some platforms to reliably obtain strong features such as relationship graphs and device information, forcing detection to be completed with weak evidence such as "text + behavior". Existing solutions also point out that some methods that rely on social relationship graphs will significantly increase data acquisition and computational overhead, making them more difficult to implement.

[0005] (3) Enhanced requirements for interpretability and auditability on the engineering side: Platform handling often requires traceable reasons and evidence summaries, while black box models have difficulty outputting verifiable evidence of "why it was judged as a robot".

[0006] Based on the above difficulties, those skilled in the art have developed the following solutions:

[0007] 1) Detection scheme based on rules, blacklists, and thresholds;

[0008] These solutions typically design features and rules based on human experience, such as: thresholds for the number of posts / comments per unit time, the proportion of duplicate content, abnormal follow-unfollowing behavior, common spam links, and keyword hits; and then combine them with blacklists (IP / device / account profiles) to block or demote accounts.

[0009] Advantages: Simple to implement, good real-time performance, and interpretable;

[0010] Disadvantages: It is sensitive to changes in strategy and can be easily bypassed by "rhythm disturbance + text rewriting"; in addition, the rule maintenance cost is high and the generalization ability is weak.

[0011] 2) Feature engineering solutions for traditional machine learning;

[0012] Based on the rules, text statistical features (word frequency, n-gram, perplexity, etc.), behavioral statistical features (interval distribution, active time period, action transition probability, etc.), and account metadata features are further input into models such as LR / SVM / GBDT / XGBoost for supervised classification.

[0013] Advantages: More robust than pure rules, and can be explained to some extent by feature importance;

[0014] Limitations: It is highly dependent on the quality of feature engineering; it is still prone to generalization degradation when faced with highly realistic generated text and collaborative group behavior.

[0015] 3) A deep semantic detection scheme based solely on text content;

[0016] This approach inputs account text into a deep model (such as CNN / RNN / Transformer / BERT) to learn a semantic representation, which is used to identify templated content, repetitions, semantic anomalies, or generation traces.

[0017] Advantages: Strong textual expression capabilities; able to recognize complex semantic patterns.

[0018] Shortcomings: It is becoming less sensitive to "text anthropomorphic" robots; at the same time, it ignores behavioral rhythm and interaction structure, resulting in insufficient identification of accounts with "normal content but abnormal behavior".

[0019] 4) A dual-modal fusion detection scheme combining text and behavior;

[0020] This approach models text semantics while simultaneously modeling behavioral temporal sequences, then performs a fusion output. Typical ideas include: using BERT semantic embedding on the text side and superimposing reconstruction error enhancement, while using a sliding window and bidirectional LSTM to capture dynamic patterns on the behavioral side, followed by cross-modal dynamic fusion, and dynamically updating the threshold to improve long-term adaptability; for example, some solutions emphasize that when the difference between text anomaly and behavioral anomaly is small, the fusion weight approaches 0.5, and dynamically adjusts and updates the threshold based on historical data distribution;

[0021] Advantages: More robust than single-text, and more sensitive to rhythmic bots;

[0022] Shortcomings: Many implementations are still static fusion methods of "feature splicing / fixed weighting", which assumes consistent modal importance and lacks adaptability; performance is prone to degradation when distributions migrate.

[0023] 5) Introduce a graph learning detection scheme based on social relationship graphs;

[0024] This approach introduces an account relationship graph (following, forwarding, commenting networks, etc.) in addition to text and behavior, uses graph neural networks such as GCN / GAT to learn structural features, and then fuses them with text / behavior for discrimination;

[0025] Advantages: More effective against "group collaborative robots" and "data-boosting networks";

[0026] Disadvantages: It relies heavily on relationship graph data, and the cost of acquiring and computing it is high in scenarios where relationship data is unavailable, privacy is restricted, or real-time scenarios, thus limiting its practical application.

[0027] 6) Explicit modeling and triage decision-making scheme for "semantic-behavioral conflict";

[0028] To address the challenge of "normal text but abnormal behavior / normal behavior but abnormal text" in bimodal scenarios, one approach proposes explicitly constructing conflict indices: characterizing the degree of conflict by combining inconsistencies in the representation layer and differences in the predicted distribution, and using conflict indices to drive adaptive path selection: low conflict follows a lightweight fusion for rapid determination, while high conflict enters a cross-modal interaction / conflict resolution path for enhanced determination.

[0029] Advantages: More robust to conflicting samples, while maintaining average cost;

[0030] Shortcomings: It still belongs to the paradigm of "passively analyzing existing data", which makes it difficult to actively supplement evidence when there is insufficient evidence; at the same time, once the diversion logic is adapted to the adversarial approach, there may still be missed detections.

[0031] 7) Intelligent agent orchestration of multi-tool forensics + evidence graph reasoning scheme;

[0032] This approach introduces a detection agent to dynamically orchestrate multiple detection tools, outputting structured evidence items (including scores, confidence levels, and source inferences), then constructing an evidence graph and performing weighted reasoning and fusion to finally output auditable conclusions and evidence chains. For example, existing solutions describe: the agent generates a detection plan and schedules tools, with the tools outputting structured evidence items (including evidence scores, confidence levels, source inferences, and evidence time windows);

[0033] Subsequently, a weighted undirected evidence graph is constructed using evidence items as nodes. The homology, time window overlap, and vector similarity are calculated to establish edges and infer evidence weights, forming a key evidence chain and a minimum evidence subset for auditing.

[0034] Advantages: Evidence is traceable and output is auditable, which is conducive to risk control closed loop;

[0035] Shortcomings: The core still relies on "existing data and tool outputs". When account spoofing results in insufficient or sparse passive evidence, it is difficult to significantly improve the detection rate.

[0036] 8) Dialogue skills enhancement programs based on social simulation / multi-role reflection (related but with different purposes);

[0037] Adjacent to "detection" is a type of solution called "enhancing the dialogue capabilities of social robots"; for example, some solutions simulate scenarios in social situations to conduct multi-role dialogues and reflect on and generate reflective dialogue data, then integrate the reflective dialogue data into the dialogue model through retrieval enhancement generation, and continue to optimize the dialogue model based on user feedback.

[0038] Advantages: Enhanced dialogue generation capabilities and scene adaptability;

[0039] Shortcomings: The goal is "capability enhancement" rather than "identification and detection", and it cannot directly solve the problems of evidence collection and judgment in platform governance.

[0040] In summary, existing technologies still have significant limitations in the governance of highly realistic social robots.

[0041] (1) When passive evidence collection leads to insufficient evidence, it is difficult to improve the reliability of the conclusion;

[0042] (2) Rapid adversarial iterations make it difficult for static strategies / fixed paths to remain stable in the long term;

[0043] (3) Data constraints (especially the unavailability of relationship graphs) make it difficult to implement some high-performance methods;

[0044] (4) Audits and interpretations are either missing or require complex evidence structures to barely support them (such as evidence diagram reasoning). Summary of the Invention

[0045] This invention aims to provide an agent-driven active interactive social robot detection method to address the problems in social platform governance scenarios where highly realistic social robot accounts exhibit textual and behavioral characteristics similar to real users through large-scale model content generation, templated operation, and collaborative control. This leads to problems such as high false negative rates, insufficient evidence collection, untraceable judgments, and slow adaptation to new types of robots in existing detection methods that rely mainly on static features and passive data analysis.

[0046] This invention provides a method for detecting agent-driven, proactive, interactive social robots, comprising the following steps:

[0047] S1. Obtain the raw data of the account to be tested, and standardize the raw data to form a structured data packet. ;

[0048] S2. Constructing an initial screening and assessment model ;

[0049] Structured data packets Input initial screening and evaluation model The initial risk score of the account to be tested is obtained. ;

[0050] Based on the initial risk score Calculate the initial uncertainty ;

[0051] Initial uncertainty Input budget mapping function To determine the maximum number of rounds of subsequent active interactive probing. ;

[0052] Output detection task status ;

[0053] S3, Based on output detection task status Generate initial detection state ;

[0054] S4, Based on the initial detection state Get the interaction record ;

[0055] S5, Interaction Recording Evidence extraction was performed to obtain the evidence vector for this round. Subsequently, based on the evidence vector Initial risk assessment With uncertainty Perform sequential updates and output the detection results when the stopping condition is met. .

[0056] Furthermore, the method for obtaining the raw data of the account to be tested is as follows:

[0057] The account identifier to be tested is set as follows The detection reference time is and the window length is ;

[0058] Based on the window length Determine the detection time window ;

[0059] During the detection time window Internal acquisition and identification of the account to be detected The relevant raw data.

[0060] Furthermore, the raw data of the account to be tested includes the account's data within the testing time window. Text records generated within the time window and accounts to be detected The behavior logs within the system and the interaction availability information based on the account to be tested form a set of interaction constraints. ;

[0061] The text records include at least one of the following: the post body, comment content, reply content, forwarded comments, and personal profile;

[0062] The behavior log records include at least the behavior type and the timestamp of the behavior occurrence. The behavior type includes at least one of the following: posting, commenting, replying, forwarding, liking, following, and unfollowing.

[0063] The set of interaction constraints At a minimum, it should include the available channel types and the maximum number of triggers per channel per unit time. When the platform field is unavailable, the interaction constraint set should be... Take the platform's default restrictions or the business-side configuration.

[0064] Furthermore, structured data packets are formed. The method is as follows:

[0065] Each text record is anonymized, and its corresponding timestamp and source context information are extracted. Then, the text is sorted by timestamp to form a text sequence. ;

[0066] Interaction object information is extracted from each behavior log, and then sorted by timestamp to form a behavior sequence. If the platform field is unavailable, it should be set to empty.

[0067] Based on the interaction availability information of the account to be tested, a set of interaction constraints is formed. ;

[0068] Based on the detection account identifier Detection reference time Detection time window Text sequence Behavioral sequence and the set of interaction constraints To form structured data packets .

[0069] Furthermore, the initial screening assessment model A dual-modal feature fusion neural network structure is adopted, and its specific construction process is as follows:

[0070] Text feature extraction branch: This involves extracting structured data packets. Text sequence in Input the pre-trained language model encoder to extract the semantic feature vector of the account to be detected. ;

[0071] Behavior sequence encoding branch: Encoding behavior sequence The behavioral type and time interval features are embedded and fed into a long short-term memory network or a gated recurrent unit layer to capture the temporal patterns of behavior and output a behavioral feature vector. ;

[0072] Feature fusion and discrimination: combining semantic feature vectors With behavioral feature vector The data is then concatenated, input into a multilayer perceptron for nonlinear feature interaction, and finally processed... The activation function maps the output, and the values ​​are obtained from the output. Initial risk score of the interval .

[0073] Furthermore, the specific process of S3 is as follows:

[0074] S3.1, From the detection task status Reading structured data packets And generated by the image generation function For structured data packets Perform statistical modeling to obtain account profiles ;

[0075] S3.2, from the memory extraction function right Perform fact extraction to obtain contextual memory. ;

[0076] S3.3, Constructing a probe library It includes four types of probes: memory consistency probes, context transfer probes, structured response probes, and style continuity probes.

[0077] Based on account profile With the set of interaction constraints Generate candidate probe set ;

[0078] S3.4, Check the task status Account profiling Contextual memory With candidate probe set Combined to form the initial detection state .

[0079] Furthermore, the specific process of S4 is as follows:

[0080] S4.1, Initial detection state Input decision function To select from the candidate probe set Select probe action for this round ;

[0081] S4.2, Constructing a probe Based on the probe actions in this round Contextual memory Generate probe request and request probe Send to the interaction channel of the account to be tested;

[0082] S4.3 Record probe requests in the interaction channel The sending time is recorded as And after receiving the response from the account to be tested, record the response content as follows: The response reception time is recorded as The interaction delay is thus obtained and denoted as . ;

[0083] S4.4, Perform the probe action in this round. Probe request Response content With interaction latency Combined to form an interactive record .

[0084] Furthermore, the specific process of S5 is as follows:

[0085] S5.1, Evidence Extraction Function Interaction Records With structured data packets Joint analysis was conducted to obtain the evidence vector for this round. ;

[0086] S5.2, The evidence vector for this round Mapped to evidence scores ;

[0087] S5.3, Based on initial risk assessment Evidence Score Perform risk updates to obtain updated risk scores. ;

[0088] S5.4, Based on the updated risk score Calculate the updated uncertainty ;

[0089] S5.5 Setting a high-risk threshold With low risk threshold and based on With uncertainty Determine the stopping condition. The stopping function is denoted as... Its output is a detection marker. ;

[0090] For detection markers With high risk threshold Low risk threshold The results were obtained by comparing the results separately. .

[0091] Furthermore, detection markers With high risk threshold Low risk threshold The comparisons are conducted separately as follows:

[0092] When satisfied At that time, the detection mark will be If it is set as a social robot, then its detection results For social robots;

[0093] When satisfied At that time, the detection mark will be If set to a normal user, the detection results will be... For normal users;

[0094] When satisfied Furthermore, the number of detection rounds in this test did not reach the maximum number of detection rounds. Then, return to S4 to execute the next round of probing;

[0095] When satisfied Furthermore, the number of detection rounds reached the maximum number of detection rounds. At that time, the detection mark will be If the test result is set to require verification, then... This needs to be reviewed.

[0096] Furthermore, the agent-driven active interactive social robot detection method further includes the following steps:

[0097] S6. Based on the detection results Generate testing reports for business processing and traceable auditing.

[0098] Furthermore, the specific process of S6 is as follows:

[0099] S6.1, Constructing the Report Object The reporting object At least include detection markers Risk classification With uncertainty Furthermore, it includes a summary of key evidence related to the detection and information on the interaction links;

[0100] Let the report generation function be denoted as ;

[0101] S6.2, From Interaction Records Extract probe action markers Probe request and response content and compare it with the detection marker Associated write report object ;

[0102] S6.3, From Interaction Records Extracting interaction latency and interaction latency Write to report object This is used to support the auditing of the interaction rhythm;

[0103] S6.4, In the reporting object Write threshold information and This is to record the judgment boundaries used in this test;

[0104] S6.5 Output test report.

[0105] The present invention also provides an electronic device including a memory, one or more processes, and one or more programs stored in the memory, said one or more programs including instructions for performing the agent-driven active interactive social robot detection method as described above.

[0106] The present invention also provides a storage medium including one or more programs executable by one or more processors of an electronic device, said one or more programs including instructions for performing an agent-driven active interactive social robot detection method as described above.

[0107] Compared with the prior art, the present invention has the following beneficial effects:

[0108] (1) This invention introduces an active interactive evidence collection mechanism for detecting intelligent agents. Based on passive initial screening, it automatically selects and initiates probe interactions to obtain incremental evidence, thereby improving the ability to identify highly realistic social robots and outputting an audit report containing key evidence and interaction links.

[0109] (2) Upgrade from passive analysis to proactive interactive evidence collection;

[0110] Instead of relying solely on existing text / behavioral data from accounts, the detection agent proactively initiates probe interactions to acquire incremental evidence under compliance and budget constraints, thereby improving the ability to identify highly realistic robots.

[0111] (3) Agent-driven sequential decision-making mechanism;

[0112] The detection process is not a one-time scoring, but rather a decision-making process based on the current detection status (risk, uncertainty, budget, profile, memory, etc.) among the candidate probe set, forming a closed loop of "detection - evidence collection - update - re-detection" to obtain sufficient evidence strength with fewer interaction rounds.

[0113] (4) Evidence extraction and sequential update based on interactive responses;

[0114] Evidence is extracted and quantitatively scored from the interaction records of each round of probes, including "probe action - probe request - response content - interaction delay". The risk score and uncertainty are sequentially updated, and the conclusion is output in combination with the stopping condition to improve the stability of the judgment and the ability to resist confrontation.

[0115] (5) Auditable output;

[0116] The final output not only provides the judgment result and confidence level, but also forms a summary of key evidence linked to the probe link to support business processing and traceable auditing.

[0117] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0118] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0119] Figure 1 This is a flowchart illustrating an agent-driven active interactive social robot detection method according to an embodiment of the present invention. Detailed Implementation

[0120] To make the above-mentioned objects, features, and advantages of the present invention clearer and easier to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the accompanying drawings of the present invention are all in a simplified form and use non-precise proportions, and are only used to facilitate and clearly illustrate the implementation of the present invention.

[0121] Example 1:

[0122] See Figure 1As shown, this invention provides an agent-driven active interactive social robot detection method. This invention is applicable to internet business scenarios involving account interaction and content publishing, such as social media, content communities, live streaming / e-commerce comment sections, online forums, instant messaging groups, and customer service / Q&A platforms. It can be used for robot account identification, mass account farming and bot control, fake interaction / brushing detection, risk control auditing, and compliance handling. Specifically, the agent-driven active interactive social robot detection method includes the following steps:

[0123] S1. Obtain the raw data of the account to be tested, and standardize the raw data to form a structured data packet. .

[0124] This embodiment uses account detection on a social media platform as an example for explanation. It receives the identifier of the account to be detected. and detection reference time .

[0125] Set window length And based on this, determine the detection time window. ;

[0126] Detection time window The expression is as follows:

[0127] ;

[0128] in, Represented as from the detection reference time To window length (excluding window length) ) to the detection reference time (Including the detection baseline time) (Time interval)

[0129] During the detection time window Internal acquisition and identification of the account to be detected The relevant raw data is processed, and standardized operations such as deduplication, timestamp unification, data masking and replacement, and missing field handling are performed on the raw data to form a structured data package. This raw data is intended for use in subsequent steps. Preferably, the raw data includes at least the following three categories:

[0130] (1) Text data acquisition and standardization. Obtain the data of the account to be tested within the testing time window. The generated text records include at least one or more of the following: post body, comment content, reply content, forwarding remarks, and personal profile. Each text record is anonymized, and its corresponding timestamp and source context information are extracted. These are then sorted by timestamp to form a text sequence. ;

[0131] Among them, text sequence The expression is as follows:

[0132] ;

[0133] in, For the first The de-identified text content of each text record; For the first The timestamp of each text record; For the first Text scene identifier for each text record; For the detection time window The number of text records within.

[0134] (2) Behavioral data acquisition and standardization. This involves acquiring the data of the account to be tested within the testing time window. The system records behavioral logs, which include at least the behavior type and the timestamp of the behavior occurrence. The behavior type includes at least one or more of the following: posting, commenting, replying, forwarding, liking, following, and unfollowing. Interaction object information is extracted from each behavioral log (if the platform field is unavailable, it is set to empty), and then sorted by timestamp to form a behavioral sequence. ;

[0135] Behavioral sequence The expression is as follows:

[0136] ;

[0137] in, For the first The behavior type of each behavior log; For the first The timestamp of the behavior occurrence in the behavior log; For the first The identifier of the interaction object in the behavior log; For the detection time window The number of behavior log entries within.

[0138] (3) Interaction Constraint Acquisition. To support subsequent proactive interactive detection, under the premise of compliance, the interaction availability information of the account to be tested is acquired, and an interaction constraint set is formed. Preferably, the set of interaction constraints At a minimum, this should include the available channel types (public comments / private messages / Q&A, etc.) and the maximum number of triggers per channel per unit of time; if the platform field is unavailable, Take the platform's default restrictions or the business-side configuration. Among them, the set of interaction constraints... Used to describe available interactive channels and frequency control limitations.

[0139] After completing the above processing, a structured data packet is output. (Used as input to S2);

[0140] Structured data packets The expression is as follows:

[0141] .

[0142] S2, based on structured data packets A passive initial screening assessment is performed on the accounts to be tested to obtain an initial risk score and initial uncertainty, and based on this, the maximum number of rounds of subsequent active interactive detection is determined.

[0143] Preferably, the specific process of S2 is as follows:

[0144] S2.1, Transform the structured data packet Input initial screening and evaluation model The initial risk score of the account to be tested is obtained. ;

[0145] S2.2, Based on initial risk score Calculate the initial uncertainty This is to characterize the degree of certainty of the preliminary screening results;

[0146] S2.3, Initial uncertainty Input budget mapping function To determine the maximum number of rounds of subsequent active interactive probing. ;

[0147] S2.4 After completing the initial screening assessment and budget determination, output the testing task status. and will As input to S3.

[0148] Preferred initial risk score The expression is as follows:

[0149] ;

[0150] in, The range of values ​​is , The larger the value, the higher the risk that the account being tested is a social bot.

[0151] Preferred initial uncertainty The expression is as follows:

[0152] ;

[0153] in, The range of values ​​is ; Represented as the natural logarithm function, Represented as natural constant Logarithm with base 0.

[0154] Preferred, maximum number of rounds The expression is as follows:

[0155] ;

[0156] Among them, the budget mapping function The output is a positive integer.

[0157] Preferably, detect task status The expression is as follows:

[0158] .

[0159] Further preferred, the initial screening evaluation model A dual-modal feature fusion neural network structure is adopted, and its specific construction process is as follows:

[0160] Text feature extraction branch: This involves extracting structured data packets. Text sequence in Input the pre-trained language model encoder to extract the semantic feature vector of the account to be detected. ;

[0161] Behavior sequence encoding branch: Encoding behavior sequence The behavioral type and time interval features are embedded and fed into a long short-term memory network or a gated recurrent unit layer to capture the temporal patterns of behavior and output a behavioral feature vector. ;

[0162] Feature fusion and discrimination: combining semantic feature vectors With behavioral feature vector The data is then concatenated, input into a multilayer perceptron for nonlinear feature interaction, and finally processed... The activation function maps the output, and the values ​​are obtained from the output. Initial risk score of the interval .

[0163] S3, Based on output detection task status Establish the detection task status and initialize the probe action set.

[0164] Obtain the detection task status output by S2 and based on Construct a detection state for active interactive probing, and generate a set of candidate probes as input to S4.

[0165] Preferably, the specific process of S3 is as follows:

[0166] S3.1, From the detection task status Reading structured data packets And generated by the image generation function right Perform statistical modeling to obtain account profiles ;

[0167] S3.2, from the memory extraction function right Perform fact extraction to obtain contextual memory. This is used for subsequent probe content construction and consistency verification;

[0168] S3.3, Constructing a probe library It includes four probe types: memory consistency probes, context transfer probes, structured response probes, and style continuity probes. Based on account profiles... With the set of interaction constraints Generate candidate probe set ;

[0169] S3.4, Check the task status Account profiling Contextual memory With candidate probe set Combined to form the initial detection state and will As input to S4.

[0170] Preferred account profiles The expression is as follows:

[0171] .

[0172] Preferred, contextual memory The expression is as follows:

[0173] ;

[0174] Among them, contextual memory By It consists of entities, events, and timeline fragments extracted from it.

[0175] Preferred candidate probe set The expression is as follows:

[0176] ;

[0177] in, For the first One candidate probe action; This represents the number of candidate probe actions.

[0178] Further preferred, for any candidate probe action Represent it as a triple, and satisfy:

[0179] ;

[0180] in, This is a probe type identifier. The value belongs to the probe library. One of the four probe types included; This is the probe content, used to initiate probes to the account to be detected in the interaction channel; The response format constraint is used to limit the response structure of the account to be tested in order to quantify the score.

[0181] Preferred initial detection state The expression is as follows:

[0182] .

[0183] S4. The detection agent generates a detection plan and performs interactive evidence collection.

[0184] Obtain the initial detection state output by S3. The detection agent selects a probe action from the candidate probe set, generates a probe request, performs interactive evidence collection, and obtains an interaction record, which serves as the input to S5.

[0185] Preferably, the specific process of S4 is as follows:

[0186] S4.1, Initial detection state Input decision function To select from the candidate probe set Select probe action for this round ;

[0187] S4.2, Constructing a probe Based on the probe actions in this round Contextual memory Generate probe request and request probe Send to the interaction channel of the account to be tested;

[0188] S4.3 Record probe requests in the interaction channel The sending time is recorded as And after receiving the response from the account to be tested, record the response content as follows: The response reception time is recorded as The interaction delay is thus obtained and denoted as . ;

[0189] S4.4, Perform the probe action in this round. Probe request Response content With interaction latency Combined to form an interactive record And record the interaction. As input to S5.

[0190] Preferably, this round of probe action The expression is as follows:

[0191] ;

[0192] in, Let it be the decision function; Belongs to the candidate probe set .

[0193] Preferably, probe request The expression is as follows:

[0194] ;

[0195] in, Used to initiate probes to the account to be detected in the interaction channel.

[0196] Preferably, the interaction delay is denoted as The expression is as follows:

[0197] .

[0198] Preferred, interaction record The expression is as follows:

[0199] .

[0200] S5, Probe Evidence Extraction, Sequential Update and Stop Determination.

[0201] Get the interaction logs output by S4 and the interaction record Evidence is extracted to obtain the evidence vector for this round; then, the risk score and uncertainty are sequentially updated based on the evidence vector, and the detection conclusion is output when the stopping condition is met.

[0202] Preferably, the specific process of S5 is as follows:

[0203] S5.1, Evidence Extraction Function Interaction Records With structured data packets Joint analysis was conducted to obtain the evidence vector for this round. ;

[0204] S5.2, The evidence vector for this round Mapped to evidence scores ;

[0205] S5.3, Based on initial risk assessment Evidence Score Perform risk updates to obtain updated risk scores. ;

[0206] S5.4, Based on the updated risk score Calculate the updated uncertainty ;

[0207] S5.5 Setting a high-risk threshold With low risk threshold and based on With uncertainty Determine the stopping condition. The stopping function is denoted as... Its output is a detection marker. ;

[0208] For detection markers With high risk threshold Low risk threshold The results were obtained by comparing the results separately. ;

[0209] S5.5 Output detection results and the test results As input to S6.

[0210] Further preferred, detection marker With high risk threshold Low risk threshold The comparisons are conducted separately as follows:

[0211] When satisfied At that time, the detection mark will be If it is set as a social robot, then its detection results For social robots;

[0212] When satisfied At that time, the detection mark will be If set to a normal user, the detection results will be... For normal users;

[0213] If neither of the above conditions is met and the current number of detection rounds has not reached the maximum number of detection rounds. If necessary, return to step S4 to execute the next round of detection;

[0214] If neither of the above conditions is met and the current number of detection rounds reaches the maximum number of detection rounds. At that time, the detection mark will be If the test result is set to require verification, then... This needs to be reviewed.

[0215] Preferred, current round of evidence vector The expression is as follows:

[0216] .

[0217] Preferred evidence score The expression is as follows:

[0218] ;

[0219] in, For evidence scoring functions; The range of values ​​is , The larger the value, the more the evidence in this round supports the claim that the account being tested is a social bot.

[0220] Preferred, updated risk score The expression is as follows:

[0221] ;

[0222] in, for function; The evidence update coefficient.

[0223] Preferred, updated uncertainty The expression is as follows:

[0224] .

[0225] Preferably, detection marker The expression is as follows:

[0226] .

[0227] Preferably, the test results The expression is as follows:

[0228] .

[0229] S6. Output the test results and a summary of auditable evidence.

[0230] Obtain the detection results output by S5 It also generates testing reports for business processing and traceability auditing.

[0231] Preferably, the specific process of S6 is as follows:

[0232] S6.1, Constructing the Report Object The reporting object At least include detection markers Risk classification With uncertainty Furthermore, it includes a summary of key evidence related to the detection and information on the interaction links;

[0233] Let the report generation function be denoted as ;

[0234] S6.2, From Interaction Records Extract probe action markers Probe request and response content and compare it with the detection marker Associated write report object ;

[0235] S6.3, From Interaction Records Extracting interaction latency and interaction latency Write to report object This is used to support the auditing of the interaction rhythm;

[0236] S6.4, In the reporting object Write threshold information and This is to record the judgment boundaries used in this test;

[0237] S6.5 Output test report.

[0238] Preferred reporting subjects The expression is as follows:

[0239] .

[0240] Example 2:

[0241] As a further embodiment of the present invention, the present invention also provides an electronic device, comprising:

[0242] One or more processors;

[0243] Storage device for storing one or more programs;

[0244] When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0245] In practical use, users can interact with servers, which are also electronic devices, via a network to receive or send messages. Terminal devices are generally various electronic devices equipped with a display and used through a human-computer interface, including but not limited to smartphones, tablets, laptops, and desktop computers. Various specific application software can be installed on these terminal devices as needed, including but not limited to web browsers, instant messaging software, social media platforms, and shopping apps.

[0246] Furthermore, the server is a network server that provides various services, such as a backend server that provides corresponding computing services for the raw data of the account to be detected transmitted from the terminal device, so as to realize the processing of the agent-driven active interactive social robot detection method, calculate the social robot detection result, and finally return it to the terminal device.

[0247] Example 3:

[0248] As a further embodiment of the present invention, the present invention also provides a storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the agent-driven active interactive social robot detection method as described above.

[0249] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting an agent-driven, proactive, interactive social robot, characterized in that, Includes the following steps: S1. Obtain the raw data of the account to be tested, and standardize the raw data to form a structured data packet. ; S2. Constructing an initial screening and assessment model ; Structured data packets Input initial screening and evaluation model The initial risk score of the account to be tested is obtained. ; Based on the initial risk score Calculate the initial uncertainty ; Initial uncertainty Input budget mapping function To determine the maximum number of rounds of subsequent active interactive probing. ; Output detection task status ; S3, Based on output detection task status Generate initial detection state ; S4, Based on the initial detection state Get the interaction record ; S5, Interaction Recording Evidence extraction was performed to obtain the evidence vector for this round. Subsequently, based on the evidence vector Initial risk assessment With uncertainty Perform sequential updates and output the detection results when the stopping condition is met. ; The raw data of the account to be tested includes the account during the testing time window. Text records generated within the time window and accounts to be detected The behavior logs within the system and the interaction availability information based on the account to be tested form a set of interaction constraints. ; The text records include at least one of the following: the post body, comment content, reply content, forwarded comments, and personal profile; The behavior log records include at least the behavior type and the timestamp of the behavior occurrence. The behavior type includes at least one of the following: posting, commenting, replying, forwarding, liking, following, and unfollowing. The set of interaction constraints At a minimum, it should include the available channel types and the maximum number of triggers per channel per unit time. When the platform field is unavailable, the interaction constraint set should be... Retrieve platform default restrictions or business-side configurations; The specific process of S3 is as follows: S3.1, From the detection task status Reading structured data packets And generated by the image generation function For structured data packets Perform statistical modeling to obtain account profiles ; S3.2, from the memory extraction function right Perform fact extraction to obtain contextual memory. ; S3.3, Constructing a probe library It includes four types of probes: memory consistency probes, context transfer probes, structured response probes, and style continuity probes. Based on account profile With the set of interaction constraints Generate candidate probe set ; S3.4, Check the task status Account profiling Contextual memory With candidate probe set Combined to form the initial detection state .

2. The agent-driven active interactive social robot detection method according to claim 1, characterized in that, The method for obtaining the raw data of the account to be tested is as follows: Let the account to be detected be identified as The detection reference time is and the window length is ; Based on the window length Determine the detection time window ; During the detection time window Internal acquisition and identification of the account to be detected The relevant raw data.

3. The agent-driven active interactive social robot detection method according to claim 2, characterized in that, Forming structured data packets The method is as follows: Each text record is anonymized, and its corresponding timestamp and source context information are extracted. Then, the text is sorted by timestamp to form a text sequence. ; Information about the interaction objects is extracted from each behavior log record, and then sorted by timestamp to form a behavior sequence. If the platform field is unavailable, it should be set to empty. Based on the interaction availability information of the account to be tested, a set of interaction constraints is formed. ; Based on the account identifier to be detected Detection reference time Detection time window Text sequence Behavioral sequence and the set of interaction constraints To form structured data packets .

4. The agent-driven active interactive social robot detection method according to any one of claims 1-3, characterized in that, The initial screening assessment model A dual-modal feature fusion neural network structure is adopted, and its specific construction process is as follows: Text feature extraction branch: This involves extracting structured data packets. Text sequence in Input the pre-trained language model encoder to extract the semantic feature vector of the account to be detected. ; Behavior sequence encoding branch: Encoding behavior sequence The behavioral type and time interval features are embedded and fed into a long short-term memory network or a gated recurrent unit layer to capture the temporal patterns of behavior and output a behavioral feature vector. ; Feature fusion and discrimination: combining semantic feature vectors With behavioral feature vector The data is then concatenated, input into a multilayer perceptron for nonlinear feature interaction, and finally processed... The activation function maps the output, and the values ​​are obtained from the output. Initial risk score of the interval .

5. The agent-driven active interactive social robot detection method according to claim 4, characterized in that, The specific process of S4 is as follows: S4.1, Initial detection state Input decision function To select from the candidate probe set Select probe action for this round ; S4.2, Constructing a probe Based on the probe actions in this round Contextual memory Generate probe request and request probe Send to the interaction channel of the account to be tested; S4.3 Record probe requests in the interaction channel The sending time is recorded as And after receiving the response from the account to be tested, record the response content as follows: The response reception time is recorded as The interaction delay is thus obtained and denoted as . ; S4.4, Perform the probe action in this round. Probe request Response content With interaction latency Combined to form an interactive record .

6. The agent-driven active interactive social robot detection method according to claim 5, characterized in that, The specific process of S5 is as follows: S5.1, Evidence Extraction Function Interaction Records With structured data packets Joint analysis was conducted to obtain the evidence vector for this round. ; S5.2, The evidence vector for this round Mapped to evidence scores ; S5.3, Based on initial risk assessment Evidence Score Perform risk updates to obtain updated risk scores. ; S5.4, Based on the updated risk score Calculate the updated uncertainty ; S5.5 Setting a high-risk threshold With low risk threshold and based on With uncertainty Determine the stopping condition; the stopping function is denoted as... Its output is a detection marker. ; For detection markers With high risk threshold Low risk threshold The results were obtained by comparing the results separately. .

7. The agent-driven active interactive social robot detection method according to claim 6, characterized in that, Detection Marker With high risk threshold Low risk threshold The comparisons are conducted separately as follows: When satisfied At that time, the detection mark will be If it is set as a social robot, then its detection results For social robots; When satisfied At that time, the detection mark will be If set to a normal user, the detection results will be... For normal users; When satisfied Furthermore, the number of detection rounds in this test did not reach the maximum number of detection rounds. Then, return to S4 to execute the next round of probing; When satisfied Furthermore, the number of detection rounds reached the maximum number of detection rounds. At that time, the detection mark will be If the test result is set to require verification, then... This needs to be reviewed.

8. The agent-driven active interactive social robot detection method according to any one of claims 5-7, characterized in that, It also includes the following steps: S6. Based on the detection results This generates testing reports for business processing and traceable auditing.

9. The agent-driven active interactive social robot detection method according to claim 8, characterized in that, The specific process of S6 is as follows: S6.1, Constructing the Report Object The reporting object At least include detection markers Risk classification With uncertainty Furthermore, it includes a summary of key evidence related to the detection and information on the interaction links; Let the report generation function be denoted as ; S6.2, From Interaction Records Extract probe action markers Probe request and response content and compare it with the detection marker Associated write report object ; S6.3, From Interaction Records Extracting interaction latency and interaction latency Write to report object This is used to support the auditing of the interaction rhythm; S6.4, In the reporting object Write threshold information and This is to record the judgment boundaries used in this test; S6.5 Output test report.

10. An electronic device, characterized in that, It includes a memory, one or more processes, and one or more programs stored in the memory, said one or more programs including instructions for performing the agent-driven active interactive social robot detection method as described in any one of claims 1-9.

11. A storage medium, characterized in that, It includes one or more programs that are executed by one or more processors of an electronic device, the one or more programs including instructions for performing the agent-driven active interactive social robot detection method as described in any one of claims 1-9.

Images