Social robot detection method based on uncertainty driving and agent dynamic arrangement, electronic device and storage medium

Abstract

This invention provides a social robot detection method, electronic device, and storage medium based on uncertainty-driven and agent dynamic orchestration. The method includes: obtaining a standardized data packet to be detected; calculating behavioral entropy and detection uncertainty based on the data packet to be detected, and generating an agent activation signal; initializing the detection agent; outputting an accumulated evidence set based on the detection agent; converting the evidence set into a weighted undirected graph in graph structure form; outputting a binary detection conclusion based on the weighted undirected graph; calculating a key evidence set; and obtaining a social robot detection conclusion based on the binary detection conclusion, the final anomaly risk score, and the key evidence set. This invention introduces behavioral entropy and detection uncertainty as triggers, using only extremely low-cost statistical rules to quickly intercept or allow low-uncertainty samples, and only activating the agent for deep detection on ambiguous high-uncertainty samples, achieving adaptive cost reduction and efficiency improvement by using resources effectively.

Description

Technical Field

[0001] This invention belongs to the field of cyberspace security and social platform governance technology, and relates to a social robot detection method, electronic device and storage medium based on uncertainty-driven and intelligent agent dynamic orchestration. Background Technology

[0002] The continued growth in user and content volume on social media platforms has enabled automated accounts (social bots) to participate in information dissemination, interaction, public opinion guidance, and marketing activities at low cost. Social bots exhibit a clear generational evolution: from early isolated and simple bots to group bots capable of simulating real social networks, and then to new types of bots with "highly realistic simulations, human-machine integration, and adversarial detection strategies," showing an overall trend of becoming increasingly difficult to identify and more adversarial to detection strategies.

[0003] Against this backdrop, detection strategies have evolved from traditional machine learning to deep learning, graph neural networks, and even more complex reinforcement learning / large model recognition. Meanwhile, the heterogeneous distribution of social bots across different user groups / communities, with significant differences in behavior and characteristics across different communities, further increases the difficulty of using a unified detection model.

[0004] Social robot detection typically faces the following challenges in engineering implementation:

[0005] 1) Fixed detection process, difficult to adapt to sample differences: Many solutions use fixed feature extraction and fusion processes. Even if the output probability is used, it is difficult to adjust the strategy under different account types, different data completeness (missing modality / missing field) and different cost constraints (real-time / offline). This results in either wasted computing power or insufficient evidence for high-risk samples.

[0006] 2) Lack of auditable evidence chain: Many deep models output "risk scores / categories" but lack verifiable "evidence chains", making it difficult to meet the requirements of "traceability, explainability, and reproducibility" in scenarios such as platform risk control, appeal review, and compliance audit.

[0007] 3) Difficulty in unified management and fusion of multi-source features: The results of tools such as text, behavior, device, interaction consistency, and collaborative signs have different granularities, scales, and confidence levels. The lack of a unified evidence expression and fusion mechanism makes the system difficult to expand and maintain.

[0008] Therefore, a new detection paradigm is needed: without relying on a "single model structure" for innovation, let the agent be the main thread, and achieve "high accuracy, low cost, and auditable" social robot detection by dynamically selecting analysis tools and constructing evidence graphs. Summary of the Invention

[0009] This invention aims to provide a social robot detection method based on uncertainty-driven and agent dynamic orchestration, in order to solve the problems of wasted computing resources, failure to detect complex disguised samples, and uninterpretable detection results in existing social robot detection technologies.

[0010] This invention provides a social robot detection method based on uncertainty-driven and agent dynamic orchestration, comprising the following steps:

[0011] Step 1: Obtain the target account Within the time range Get the text content sequence Interaction sequence and publicly available portrait metadata The data is then subjected to structured parsing to obtain standardized data packets to be detected. ;

[0012] Step 2: Based on the data packet to be detected Calculate behavioral entropy and detection uncertainty Generate agent activation signal ;

[0013] Step 3: Based on agent activation signals Initialize the detection agent;

[0014] Based on the current uncertainty state, and under computational cost constraints, the detection agent uses a pre-built set of logic detection tools. The system dynamically selects and executes the optimal tool to output structured evidence items. ;

[0015] Based on structured evidence items The probability distribution of the current detection result is updated using a Bayesian update formula or weighted fusion logic, and the new uncertainty is recalculated. ;

[0016] Based on the new uncertainty Determine if the iteration termination condition is met; if so, output the accumulated evidence set. ;

[0017] Step 4: Gather the evidence Transform into a weighted undirected graph in graph structure form ;

[0018] Compute a weighted undirected graph Any pair of nodes in Determine the correlation strength and establish edge connections, and calculate the weighted undirected graph. The final influence weight of each node in the process Simultaneously, the weighted undirected graph is obtained. The final influence weight of each node in the process The highest weight in the process One evidence node;

[0019] Based on final influence weight Calculate the final abnormal risk score of the account to be detected. ;

[0020] The final abnormal risk classification Compared with the preset judgment threshold Compare and output binary detection conclusions. ;

[0021] Step 5: Based on the highest weight A set of evidence nodes and their influence weights All evidence nodes are weighted according to their final influence. Arrange the evidence in descending order to obtain the key evidence set. ;

[0022] Based on binary detection conclusions Final abnormal risk classification and key evidence set The social robot detection results were obtained.

[0023] Furthermore, the specific process for obtaining observable interaction data and publicly available profile metadata of the target account is as follows:

[0024] Step 1.1: Receive the unique identifier of the account to be tested. Compared with the detection baseline time ;

[0025] Set the backtracking time window length to This determines the time range covered by this test. ;

[0026] Step 1.2, targeting the account Within the time range Get the text content sequence Interaction sequence and publicly available portrait metadata The data packet to be detected is then subjected to structured parsing. ;

[0027] Step 1.3: Process the collected text content sequence Interaction sequence and public image metadata All data packets undergo standardization processing, resulting in standardized data packets for testing. .

[0028] Furthermore, the text content sequence Including publication timestamp Cleaned text content and content type identifier ;

[0029] The interaction behavior sequence Including the time of the behavior Public interaction type and the public identifier of the interaction object ;

[0030] The aforementioned public portrait metadata Including registration time Statistical eigenvectors and public client tags .

[0031] Furthermore, the specific process of step two is as follows:

[0032] Step 2.1: Construct behavioral time series and calculate time entropy;

[0033] From the data packet to be detected Text content sequence in With interaction behavior sequence Extract the timestamps from all records, merge them, and sort them in ascending order to construct a comprehensive time series. ;

[0034] Based on comprehensive time series Calculate the time interval sequence of adjacent actions ;

[0035] Calculating behavioral temporal entropy using Shannon entropy ;

[0036] Step 2.2: Extract the initial screening feature vector;

[0037] Construct a lightweight initial screening feature vector ;

[0038] Step 2.3: Initially screen the feature vectors Input a pre-defined lightweight initial screening classifier, and output the initial prediction probability that the account is a social bot. ;

[0039] Step 2.4: Predict probability based on initial screening The detection uncertainty is calculated using the binary entropy formula. ;

[0040] Step 2.5: Calculate the detection uncertainty. With the preset trigger threshold The comparison is performed, and two types of splitting logic are executed: direct output and agent-triggered output.

[0041] Furthermore, the calculated detection uncertainty With the preset trigger threshold The specific process of comparison is as follows:

[0042] like This indicates that the initial screening features are sufficient to make a credible judgment; at this point, the probability predicted directly from the initial screening can be used. The detection process ends when the detection result is output without triggering subsequent steps. If identified as a robot, then it is a real person.

[0043] like This indicates that it is difficult to make a qualitative judgment based solely on statistical characteristics; at this point, the activation signal of the generating agent is generated. and the current data packet to be detected With detection uncertainty Pass it to the next step.

[0044] Furthermore, the specific process of step three is as follows:

[0045] Step 3.1: Pre-set a set of logical analysis tools for anomaly characteristics in different dimensions. Each tool Based on the data packet to be detected get;

[0046] Step 3.2: Based on the state vector of the detection agent Calculate in each iteration The detection intelligent agent is based on a set of logic analysis tools. The tool not yet used in China for calculating expected utility scores ;

[0047] The detection agent employs a greedy strategy, selecting the expected utility score. The highest tool as the current action Execute;

[0048] Step 3.3: Execute the selected tool Output structured evidence items ;

[0049] Step 3.4: Based on structured evidence items The probability distribution of the current detection result is updated using a Bayesian update formula or weighted fusion logic, and the new uncertainty is recalculated. ;

[0050] Check if the iteration termination condition is met; if the termination condition is met, output the accumulated evidence set. Proceed to step four; if not satisfied, return to step 3.2.

[0051] Furthermore, each tool Includes the following logical units:

[0052] Semantic consistency analysis tools ;

[0053] Input: Text sequence ;

[0054] Logic: Utilize a lightweight natural language reasoning model to detect whether there are logical contradictions in the historical tweets posted by the account under test;

[0055] Output: Semantic conflict score ;

[0056] Circadian rhythm anomaly detection tool ;

[0057] Input: From the sequence of interaction behaviors A sequence of behavioral timestamps;

[0058] Logic: Map timestamps to a 24-hour clock disk and calculate the overlap and dispersion of active time distribution with the normal human sleep cycle;

[0059] Output: Rhythm Anomaly Score ;

[0060] Interactive homogeneity detection tool ;

[0061] Input: From the sequence of interaction behaviors A list of interactive objects;

[0062] Logic: Analyze whether the target audience of the account's public interactions exhibits a high degree of homogeneity;

[0063] Output: Network homogeneity score .

[0064] Furthermore, the iteration termination condition must satisfy any of the following:

[0065] Confidence level meets the standard: ;

[0066] Budget exhausted: , This is represented as the preset upper limit. This represents the estimated cost of the next candidate tool;

[0067] Tools exhausted: All available tools have been executed.

[0068] Furthermore, the specific process of step four is as follows:

[0069] Construct a weighted undirected graph ;

[0070] Based on weighted undirected graph Collection of evidence Each piece of evidence in the graph is mapped to a graph node. ;

[0071] Compute a weighted undirected graph Any pair of nodes in Conflict level ;

[0072] Conflict level With preset threshold Compare; if This indicates that the node has... If the evidence is consistent with the conclusion, then establish positive support edges and assign their weights. Set to a positive value; if This indicates that the node has... The evidence contradicts the conclusions; therefore, a negative inhibition edge is established, and its weight is... Set to a negative value or use a special marker;

[0073] The weighted undirected graph is updated using a confidence-weighted propagation algorithm. The final influence weight of each node in the process At the same time, the one with the highest weight in the calculation process is obtained. One evidence node;

[0074] Based on the final influence weight of each node Calculate the final abnormal risk score of the account to be detected. ;

[0075] The final abnormal risk classification Compared with the preset judgment threshold Compare and output binary detection conclusions. .

[0076] Furthermore, based on the binary detection conclusions Final abnormal risk classification and key evidence set The specific process for obtaining the social robot detection conclusion is as follows:

[0077] Using a pre-built evidence interpretation template library, the binary detection conclusions are presented. Final abnormal risk classification and key evidence set Transform the evidence items in the document into natural language descriptions;

[0078] The content described in natural language is encapsulated into a standard JSON format audit log structure. And generate a unique hash fingerprint;

[0079] The generated audit log The results are written to an immutable log storage medium and the detection conclusions are returned to the business side via an API interface; if If so, it is determined to be a robot.

[0080] As a further aspect of the present invention, the social robot detection method based on uncertainty-driven and agent dynamic orchestration further includes the following steps:

[0081] Step 6: Feedback loop and adaptive parameter update.

[0082] As a further aspect of the present invention, the social robot detection method based on uncertainty-driven and agent dynamic orchestration, wherein step six is ​​specifically as follows:

[0083] Receive for the account to be tested Final manual review of labels or business processing feedback In this context, 0 represents a confirmed real person, and 1 represents a confirmed anomaly / bot.

[0084] Encapsulate the complete context of this detection task into an archived entry. Stored in the historical sample database;

[0085] For each tool that actually participated in this test Evaluate its output With vacuum label Consistency;

[0086] The tool's base confidence level in the knowledge base is updated using an exponential moving average algorithm. ;

[0087] Periodically calculate the false alarm rate and false negative rate within the sliding window; if the false alarm rate exceeds the preset warning threshold, automatically adjust the judgment threshold in step S4. Fine-tuning was performed to obtain the updated judgment threshold. After the update is complete, the latest parameter set will be provided. Compared with the updated judgment threshold Write it to the configuration center for use in the next step of the detection process from step one to step five.

[0088] As a further aspect of the present invention, 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 executing the social robot detection method based on uncertainty-driven and agent dynamic orchestration as described above.

[0089] As a further aspect 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 executing the social robot detection method based on uncertainty-driven and agent dynamic orchestration as described above.

[0090] Compared with the prior art, the present invention has the following beneficial effects:

[0091] (1) This invention introduces behavioral entropy and detection uncertainty as triggers. For simple samples (low uncertainty), it uses only low-cost statistical rules to quickly intercept or allow them. Only for complex samples (high uncertainty) that are "ambiguous" is the agent activated for deep detection, achieving adaptive cost reduction and efficiency improvement by "using the best resources where they are most needed". This invention constructs a detection agent, which, based on the utility function of "maximizing information gain / cost", dynamically selects what to check next like a detective (e.g., only checking semantic logic if suspicious behavior is found). This serial reasoning decision-making mechanism significantly improves the ability to capture new high-simulation robots. In this invention, by constructing a dynamic evidence graph, graph theory algorithms are used to automatically handle contradictions between evidence (conflict resolution), and finally outputs a natural language audit report containing "reasoning trajectory", realizing complete transparency and traceability of the detection process. The overall detection architecture of this invention protects a two-stage detection method of "initial screening-precision testing", that is, first calculate the detection uncertainty based on behavioral time entropy, and only when the uncertainty exceeds the threshold is the agent initialized and enter the complete process of dynamic orchestration.

[0092] (2) Decision-making logic of the agent (core algorithm):

[0093] The strategy for protecting intelligent agents is to select detection tools based on the "expected utility function." That is: (The expected reduction in uncertainty divided by the computational cost), this "cost-effectiveness" decision-making logic defined by a mathematical formula is the soul of the algorithm in this invention.

[0094] (3) Construction and reasoning methods of evidence diagrams:

[0095] This paper protects the mechanism for handling multimodal conflicts using graph structures. In particular, it defines the construction rules for "supporting edges" (feature consistency) and "conflicting edges" (feature contradiction), as well as a weight update algorithm based on confidence propagation.

[0096] (4) Privacy-compliant data processing paradigm:

[0097] Protect an analytical method "based on publicly available behavioral logical characteristics". This is a technical approach that does not rely on user privacy data (such as device fingerprints or IP addresses), but instead identifies anomalies by analyzing the logical characteristics of publicly available data (such as "semantic consistency" and "circadian rhythm overlap").

[0098] 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

[0099] 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:

[0100] Figure 1 This is a flowchart illustrating a social robot detection method based on uncertainty-driven and agent dynamic orchestration in an embodiment of the present invention. Detailed Implementation

[0101] 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.

[0102] Example 1:

[0103] See Figure 1 As shown, the present invention provides a social robot detection method based on uncertainty-driven and agent dynamic orchestration, comprising the following steps:

[0104] S1: Obtain the target account Within the time range Get the text content sequence Interaction sequence and publicly available portrait metadata The data is then subjected to structured parsing to obtain standardized data packets to be detected. .

[0105] This step aims to establish the basic data representation of the account to be tested. All data collection processes are based on the standard data interfaces provided by the social platform or publicly visible front-end information, and do not involve reading the user's private device information.

[0106] Preferably, the specific process for obtaining observable interaction data and publicly available profile metadata of the target account is as follows:

[0107] S1.1 Determine the detection window;

[0108] Receive the unique identifier of the account to be tested Compared with the detection baseline time (Usually the current moment). Set the backtracking time window length to... This determines the time range covered by this test. ;

[0109] Time range The expression is as follows:

[0110] ;

[0111] S1.2 Obtain and parse the raw data;

[0112] Target account Within the time range Get the text content sequence Interaction sequence and publicly available portrait metadata The data packet to be detected is then subjected to structured parsing. ;

[0113] Data packets to be detected The expression is as follows:

[0114] ;

[0115] in, It is represented as a sequence of text content, which is a record of content published by the account to be tested, obtained from the historical content database of the network platform or through public timelines; Represented as an interaction behavior sequence, it consists of observable operational behaviors of the account to be detected extracted from the public interaction logs of the network platform; This refers to public profile metadata, which is non-privacy attribute information extracted from the public personal homepage of the account to be tested. Public profile metadata is used to construct the basic profile of the account to be tested.

[0116] Preferred, text content sequence The expression is:

[0117] :

[0118] in, The total number of text content records of the account to be tested. Represented as the first A record of text content.

[0119] Preferably, in this embodiment, the text content sequence Including publication timestamp (Converted to UTC standard time), cleaned text content (Remove HTML tags and invisible characters, but keep emojis) and content type identifiers. (Original / Forwarded / Reply).

[0120] Preferred, interaction behavior sequence The expression is as follows:

[0121] ;

[0122] in, The total number of interaction records of the account to be tested. Represented as the first A record of interactive behavior.

[0123] Preferably, in this embodiment, the interaction behavior sequence Including the time of the behavior Public interaction type (Including: likes, follows, unfollows, favorites) and public identifiers of the interacting objects. (For example, the ID of the post that was liked or the ID of the user that was followed). Specifically, in this embodiment, only explicit interactive behaviors visible to other users on the social network are collected for the content in the interaction behavior sequence, excluding users' private click streams or browsing history.

[0124] Preferably, in this embodiment, the publicly disclosed image metadata... Including registration time Statistical eigenvectors and public client tags ;

[0125] Registration time The timestamp created for the account to be tested is used to calculate the account's "network age".

[0126] Statistical eigenvectors This includes publicly displayed metrics such as number of followers, number of fans, and total number of tweets posted.

[0127] Public client tags This refers to the source identifier string that social media platforms publicly display below tweets (e.g., "Twitter for iPhone", "Twitter Web App", "Android Client", etc.).

[0128] S1.3 Data standardization and preprocessing;

[0129] The collected text content sequence Interaction sequence and public image metadata All data packets undergo standardization processing, resulting in standardized data packets for testing. This serves as the input for calculating the uncertainty in step S2.

[0130] Preferably, in this embodiment, the standardization process includes:

[0131] ① Handling missing values: If If empty, mark it as "Unknown"; if statistical features are missing, pad with zeros.

[0132] ② Text encoding: for The text is segmented and stop word removed, and pre-trained language models (such as BERT or lightweight Word2Vec) are used to convert the text into semantic vector representations.

[0133] ③ Time alignment: Verify all timestamps Is it in the window? Within the data, dirty data from abnormal time points is removed.

[0134] S2: Based on the data packet to be detected Calculate behavioral entropy and detection uncertainty Generate agent activation signal .

[0135] This step aims to process the data packet to be detected output in step S1. A low-cost, rapid assessment is conducted by quantifying the temporal regularity (entropy) of the behavior of the accounts to be detected and the prediction confidence of the initial screening model to calculate the detection uncertainty, which serves as the basis for deciding whether to activate subsequent high-cost intelligent agent detection.

[0136] Preferably, the specific process of S2 is as follows:

[0137] S2.1 From the data packet to be detected Text content sequence in With interaction behavior sequence Extract the timestamps from all records, merge them, and sort them in ascending order to construct a comprehensive time series. ;

[0138] Composite Time Series The expression is as follows:

[0139] ;

[0140] in, For time window Total number of active users within the app.

[0141] S2.2, Based on comprehensive time series Calculate the time interval sequence of adjacent actions ;

[0142] Time interval sequence of adjacent actions The expression is as follows:

[0143] ;

[0144] .

[0145] S2.3 To quantify the regularity of behavior, Shannon entropy is used to calculate the temporal entropy of behavior. The specific calculation method is as follows: The time interval sequence of adjacent actions... Discretize the data into buckets and statistically analyze the probability distribution of the occurrence of each bucket in each time interval. ;

[0146] Time entropy The expression is as follows:

[0147] ;

[0148] in, It is represented as a logarithmic operation with base 2.

[0149] Technical principle: Automated scripts typically exhibit extremely low time entropy (mechanically timed sending) or extremely high time entropy (pseudo-random sending), while the time interval distribution of real users usually conforms to a Pareto distribution or a specific physiological rhythm, and their entropy value is in the middle range.

[0150] S2.4 Extract the initial screening feature vector;

[0151] To perform rapid risk assessment, a lightweight initial screening feature vector is constructed. This vector consists only of statistical features with extremely low computational cost, and does not contain complex semantic encoding or graph features;

[0152] Initial screening of feature vectors The expression is as follows:

[0153] ;

[0154] in: Expressed as an activity rate ratio, (Total number of active users divided by window length); Represented as normalized metadata statistical features (taken from the statistical feature vector in step S1) ); This represents the one-hot encoding of the client tag, corresponding to the public client tag in S1. .

[0155] S2.5 Calculate the initial screening probability and uncertainty; the specific process is as follows:

[0156] S2.5.1 Construct a lightweight gradient boosting decision weight model based on ensemble learning. This model has the characteristics of low computational complexity, is suitable for processing structured statistical features, and can meet the millisecond-level initial screening requirements.

[0157] Specifically, the lightweight gradient boosting decision weight model based on ensemble learning includes an input adaptation layer, an ensemble inference layer, and a probability output layer.

[0158] The input adaptation layer includes a feature sparsification processing module, which is used to optimize the sparse matrix for "one-hot encoding of client tags" and directly transmit "behavioral temporal entropy" and "normalized statistical features" to combine scalars of different dimensions into a feature matrix that the model can read.

[0159] The integrated inference layer is the core computational layer of this model, consisting of K regression decision trees; each tree is based on the split points of the features (e.g.: The samples are divided into specific leaf nodes, and each leaf node is assigned a weight score. This layer does not perform complex matrix convolution operations, but only extremely fast Boolean checks and addition operations, ensuring a millisecond-level response time;

[0160] The probability output layer includes a Sigmoid activation function module, which is used to sum the weights of the leaf nodes of all trees in the ensemble inference layer. As input ,pass Calculate the final probability.

[0161] S2.5.2, Initially screen the feature vectors The input is fed into a pre-built lightweight initial screening classifier (such as a logistic regression model or a lightweight random forest) in the input adaptation layer, and the probability output layer outputs the initial screening prediction probability that the account to be detected is a social bot. ;

[0162] Initial screening prediction probability The expression is as follows:

[0163] ;

[0164] .

[0165] S2.5.3 Construct an uncertainty assessment and diversion controller;

[0166] Predicting the probability of initial screening In the input uncertainty assessment and diversion controller, the detection uncertainty, which measures the model's hesitation towards the current sample, is calculated using the binary entropy formula (binary entropy function). .

[0167] Detection uncertainty The expression is as follows:

[0168] ;

[0169] Numerical Explanation: Initial Prediction Probability The detection uncertainty is close to 0 (confident to be a real person) or 1 (confident to be a robot). Approaching 0; the initial screening prediction probability When the value is close to 0.5 (ambiguous), the detection uncertainty is... It reached its maximum value of 1.

[0170] S2.5.4 Uncertainty Assessment and Comparison Logic Gate in the Diversion Controller Calculate the Detection Uncertainty With the preset trigger threshold (Specifically, in this embodiment, the preset trigger threshold) The value is set to 0.6 for comparison, and a high-level signal or a low-level signal is output.

[0171] The routing distributor in the uncertainty assessment and diversion controller determines the data flow direction based on a high-level or low-level signal; specifically, a low-level signal triggers the direct output interface, thus executing the direct output (low uncertainty branch); a high-level signal triggers the agent initialization interface, thus executing the agent trigger (high uncertainty branch).

[0172] Preferably, the uncertainty assessment and diversion controller is deployed after the output of the initial screening model and between the input of the detection agent, and is responsible for receiving the initial screening prediction probability from the initial screening model. And determine the subsequent calculation path.

[0173] More preferably, the uncertainty assessment and routing controller includes an uncertainty calculation unit, a threshold comparison logic gate, and a routing distributor;

[0174] The input to the uncertainty calculation unit is the initial screening prediction probability output by the initial screening model. This is used to load a preset binary entropy formula and determine the initial screening prediction probability. Convert to detection uncertainty ;

[0175] The input to the threshold comparison logic gate is the detection uncertainty. With the preset trigger threshold This is used to perform numerical comparisons and judgments; if If the output logic level is "0", then the output is low risk or low uncertainty; if If the output logic level is "1", then the output is high uncertainty.

[0176] The input to the routing distributor is the result of a threshold comparison logic gate, used to perform data flow control; if the received result of the threshold comparison logic gate is "0", the direct output interface is activated, and the initial prediction probability is... The result is directly mapped to the final probability; if the result of the received threshold comparison logic gate is "1", the agent initialization interface is activated to generate the agent activation signal. and the data packet to be detected With detection uncertainty Both routes are routed to S3.

[0177] Preferably, the calculated detection uncertainty is... With the preset trigger threshold The specific process of comparison is as follows:

[0178] like This indicates that the initial screening features are sufficient to make a reliable judgment. At this point, the probability prediction can be directly based on the initial screening. Output the detection results without triggering subsequent steps; this detection process ends. If identified as a robot, then it is a real person.

[0179] like This indicates that it is difficult to make a definitive judgment based solely on statistical characteristics (e.g., belonging to a robot with sophisticated disguise or a real person exhibiting abnormal behavior). At this point, an agent activation signal is generated. and the current data packet to be detected With detection uncertainty This is then passed to step S3.

[0180] S3: Based on agent activation signals Initialize the detection agent;

[0181] Based on the current uncertainty state, and under computational cost constraints, the detection agent uses a pre-built set of logic detection tools. The system dynamically selects and executes the optimal tool to output structured evidence items. ;

[0182] Based on structured evidence items The probability distribution of the current detection result is updated using a Bayesian update formula or weighted fusion logic, and the new uncertainty is recalculated. ;

[0183] Check if the iteration termination condition is met. If it is, output the accumulated evidence set. .

[0184] Preferably, step S3 specifically includes the following sub-steps:

[0185] S3.1 Define a logic testing tool library;

[0186] The component library includes a pre-built set of logical analysis tools for anomalies of different dimensions. Each tool Relying solely on the data packets to be detected obtained in step S1 It does not involve the acquisition of external privacy data.

[0187] Each tool Specifically, it includes (but is not limited to) the following logical units:

[0188] Semantic consistency analysis tools ;

[0189] Input: Text sequence ;

[0190] Logic: Utilize a lightweight Natural Language Inference (NLI) model to detect whether there are logical contradictions in the historical tweets posted by the account under test (e.g., the same account claims to be in two geographically distant locations within a very short period of time, or the identity descriptions are inconsistent).

[0191] Output: Semantic conflict score ;

[0192] Circadian rhythm anomaly detection tool ;

[0193] Input: A sequence of behavior timestamps (from a sequence of interaction behaviors) );

[0194] Logic: Map timestamps to a 24-hour clock disk and calculate the overlap and dispersion of active time distribution with normal human sleep cycles (such as 02:00-06:00 in the UTC+8 interval); used to detect the existence of machine characteristics such as "24-hour non-stop" or "strict fixed-point operation";

[0195] Output: Rhythm Anomaly Score ;

[0196] Interactive homogeneity detection tool ;

[0197] Input: List of interactive objects (from the sequence of interactive behaviors) );

[0198] Logic: Analyze whether the target audience of public interactions (likes / shares) of an account is highly homogeneous (e.g., the interaction targets are all new accounts with similar registration times, or the interaction targets all belong to a known high-risk list).

[0199] Output: Network homogeneity score .

[0200] Preferably, in this embodiment, the lightweight natural language reasoning model adopts a Siamese network architecture based on knowledge distillation. Its construction process is as follows:

[0201] Choose a lightweight pre-trained model (such as DistilBERT or Albert) as the base;

[0202] The model was fine-tuned using publicly available natural language inference datasets.

[0203] The focus of training is on the model's ability to identify the "Contradiction" label, classifying "Implication" and "Neutral" as non-conflict categories, thereby outputting a single semantic conflict score.

[0204] In a further preferred embodiment, the lightweight natural language reasoning model logically comprises three sequentially connected module layers: a semantic encoding layer, an interactive reasoning layer, and a conflict resolution layer.

[0205] The semantic encoding layer consists of two lightweight Transformer encoders with shared weights; it receives two texts to be compared (e.g., text A "I am in Beijing", text B "I am in New York now") and maps them into high-dimensional semantic vectors. and This layer is responsible for converting unstructured text into machine-computable mathematical vectors;

[0206] The interactive reasoning layer includes a vector concatenation and difference calculation unit, which is used to fuse two semantic vectors to capture the distance and difference features of the two texts in the semantic space.

[0207] The conflict resolution layer includes a fully connected layer and a Softmax activation function; it is used to map the output of the interaction layer to... The probability value; its output is directly used as the "semantic conflict score". The closer the score is to 1, the higher the probability that the two texts are logically contradictory.

[0208] S3.2, Strategies for evaluating and selecting the utility of intelligent agents;

[0209] The detection agent maintains a state vector ,in: This represents the uncertainty at the current moment. This represents the computational cost already incurred.

[0210] In each iteration In the middle, the detection agent targets the logic analysis toolset Tools not yet used in the calculation of expected utility scores ;

[0211] Expected utility score The expression is as follows:

[0212] ;

[0213] in:

[0214] Represented as tools The expected reduction in uncertainty (based on the historical average information gain estimate of the tool).

[0215] Represented as tools The estimated cost (such as time complexity or API call cost).

[0216] The detection agent employs a greedy strategy, selecting the expected utility score. The highest tool as the current action Execute.

[0217] S3.3, Tool Execution and Evidence Generation;

[0218] Execute the selected tool Output structured evidence items ;

[0219] Structured evidence items The expression is as follows:

[0220] ;

[0221] in: Indicated as a tool identifier, This is represented as the abnormal score output by the tool (specifically, in this embodiment, Set to 0.9). This is represented by the credibility weight of the tool itself. (These are preset constants, such as a weight of 0.8 for text analysis tools and 0.6 for time analysis). It is expressed as a natural language description (e.g., "High-frequency active behavior was detected between 3 and 5 a.m.").

[0222] S3.4 Uncertainty Update and Termination Decision;

[0223] Based on structured evidence items The probability distribution of the current detection result is updated using a Bayesian update formula or weighted fusion logic, and the new uncertainty is recalculated. .

[0224] Check if any of the following iteration termination conditions are met:

[0225] Confidence level meets the standard: (Uncertainty is below the stopping threshold, for example, sufficient evidence has been obtained to determine that it is a robot);

[0226] Budget exhausted: (Total computing cost) Reaching the preset limit , (represented as the estimated cost of the next candidate tool).

[0227] Tools exhausted: All available tools have been executed.

[0228] If the termination condition is met, output the accumulated set of evidence. Proceed to step S4; otherwise, return to S3.2 to proceed to the next iteration.

[0229] S4: Constructing dynamic evidence graphs and conflict resolution.

[0230] This step receives the evidence set output from step S3 through a pre-built evidence graph construction module (which instantiates a graph object in memory, traverses the evidence set, instantiates each evidence item as a node object of the graph, and initializes the adjacency matrix). This is transformed into a graph structure, and graph theory methods are used to handle the support and conflict relationships between evidence from different sources, calculating the final weight of each piece of evidence to arrive at a high-confidence detection conclusion. Step S4 specifically includes the following sub-steps:

[0231] S4.1 Construct a weighted undirected graph ;

[0232] Based on weighted undirected graph Collection of evidence Each piece of evidence in the graph is mapped to a graph node. .

[0233] Each node contains an attribute vector. ;

[0234] In step four, Specifically refers to the first term in a weighted undirected graph. One piece of evidence, Less than or equal to the total number of evidence nodes; Specifically refers to the first term in a weighted undirected graph. One piece of evidence, The total number of evidence nodes is less than or equal to the total number of evidence nodes.

[0235] Weighted undirected graph The expression is:

[0236] ;

[0237] in, Represented as a set of nodes; Represented as a set of edges;

[0238] Attribute vector The expression is as follows:

[0239] ;

[0240] in, This represents the probability of an anomaly output by the tool (0 represents complete normality, and 1 represents complete anomaly). This represents the base confidence level of the tool.

[0241] S4.2 Calculate the weighted undirected graph Any pair of nodes in The correlation strength is determined and edge connections are established; where, Edge weights This reflects the degree of consistency among the evidence.

[0242] Compute a weighted undirected graph Any pair of nodes in Conflict level ;

[0243] Conflict level The expression is as follows:

[0244] ;

[0245] Conflict level With preset threshold (Specifically, in this embodiment, the threshold) Set to 0.2) for comparison;

[0246] like This indicates that the node has... If the evidence is consistent (mutually supporting), establish a positive support edge and assign its weight. Set to a positive value (e.g.) , (represented as the conflict degree of the node pair).

[0247] like This indicates that the node has... If the evidence and conclusions contradict each other (there is a conflict), a negative inhibition edge is established, and its weight is... Set to a negative value or add a special marker.

[0248] The preferred logical example of a positive support edge is as follows:

[0249] The simultaneous occurrence of "abnormal circadian rhythm" and "high-frequency mechanical posting" enhances each other's credibility.

[0250] The logical example of a negative suppression edge is as follows:

[0251] There is a conflict between "natural text semantics (like a real person)" and "24 / 7 operation (like a robot)".

[0252] S4.3 Conflict resolution based on graph reasoning;

[0253] Utilize graph reasoning algorithms (such as simplified graph attention mechanisms or weighted voting) for weighted undirected graphs. The importance of any node in the process is reassessed in order to resolve conflicts.

[0254] The final influence weight of each node is updated using a confidence-weighted propagation algorithm. At the same time, the one with the highest weight in the calculation process is obtained. One evidence node;

[0255] Final Influence Weight The expression is as follows:

[0256] ;

[0257] in, Represented as nodes There exists a set of neighboring nodes directly connected by edges.

[0258] Resolution principle: When a conflict occurs, the base confidence level... Evidence that is higher in rank and receives more positive support from neighboring nodes will ultimately have a higher influence weight. Evidence that is isolated and of low confidence will be amplified; conversely, the weight of isolated and low-confidence evidence will be suppressed.

[0259] For example, in the aforementioned conflict of "text as real as a human" but "work and rest as robotic," the increasing sophistication of "LLM (Limited Language Management) technology" has improved the functionality of text tools. The levels are lower, and "biological rhythms" are difficult to physically fake. The weight of the text evidence is too high, so the algorithm will automatically reduce the weight of the text evidence and tend to judge it as a robot.

[0260] S4.4, Generate the final risk score and judgment;

[0261] Based on the final influence weight of each node Calculate the final abnormal risk score of the account to be detected. ;

[0262] Final abnormal risk score The expression is as follows:

[0263] ;

[0264] The final abnormal risk classification Compared with the preset judgment threshold (Specifically, in this embodiment, the determination threshold) The comparison is set to 0.75, and the binary detection conclusion is output. ;

[0265] Binary detection conclusion The expression is as follows:

[0266] ;

[0267] After completing the calculation, the binary detection results will be... Final abnormal risk classification and the highest weight in the calculation process Each evidence node (i.e., the final influence weight) The largest node is passed to step S5 to generate an audit report.

[0268] S5: Output the detection results and an auditable chain of evidence.

[0269] This step aims to transform the mathematical reasoning results generated in step S4 into a human-readable structured audit report. By extracting high-weight evidence nodes, a complete chain of evidence is constructed, providing interpretable evidence for subsequent manual review, user complaint handling, or compliance audits. Step S5 specifically includes the following sub-steps:

[0270] S5.1 Extraction and sorting of key evidence;

[0271] Receive the final evidence graph nodes and their influence weight set output from step S4. .

[0272] Influence weight set The expression is as follows:

[0273] ;

[0274] in, This represents the total number of evidence nodes. Corresponding to the The final influence weight of each evidence node.

[0275] In order to focus on core risk characteristics, key evidence screening is performed:

[0276] All evidence nodes are weighted according to their final influence. Sort in descending order.

[0277] Before selection Nodes (e.g.) (as key evidence set) ;

[0278] Key Evidence Set The expression is as follows:

[0279] ;

[0280] in, This represents the primary evidence with the highest weight. This is represented by the main evidence, with the weights decreasing sequentially.

[0281] S5.2 Natural Language Report Generation;

[0282] By using a pre-built library of evidence interpretation templates, structured key evidence can be transformed into natural language descriptions.

[0283] The generation logic is as follows:

[0284] Conclusion sentence generation: Based on the binary detection conclusion and final abnormal risk classification Generate header conclusions.

[0285] Template: "Account" Determined as The risk confidence level is .

[0286] Evidence chain assembly: Traversing the key evidence set The evidence item in the document, read its description field ( ) and source tools ( (The paragraphs are then arranged according to their importance to form reasoning paragraphs.)

[0287] Logical connectives: For "supporting edge" in S4, use "and" and "further confirmation"; for "winning evidence after conflict resolution", use "although...but the agent detection found...".

[0288] Preferably, the pre-built evidence interpretation template library specifically includes the following three types of mapping templates:

[0289] Global conclusion template: Used to describe the final binary detection results and risk level.

[0290] Example: "Account {Target_ID} was identified by the system as {Label} (social bot / normal user), with a comprehensive risk score of {Final_Score} and a confidence level of {Confidence_Level}."

[0291] Tool-level feature description template: For each specific tool in the detection toolkit (such as the tool defined in S3.1), a corresponding parameterized text slot is preset.

[0292] For circadian rhythm tools: If an anomaly is detected, the template "This account exhibits abnormally high frequency of activity during the inactive period {Sleep_Window} of {Time_Zone}, with an activity level of {Activity_Rate}, which is inconsistent with human physiological characteristics" should be invoked.

[0293] For semantic consistency tools: If an anomaly occurs, call the template "The account's published historical content contains logical contradictions at {Conflict_Count}, for example, claiming {Location_A} in {Time_A} but claiming {Location_B} in {Time_B}".

[0294] For homogenized tools: If an anomaly is detected, call the template "The interaction objects of this account are highly homogenized, and {Percent}% of the interaction targets belong to known risk groups".

[0295] Logical association and conflict resolution template: Based on the edge attributes (supporting edge / suppressing edge) of the graph structure in step S4, generate connectors to construct fluent reasoning paragraphs.

[0296] Supporting Relationship (Positive Edge): When two high-weighted pieces of evidence are consistent, use the template "Furthermore, the findings of {Evidence_A} further corroborate the conclusion of {Evidence_B}, forming a complete chain of evidence."

[0297] Conflict resolution (negative edge): When high-weighted evidence overwhelms low-weighted evidence, the template "Although {Evidence_Weak} shows some normal characteristics, based on the higher confidence of {Evidence_Strong}, the system determines that the former is a fake feature and ultimately accepts the latter" is used.

[0298] S5.3, Construct tamper-proof audit logs;

[0299] To meet auditing requirements, the above information is encapsulated into a standard JSON format audit log structure. And generate a unique hash fingerprint.

[0300] Audit Log Structure The definition is as follows:

[0301] {

[0302] "trace_id": "UUID-Timestamp", / / Unique tracking ID

[0303] "decision": {

[0304] "label": "Social_Bot", / / Detection conclusion Y

[0305] "score": 0.85 / / Final risk score R_final

[0306] },

[0307] "evidence_chain": [ / / Key evidence chain E_key

[0308] {

[0309] "rank": 1,

[0310] "tool": "T2_Circadian", / / Source tool

[0311] "reason": "The account's activity time is highly concentrated between 02:00-04:00 AM, which violates normal physiological rhythms."

[0312] "weight": 0.45 / / Weight alpha

[0313] },

[0314] {

[0315] "rank": 2,

[0316] "tool": "T1_Semantic",

[0317] "reason": "The published content contains a large number of repetitive, templated text snippets",

[0318] "weight": 0.30

[0319] }

[0320] ],

[0321] "raw_ref": "Hash(D)" / / A hash reference to the original data packet (does not store the original data, only the fingerprint for verification).

[0322] }

[0323] S5.4 Result Output and Storage;

[0324] The generated audit log The results are written to an immutable log storage medium and the detection conclusions are returned to the business side via an API interface.

[0325] like If the tweet is identified as a bot, the corresponding action will be taken (such as folding the tweet, requiring two-factor authentication, or banning the user).

[0326] Meanwhile, the log can be pushed to the manual review platform as a "work order". Reviewers can quickly determine whether the model has been mistakenly deleted based on the evidence_chain field, thereby greatly improving the efficiency of manual review.

[0327] S6, Feedback closed loop and adaptive parameter update;

[0328] This step aims to establish a long-term evolutionary mechanism for the detection system. By receiving real labels from manual review or business feedback, the core parameters upon which the agent's decision-making depends are adaptively calibrated, enabling the detection method to continuously learn and resist drift as samples accumulate. Step S6 specifically includes the following sub-steps:

[0329] S6.1 Sample archiving and label alignment;

[0330] The system receives data for the account to be tested. Final manual review of labels or business processing feedback (Where 0 represents confirmed real person, and 1 represents confirmed abnormal / robot).

[0331] Encapsulate the complete context of this detection task into an archived entry. Stored in the historical sample database;

[0332] Archived entries The expression is as follows:

[0333] ;

[0334] in, This is a subset of tools that the agents in S3 have actually invoked.

[0335] S6.2 Dynamic calibration of tool confidence;

[0336] For each tool that actually participated in this test Evaluate its output With real labels Consistency.

[0337] The tool's base confidence level in the knowledge base is updated using an exponential moving average algorithm. ;

[0338] Baseline confidence level The expression is as follows:

[0339] ;

[0340] in: The smoothing coefficient (e.g., 0.01) is used to control the update rate and avoid drastic parameter fluctuations.

[0341] The indicator function is set to 1 when the deviation between the tool's judgment result and the true label is less than 0.5 (i.e., the judgment is correct), and 0 otherwise.

[0342] Technical effect: Through this mechanism, the confidence weight of stable tools will gradually increase, while the weight of tools that are proven to have made misjudgments will automatically decrease, thereby realizing the "survival of the fittest" of the tool library by the agent.

[0343] S6.3, Decision threshold drift correction;

[0344] To address the evolving patterns of social bot attacks, the system periodically (e.g., every...) The false positive rate (FPR) and the missed positive rate (TPR) within a sliding window are statistically analyzed for each sample.

[0345] If the false alarm rate exceeds the preset warning threshold, the judgment threshold in step S4 will be automatically adjusted. Fine-tuning was performed to obtain the updated judgment threshold. ;

[0346] Updated judgment threshold The expression is as follows:

[0347] ;

[0348] in, This indicates the adjustment step size. This is represented as the false alarm rate for the current window. This is expressed as the target false alarm rate.

[0349] After the update is complete, the latest parameter set will be provided. Compared with the updated judgment threshold Write it to the configuration center for use in the next S1-S5 detection process, thus completing the closed loop.

[0350] Example 2:

[0351] As a further embodiment of the present invention, the present invention also provides an electronic device, comprising:

[0352] One or more processors;

[0353] Storage device for storing one or more programs;

[0354] When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0355] 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.

[0356] 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 target account transmitted from the terminal device, so as to realize the processing of the social robot detection method based on uncertainty-driven and intelligent agent dynamic orchestration, calculate the social robot detection results, and finally return them to the terminal device.

[0357] Example 3:

[0358] 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 social robot detection method based on uncertainty-driven and agent dynamic orchestration as described above.

[0359] 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 social robot detection method based on uncertainty-driven and agent dynamic orchestration, characterized in that, Includes the following steps: Step 1: Obtain the target account Within the time range Get the text content sequence Interaction sequence and publicly available portrait metadata The data is then subjected to structured parsing to obtain standardized data packets to be detected. ; Step 2: Based on standardized data packets to be detected Calculate behavioral entropy and detection uncertainty Generate agent activation signal ; Step three, activating the agent based on the signal , initializing the detection agent; Based on the current uncertainty state, and under computational cost constraints, the detection agent uses a pre-built set of logic detection tools. The system dynamically selects and executes the optimal tool to output structured evidence items. ; Based on structured evidence items The probability distribution of the current detection result is updated using a Bayesian update formula or weighted fusion logic, and the new uncertainty is recalculated. ; Based on the new uncertainty Determine if the iteration termination condition is met; if so, output the accumulated evidence set. ; Step 4: Gather the evidence Transform into a weighted undirected graph in graph structure form ; Compute a weighted undirected graph Any pair of nodes in Determine the correlation strength and establish edge connections, and calculate the weighted undirected graph. The final influence weight of each node in the process Simultaneously, the weighted undirected graph is obtained. The final influence weight of each node in the process The highest weight in the process One evidence node; Based on the final influence weight , calculate the final abnormal risk score of the to-be-detected account ; The final abnormal risk classification Compared with the preset judgment threshold Compare and output binary detection conclusions. ; Step 5: Based on the highest weight A set of evidence nodes and their influence weights All evidence nodes are weighted according to their final influence. Arrange the evidence in descending order to obtain the key evidence set. ; based on the binary detection conclusion , the final abnormal risk score and the key evidence set , a social robot detection conclusion is obtained; The specific process of step two is as follows: Step 2.1: Construct behavioral time series and calculate time entropy; From standardized data packets to be detected Text content sequence in With interaction behavior sequence Extract the timestamps from all records, merge them, and sort them in ascending order to construct a comprehensive time series. ; Based on integrated time series Computing a sequence of time intervals of adjacent behaviors ; Shannon entropy is used to calculate behavioral time entropy ; Step 2.2: Extract the initial screening feature vector; Constructing a lightweight preliminary screening feature vector ; Step 2.3: Initially screen the feature vectors Input a pre-defined lightweight initial screening classifier, and output the initial prediction probability that the account is a social bot. ; Step 2.4: Predict probability based on initial screening The detection uncertainty is calculated using the binary entropy formula. ; Step 2.5: Calculate the detection uncertainty. With preset trigger threshold The comparison is performed, and two types of splitting logic are executed: direct output and agent-triggered output. The specific process of step three is as follows: The specific process of step three is as follows: Step 3.1: Pre-set a set of logical analysis tools for anomaly characteristics in different dimensions. Each tool Based on standardized data packets to be detected get; Step 3.2: Based on the state vector of the detection agent Calculate in each iteration The detection intelligent agent is based on a set of logic analysis tools. The tool not yet used in China for calculating expected utility scores ; The detection agent employs a greedy strategy, selecting the expected utility score. The highest tool as the current action Execute; Step 3.3, executing the selected tool , output structured evidence items ; Step 3.4: Based on structured evidence items The probability distribution of the current detection result is updated using a Bayesian update formula or weighted fusion logic, and the new uncertainty is recalculated. ; Check if the iteration termination condition is met; if the termination condition is met, output the accumulated evidence set. Proceed to step four; if not satisfied, return to step 3.

2.

2. The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to claim 1, characterized in that, The specific process for obtaining observable interaction data and publicly available profile metadata of the target account is as follows: Step 1.1: Receive the unique identifier of the account to be tested. Compared with the detection baseline time ; Set the backtracking time window length as , thereby determining the time range covered by the current detection ; Step 1.2, targeting the account Within the time range Get the text content sequence Interaction sequence and publicly available portrait metadata The data packet to be detected is then subjected to structured parsing. ; Step 1.3: Process the collected text content sequence Interaction sequence and public image metadata All data packets undergo standardization processing, resulting in standardized data packets for testing. . 3.The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to claim 2, characterized in that, The text content sequence Including publication timestamp Cleaned text content and content type identifier ; The interaction behavior sequence Including the time of the behavior Public interaction type and the public identifier of the interaction object ; The aforementioned public portrait metadata Including registration time Statistical eigenvectors and public client tags .

4. The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to claim 2 or 3, characterized in that, The calculated detection uncertainty With the preset trigger threshold The specific process of comparison is as follows: If , indicates that the initial screening characteristics are sufficient to make a reliable determination; At this time, directly according to the primary screening prediction probability Output the detection result, do not trigger the subsequent steps, and the detection process ends this time; and Determine as a robot, otherwise as a real person; If , it indicates that it is difficult to determine the nature only by statistical characteristics; At this time, the agent activation signal is generated and the current data packet to be detected and the detection uncertainty are passed to the next step.

5. The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to claim 4, characterized in that, Each tool includes the following logical units: Semantic consistency analysis tools ; Input: Text sequence ; Logic: Utilize a lightweight natural language reasoning model to detect whether there are logical contradictions in the historical tweets posted by the account under test; Output: Semantic conflict score ; Circadian rhythm abnormality detection tool ; Input: sequence of behavior timestamps from a sequence of interaction behaviors ; Logic: Map timestamps to a 24-hour clock disk and calculate the overlap and dispersion of active time distribution with the normal human sleep cycle; Output: Rhythm abnormality score ; Interaction homogeneity probing tool ; Input: list of interaction objects from a sequence of interaction behaviors ; Logic: Analyze whether the target audience of the account's public interactions exhibits a high degree of homogeneity; Output: Network homogeneity score .

6. The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to claim 5, characterized in that, The iteration termination condition must satisfy any of the following: Confidence threshold: ; Budget depletion: , is represented as a preset upper limit, is represented as an estimated computational cost of the next candidate tool; Tools exhausted: All available tools have been executed.

7. The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to claim 6, characterized in that, The specific process of step four is as follows: constructing a weighted undirected graph ; Weighted undirected graph based Mapping each evidence item in a set of evidence to a graph node ; Compute a weighted undirected graph Any pair of nodes in Conflict level ; Conflict level With preset threshold Compare; if This indicates that the node is paired with... If the evidence is consistent with the conclusion, then establish positive support edges and assign their weights. Set to a positive value; if This indicates that the node is paired with... The evidence contradicts the conclusions; therefore, a negative inhibition edge is established, and its weight is... Set to a negative value or use a special marker; The weighted undirected graph is updated using a confidence-weighted propagation algorithm. The final influence weight of each node in the process At the same time, the one with the highest weight in the calculation process is obtained. One evidence node; Based on the final influence weight of each node Calculate the final abnormal risk score of the account to be detected. ; The final abnormal risk score is calculated compared with a preset determination threshold and a binary detection conclusion is output . 8.The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to claim 7, wherein, Based on the binary detection conclusion , the final abnormal risk score and the key evidence set , the specific process of obtaining the social robot detection conclusion is as follows: The preset evidence interpretation template library is adopted to convert the binary detection conclusion , final abnormal risk score and evidence item in key evidence set into natural language description; The content described in natural language is encapsulated into a standard JSON format audit log structure. And generate a unique hash fingerprint; The generated audit log The results are written to an immutable log storage medium and the detection conclusions are returned to the business side via an API interface; if If so, it is determined to be a robot.

9. The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to any one of claims 1-8, characterized in that, It also includes the following steps: Step 6: Feedback loop and adaptive parameter update.

10. The social robot detection method based on uncertainty-driven and agent dynamic orchestration according to claim 9, characterized in that, The specific process of step six is ​​as follows: Receiving a final human review label or business handling feedback for the account to be detected wherein 0 represents confirming as a real person and 1 represents confirming as an abnormality / robot;​ encapsulating the complete context of the present detection task as an archival entry into a historical sample library; For each tool that actually participated in this test Evaluate its output With real labels Consistency; The tool's base confidence level in the knowledge base is updated using an exponential moving average algorithm. ; Periodically calculate the false alarm rate and false negative rate within the sliding window; if the false alarm rate exceeds the preset warning threshold, automatically adjust the judgment threshold in step S4. Fine-tuning was performed to obtain the updated judgment threshold. After the update is complete, the latest parameter set will be provided. Compared with the updated judgment threshold Write it to the configuration center for use in the next step of the detection process from step one to step five.

11. An electronic device, comprising: 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 executing the social robot detection method based on uncertainty-driven and agent dynamic orchestration as described in any one of claims 1-10.

12. A storage medium characterized by Includes one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the social robot detection method based on uncertainty-driven and agent dynamic orchestration as described in any one of claims 1-10.

Images