A malicious social account identification method and system based on two-agent cooperation

Abstract

This invention discloses a method and system for identifying malicious social media accounts based on two-agent collaboration, belonging to the field of social media security technology. This invention addresses the problems of insufficient identification capability and high false positive rate in existing malicious social media account identification methods under highly disguised, highly concealed, and information-sparse scenarios. It obtains an enhanced feature set by performing content analysis and feature expansion on target account information; uses a recognition model built based on a large language model to perform malice identification and reasoning analysis on the enhanced feature set, obtaining account classification results and an assessment report; and fine-tunes the recognition model based on the account classification results, the assessment report, and expert-annotated data to obtain an updated recognition model for continuous application. This invention can effectively capture highly concealed malicious disguise features, correct the risk of misjudgment of ambiguous accounts through expert experience alignment, achieve dynamic iteration and logical closed loop of identification capabilities, and significantly improve the governance efficiency of social media risk control.

Description

Technical Field

[0001] This invention discloses a method and system for identifying malicious social media accounts based on two-agent collaboration, belonging to the field of social media security technology. Background Technology

[0002] With the explosive growth of large language model technologies, exemplified by ChatGPT, the security boundaries of social networks are undergoing a profound reconstruction. Generative artificial intelligence not only empowers machines to generate high-quality, logically coherent, and emotionally rich text, but also provides powerful "intelligent weapons" for cybercrime groups. This technological revolution has given rise to a new generation of "major social media accounts empowered by large models." These accounts are extremely complex: they include both evolved "fully automated malicious bots" that utilize large model intelligent agent technology to achieve 24 / 7 human-like automated attacks, and "real-person malicious accounts" or "human-machine hybrid controlled accounts" that use large models to assist in generating text. In the social network environment, whether automated or manually operated, malicious social media accounts share a core commonality: a clear destructive intent. They achieve illegal profit-making, dissemination of false information, and manipulation of public opinion through actions such as inciting conflict and publishing harmful content. In the current environment, the evolution of malicious social media accounts from simple "malicious script tools" to "intelligent entities with malicious intent" poses a severe challenge to traditional rule-based or simple automated defense methods.

[0003] In recent years, existing technical solutions for identifying malicious social media accounts can be categorized into the following three types: The first category is traditional machine learning methods based on rule statistics and feature engineering. These methods aim to identify anomalous social media accounts by constructing intuitive user behavior profiles. The approach primarily relies on expert experience to manually extract metadata features such as posting frequency, registration time, and follower ratio, which are then used to train classifiers like random forests and SVMs. Their effectiveness lies in strong model interpretability, low computational cost, and the ability to effectively filter out early-stage, low-level anomalous accounts. However, these methods are highly dependent on expert experience, feature updates are lagging, making them ill-equipped to handle continuously evolving camouflage strategies; furthermore, they lack a deep understanding of account content information and social behavior characteristics, making it difficult to identify malicious accounts.

[0004] The second category is based on deep learning and graph neural networks. These methods aim to automatically extract high-dimensional semantic features and uncover complex network relationships. Regarding account content information, these methods first utilize pre-trained language models such as BERT to deeply understand user-posted content, achieving semantic representation of user posts. Simultaneously, regarding account behavior information, they construct a social behavior graph of the account, treating accounts as nodes in the graph and modeling interactions such as following and forwarding between accounts as edges on the social behavior graph. Graph neural networks are then used for feature aggregation to capture user social behavior characteristics. In terms of feature technology, graph anomaly detection is used to predict whether an account node is a malicious social account. A typical method is the BotRGCN model, which constructs a heterogeneous social graph containing "user-user" (e.g., follow relationships) and "user-tweet" (e.g., posting relationships). Account node representations are initialized using account metadata, numerical features, and textual features, and then an RGCN iteratively updates the account representations and makes predictions. Although these methods take into account the content and behavioral information of the account, they are limited by the lack of text representation capabilities, making it difficult to effectively utilize the long-term historical information of the account and to identify the obscure and disguised content generated by large models.

[0005] The third category is based on large language models. These methods leverage the semantic understanding and reasoning capabilities of large language models to achieve intelligent identification of malicious accounts. At the content analysis level, targeted Prompt templates are designed to guide the large language model to conduct fine-grained reviews of account-posted content from dimensions such as semantic style, logical coherence, and emotional expression, identifying human-machine differences. At the behavioral feature level, some studies supplement the behavioral features of accounts that directly interact with the main account, thus considering the account's social information. However, limited by the finite window of the large model, it is difficult to capture the features of distant social accounts with collaborative relationships; furthermore, most current malicious account identification methods employ static data processing methods, making it difficult to dynamically capture account information on the internet according to identification needs, resulting in incomplete and insufficiently utilized account information, thus limiting the improvement of malicious social account identification performance.

[0006] The existing technology has the following drawbacks: (1) Current methods employ fixed text content or relational topology encoders, lacking effective contextual reasoning mechanisms, making it difficult to identify malicious accounts in scenarios involving intent spoofing and feature ambiguity. To achieve malicious purposes, current malicious social media accounts are often generated through human manipulation or by leveraging large models, resulting in highly anthropomorphic malicious accounts with adversarial and deceptive characteristics. Existing technologies mainly utilize fixed text or relational topology encoding features, which present difficulties in identifying complex behavioral patterns and adversarial content generation. As the machine traces of automated accounts are significantly weakened, the deceptiveness of human accounts is further amplified, making it difficult for second agents relying on shallow semantic features to accurately identify their potential malicious intent.

[0007] (2) Current methods mainly focus on the recent content and behavior information of target accounts on specific platforms, making it difficult to dynamically obtain long-term historical information, cross-platform account information, and information on distantly related accounts as needed for account identification. Existing methods mostly rely on the current snapshot data of the account (such as the latest 100 tweets and basic account metadata), and cannot actively supplement historical behavior, social topology, or cross-platform evidence. When malicious accounts adopt strategies such as "content dilution" and "persona disguise," they are very likely to be missed due to feature sparsity or misjudged due to local anomalies. In addition, real malicious behavior often appears in the form of groups, with collaborative behavior in social behavior networks. Most existing methods take directly related accounts such as comments and reposts as account features, lacking information perception of potential distant collaborative accounts, which limits the improvement of the ability to identify collaborative malicious social accounts.

[0008] (3) Most existing methods use fixed model parameters, which cannot continuously learn from expert experience, resulting in information sparsity and insufficient feature support in emerging scenarios: once deployed, mainstream solutions remain fixed and cannot learn from newly emerging adversarial examples or expert corrections. Although reinforcement learning can introduce feedback, it requires training an independent reward model, which is complex and unstable. The DPO direct preference optimization mechanism adopted in this invention can directly align with human preferences. By collecting the account identification results data in this invention as preference optimization data, it enhances the reasoning and recognition capabilities of the intelligent agent model and continuously iterates and optimizes over time to achieve continuous account updates, thereby adapting to the identification capabilities of emerging malicious social accounts in constantly changing scenarios. Summary of the Invention

[0009] The purpose of this invention is to address the problems of insufficient identification ability and high false positive rate of existing malicious social media account identification methods in scenarios with high camouflage, strong concealment and sparse information. This invention proposes a malicious social media account identification method and system based on two-agent collaboration. By using a two-stage progressive judgment mechanism and direct preference optimization technology, a dynamic, high-dimensional account feature profile and judgment framework with self-evolution capabilities are constructed, which significantly improves the accuracy, robustness and interpretability of the inference results in complex scenarios.

[0010] To achieve the above objectives, the present invention adopts the following technical solution.

[0011] A method for identifying malicious social media accounts based on two-agent collaboration includes the following steps: The first intelligent agent is used to perform content analysis and feature expansion on the target account information to obtain an enhanced feature set; The second intelligent agent is used to perform malice identification and reasoning analysis on the enhanced feature set to obtain account classification results and an assessment report. The second agent is fine-tuned and trained based on the account classification results, analysis reports, and expert annotation data to obtain an updated second agent, which is then applied to subsequent malicious social media account identification.

[0012] Furthermore, the first intelligent agent is used to perform content analysis and feature expansion on the target account information to obtain an enhanced feature set, including: The first intelligent agent is used to perform a quantitative evaluation of the information completeness of the target account information, and a comprehensive completeness score is obtained. Based on the comprehensive completeness score, the historical behavior trajectory of the target account is traced back using the first intelligent agent to obtain historical information features; Based on the comprehensive completeness score, the first agent traverses the multi-hop association network in the social behavior graph to obtain the features of long-distance associated accounts. Based on the comprehensive completeness score, the first intelligent agent is used to identify the corresponding entities of the target account on other social platforms through cross-domain identity association, thereby obtaining cross-platform account features.

[0013] Furthermore, the first intelligent agent performs a quantitative assessment of the information completeness of the target account information to obtain a comprehensive completeness score, including: Calculate the objective rule statistics score based on the total number of tweets published by the target account and the effective population ratio of metadata fields; A first intelligent agent based on a large language model is used to semantically score whether the existing information is sufficient to characterize the true personality of the account, and a semantic inference score is obtained. The objective rule statistical score and the semantic inference score are weighted and fused to obtain a comprehensive completeness score.

[0014] Furthermore, based on the comprehensive completeness score, the first intelligent agent is used to trace the historical behavior trajectory of the target account to obtain historical information features, including: When the comprehensive completeness score is lower than a preset threshold, the first intelligent agent dynamically determines the historical time range and behavior type to be collected for the target account based on the current observation and thinking results. Based on the determined historical time range and behavior type, a stratified sampling strategy is used to obtain fine-grained historical content of the target account during abnormally active periods through social platform interfaces; The acquired fine-grained historical content is subjected to feature extraction and structuring to obtain the historical information features composed of content evolution features, behavioral rhythm features, and interaction authenticity features.

[0015] Furthermore, based on the comprehensive completeness score, the first agent traverses the multi-hop association network in the social behavior graph to obtain long-distance associated account features, including: Using the target account as the root node, construct an initial adjacency graph based on interactive behaviors; The system uses a dual-constraint pruning strategy of semantics and structure to filter irrelevant nodes and identifies potential collaborative malicious groups through the first agent. The topological features, behavioral consistency features, and identity disguise clues of the coordinated malicious group are extracted to obtain the features of long-distance related accounts.

[0016] Furthermore, based on the comprehensive completeness score, the first intelligent agent identifies the corresponding entities of the target account on other social platforms through cross-domain identity association, obtaining cross-platform account features, including: Extract the target account's metadata and cross-platform query seed set from the published content using the first intelligent agent; Multimodal matching models are used to perform text matching, visual matching, and behavioral verification on the target platform to confirm cross-platform identity association. Obtain the metadata and published content of the associated identity on other platforms to generate cross-platform features.

[0017] Furthermore, a second intelligent agent is used to perform maliciousness identification and reasoning analysis on the enhanced feature set to obtain account classification results and an assessment report, including: The historical tweet sequences in the enhanced feature set are scanned one by one, and a suspicion score is assigned. For accounts with high scores, a second intelligent agent is used to make a comprehensive judgment by combining features from various dimensions in the enhanced feature set, and the account classification results and judgment report are output.

[0018] Furthermore, for the top-scoring accounts, a second intelligent agent is used to comprehensively determine the accounts based on features from various dimensions of the enhanced feature set, outputting account classification results and an analysis report, including: By using a second intelligent agent to analyze and enhance features, objective outliers in the expert rule base were accurately identified. The systemic malice of the anomalies is analyzed by using a second intelligent agent in conjunction with social context, and fault tolerance verification is performed to find reasonable explanations to refute the malicious hypothesis. Weighing the positive and negative evidence, we arrive at a judgment and generate an analysis report that includes the analytical process.

[0019] Furthermore, the second intelligent agent is fine-tuned and trained based on the account classification results, analysis reports, and expert-annotated data, including: Based on the account classification results and the analysis report, the account feature information of the misjudged samples is extracted as input samples, and the correct reasoning process rewritten by the expert and the incorrect reasoning report generated by the second agent are used as winning response and losing response, respectively, to construct a triplet preference dataset. The network weights of the second agent are updated by minimizing the direct preference optimization loss function using the triplet preference dataset.

[0020] A malicious social media account identification system based on two-agent collaboration includes: The feature expansion module is used to perform content analysis and feature expansion on the target account information using the first intelligent agent to obtain an enhanced feature set. The malicious identification module is used to perform maliciousness identification and reasoning analysis on the enhanced feature set using a second intelligent agent, and to obtain account classification results and judgment reports. The model optimization module is used to fine-tune the training of the second agent based on the account classification results, the judgment report, and expert annotation data to obtain an updated second agent, and then apply the updated second agent to subsequent malicious social account identification.

[0021] The present invention has achieved the following beneficial effects.

[0022] 1. This invention decouples the identification task into a two-agent collaborative architecture and utilizes the reasoning capabilities of a large language model to achieve intelligent analysis from surface-level account information to deep-level dimensions. It can not only deeply understand obscure and disguised content and complex behavioral patterns, but also eliminate the illusion of reasoning caused by input sparsity through the completeness assessment in the first stage, significantly improving the interpretability of the identification results and the logical quality of the judgment report.

[0023] 2. This invention adopts a dynamic investigation strategy based on the ReAct paradigm, which can spontaneously call tools to fill in missing information in an open environment according to the recognition needs. By multi-source fusion of long-distance historical behavior trajectory, multi-hop long-distance correlation network and cross-platform profile, a high-dimensional and dynamic digital profile of the account is constructed, which completely breaks through the limitation of traditional recognition methods that rely only on static data from a single platform.

[0024] 3. This invention introduces a fine-grained evaluation mechanism for single content. By performing multi-dimensional suspicion scoring and screening on massive historical tweets, it effectively defends against the "content dilution" strategy of malicious accounts using trivial daily content to cover up violations. While preventing core clues from being diluted by redundant data, it significantly improves the capture accuracy and system computing efficiency of the second intelligent agent in complex business scenarios.

[0025] 4. This invention constructs a continuous evolution mechanism based on direct preference optimization. By transforming the intuition of human experts in judging fuzzy cases into model preference signals, it realizes closed-loop correction of false positives and false negatives. This enables the second agent to continuously iterate and optimize its internal probability distribution as data accumulates, thereby effectively adapting to the dynamic evolution of malicious social account characteristics and possessing rapid learning and long-term robustness in identifying new spoofing strategies. Attached Figure Description

[0026] Figure 1 This is a simplified flowchart of a malicious social media account identification method based on two-agent collaboration in an embodiment. Figure 2 This is a framework diagram of a malicious social media account identification method based on two-agent collaboration in an embodiment. Figure 3 This is a block diagram of a malicious social media account identification system based on two-agent collaboration, as shown in the embodiment. Detailed Implementation

[0027] To make the various technical features, advantages, or effects of the present invention more apparent and understandable, detailed descriptions are provided below through embodiments.

[0028] This invention provides a method for identifying malicious social media accounts based on two-agent collaboration, such as... Figure 1 and Figure 2 As shown, the specific processing steps are explained below.

[0029] Step S1: Use the first intelligent agent to perform content analysis and feature expansion on the target account information to obtain an enhanced feature set.

[0030] Specifically, this step first checks the completeness of the social media account's content based on the given target account information, and automatically obtains the account's remote account information and cross-platform account information. Remote account information refers to multi-hop neighbor accounts on the social behavior graph that are associated with the target account through forwarding and commenting relationships; cross-platform account information refers to the target account's account information on other social media platforms.

[0031] In an optional embodiment of the present invention, this step aims to perform a preliminary analysis of the account content using a first intelligent agent built on a Large Language Model (LLM) based on the given target account information, and to determine whether the completeness of the account content data is sufficient to complete the determination of the account's malice.

[0032] In an optional embodiment of the present invention, the platform where the target account is located can be Weibo, and other platforms can be Xiaohongshu, Douyin, etc.

[0033] In an optional embodiment of the present invention, step S1 may include: Step S11: Use the first intelligent agent to perform a quantitative evaluation of the information completeness of the target account information and obtain a comprehensive completeness score.

[0034] Specifically, this step is used to block the illusion of recognition caused by extremely sparse input features. After receiving the initial feature set of the account to be tested, this step first performs a dual information completeness quantification assessment on the target account information and provides reliable evidence through two assessment scores. In terms of specific calculation methods, LLM achieves this by integrating objective statistical indicators with large model semantic evaluation to calculate a comprehensive completeness score.

[0035] In an optional embodiment of the present invention, step S11 may include: Step S111: Calculate the objective rule statistics score based on the total number of tweets published by the target account and the effective filling ratio of metadata fields.

[0036] Step S112: Use an LLM-based first agent to perform a semantic score on whether the existing information is sufficient to characterize the true personality of the account, and obtain a semantic inference score.

[0037] Step S113: The objective rule statistical score and the semantic inference score are weighted and fused to obtain a comprehensive completeness score.

[0038] Specifically, the formula for calculating the comprehensive completeness score is constructed as follows: in, This represents the objective rule-based statistical score; The total number of tweets posted by this account; This is a normalization function used to evaluate the richness of the content posted by an account; This represents the effective population ratio of the account's metadata fields, reflecting the completeness of the basic information. The metadata fields include nickname, bio, number of followers, region, and age. and The corresponding statistical item weights; The semantic inference score representing LLM is the score given by the model based on the criteria for identifying malicious accounts, which determines whether the existing information is sufficient to characterize the true personality of the user of the account. This is a harmonic weighting coefficient used to balance the constraints of hard objective rules with the results of soft semantic inference; This indicates the overall completeness score.

[0039] In an optional embodiment of the present invention, when the calculated comprehensive completeness score When the threshold is lower than the preset cutoff threshold, this step will trigger the LLM's dynamic reconnaissance loop based on the Observation-Thought-Action (ReAct) paradigm, thereby actively invoking the historical information collection step (S12), the remote associated account discovery step (S13), and the cross-platform account discovery step (S14), thus improving the target account information. Traditional pipeline-style detection often reads data only once, and if the information is insufficient, it will directly report an error or produce a false alarm. However, the ReAct loop used in this invention achieves the adaptability of account information supplementation: (1) Observation: LLM first reads the currently acquired account metadata, content information, topology, etc. as the initial state.

[0040] (2) Thinking: The model asks itself questions based on the current information and quantitative assessment to determine whether the chain of evidence is complete.

[0041] (3) Action: Based on the thinking, LLM selects tools from the tool library and initiates the call. The tools include the historical information collection tool for executing step S12, the remote associated account discovery tool for executing step S13, and the cross-platform account discovery tool for executing step S14.

[0042] (4) Loop and Termination: The new data returned by the tool is updated in the status, and LLM enters the next round of observation. Based on the observation and analysis results and the completeness score, it is determined whether the target account information is complete or the maximum number of steps has been reached. Finally, the enhanced feature set is output.

[0043] In an optional embodiment of the present invention, during the thinking phase, the model will reason as follows: The account profile claims to be a local lifestyle blogger, but currently there is only one retweeted tweet, which is insufficient to verify its account category. More historical comment information and historical posting frequency information need to be checked. Faced with simple samples, LLM determines that deep reconnaissance is unnecessary; however, faced with complex disguised samples, LLM will automatically perform multiple rounds of in-depth information mining, avoiding the resource waste caused by performing full-scale in-depth mining on all accounts while ensuring detection accuracy.

[0044] Step S12: Based on the comprehensive completeness score, the first intelligent agent is used to trace back the historical behavior trajectory of the target account to obtain historical information features.

[0045] Specifically, this step is used by the LLM to proactively trace the historical behavior of the target account when information completeness is insufficient, in order to supplement its dynamic characteristics over time. Malicious accounts often exhibit behavioral patterns such as short-term concentrated outbreaks, highly repetitive content, or deliberate disguise, making accurate identification difficult based solely on current snapshot data. Therefore, this step retrieves complete interaction logs of the target account within a specified time window by calling historical interfaces provided by social media platforms. These social media platform interfaces include Weibo API and Douyin historical posting records, among others.

[0046] In an optional embodiment of the present invention, these features are not only used to enhance the completeness of the current account profile, but also serve as key input for subsequent second-stage maliciousness inference. It is worth noting that historical information collection is not performed on all accounts at once, but is triggered by the information completeness assessment step and initiated on demand by the LLM, thereby ensuring detection depth while maintaining system efficiency and compliance, and avoiding excessive crawling by the model.

[0047] In an optional embodiment of the present invention, step S12 may include: Step S121: When the comprehensive completeness score is lower than a preset threshold, the first intelligent agent dynamically determines the historical time range and behavior type to be collected for the target account based on the current observation and thinking results.

[0048] In an optional embodiment of the present invention, if it is initially determined that the account is suspected to be an abnormal account, then the focus is on collecting high-frequency interaction records in the past 7 days; if it is suspected to be a fake marketing account, then the scope is expanded to include sales-related content and comments in the past 30 days.

[0049] Step S122: Based on the determined historical time range and behavior type, use a hierarchical sampling strategy to obtain fine-grained historical content of the target account during abnormally active periods through the social platform interface.

[0050] Specifically, this step employs a tiered sampling strategy. First, it obtains a coarse-grained time distribution histogram, such as the daily posting volume, to identify abnormally active periods. Then, it focuses on high-density intervals for fine-grained content extraction. This strategy improves efficiency and reduces platform access costs.

[0051] Step S123: Perform feature extraction and structuring processing on the acquired fine-grained historical content to obtain the historical information features composed of content evolution features, behavioral rhythm features, and interactive authenticity features.

[0052] Specifically, the collected historical data will be structured into historical information features. Among these, content evolution features include topic drift, keyword stability, and text similarity sequence; behavioral rhythm features include posting interval distribution, nighttime activity ratio, and sudden behavior index; and interaction authenticity features include comment reply rate, fan interaction quality, and self-interaction ratio, such as self-evaluation and self-praise.

[0053] Step S13: Based on the comprehensive completeness score, the first agent traverses the multi-hop association network in the social behavior graph to obtain the long-distance associated account features.

[0054] Specifically, this step aims to overcome the limitations of traditional social relationship analysis based on direct neighbors (i.e., one-hop neighbors). By constructing and traversing a multi-hop association network of the target account within the social behavior graph, it aims to uncover potential collaborative malicious groups or disguised propagation chains, thereby supplementing the target account's social behavior information. Malicious accounts often exist in the form of abnormal account matrices or puppet account groups. Their individual characteristics may be carefully disguised, but their group behavior often exhibits abnormal topological patterns in the graph structure, such as high cohesion, low external connections, and synchronous operations. This step involves LLM proactively initiating graph expansion exploration to discover and aggregate features of long-distance associated accounts within the multi-hop range of the target account.

[0055] In an optional embodiment of the present invention, the discovered long-distance related account content and its aggregation features are injected as contextual evidence into the second-stage maliciousness comprehensive identification step, significantly improving the ability to identify covert, group-based malicious behavior. The entire process is triggered on demand by LLM in the ReAct loop, ensuring that graph exploration is only carried out when necessary, balancing accuracy and computational efficiency.

[0056] In an optional embodiment of the present invention, step S13 may include: Step S131: Using the target account as the root node, construct an initial adjacency graph based on the interaction behavior.

[0057] Specifically, this step uses the target account as the root node and constructs an initial adjacency graph based on its known interactive behaviors, including forwarding, commenting, mentioning, and participating in topics together. Each edge is accompanied by the behavior type, timestamp, and frequency weight.

[0058] Step S132: Use a semantic and structural dual-constraint pruning strategy to filter irrelevant nodes and identify potential collaborative malicious groups through the first intelligent agent.

[0059] Specifically, LLM dynamically determines the traversal depth and breadth based on the current information completeness assessment results. To avoid explosive growth of the graph, this step introduces a semantic-structural dual-constraint pruning strategy. The structural constraint retains only edges that meet the minimum interaction strength threshold; the semantic constraint uses an embedding model to calculate content similarity between nodes and filters out semantically irrelevant weakly related nodes. The embedding model can be Sentence-BERT.

[0060] In an optional embodiment of the present invention, the minimum interaction intensity threshold is three or more mutual comments. If it is initially detected that a large number of reposts from the same niche account are found to be from the target account, a depth-first search is performed along the reposting chain; if multiple accounts with similar wording are detected in the comment section, a breadth-first diffusion is initiated to identify potential accomplices.

[0061] Step S133: Extract the topological features, behavioral consistency features, and identity disguise clues of the collaborative malicious group to obtain the features of remotely related accounts.

[0062] Specifically, this step extracts group-level features from the final retained set of long-distance associated accounts. These features include topological features such as subgraph density, average path length, and core-periphery structure index; behavioral consistency features such as posting time synchronization rate, text template reuse rate, and similarity of target account interaction preferences; and identity spoofing clues such as similar batch registration times, highly similar avatars or nicknames, and consistent patterns of missing metadata fields.

[0063] Step S14: Based on the comprehensive completeness score, the first intelligent agent is used to identify the corresponding entities of the target account on other social platforms through cross-domain identity association, thereby obtaining cross-platform account features.

[0064] Specifically, this step aims to break down data silos across single social media platforms. Through cross-domain identity association technology, it identifies potential corresponding entities of target accounts on other mainstream social media platforms, thereby constructing a more comprehensive digital profile of users. These social media platforms include Weibo, Xiaohongshu, Douyin, Zhihu, and Bilibili. In reality, malicious account operators often employ multi-platform distribution and single-point traffic generation strategies, playing different roles on different platforms to evade regulation. Relying solely on data from a single platform can easily lead to misjudgments or missed detections.

[0065] In an optional embodiment of the present invention, this step involves the LLM proactively initiating a cross-platform related search based on clues. The clues include: (1) Explicit clues: Other platform IDs, homepage links, QR codes, etc. contained in the target account profile.

[0066] (2) Implicit clues: profile picture, nickname keywords, personal profile semantics, watermarks or exclusive wording in published content, etc.

[0067] In an optional embodiment of the invention, cross-platform discovery is also invoked on demand by the LLM in the ReAct loop, avoiding costly full-network scanning of all accounts. This step significantly enhances the ability to identify malicious accounts disguised across platforms. For example, an account may appear as a regular user on Weibo but post a large amount of fake product recommendations on Xiaohongshu; this step can supplement the account's content information with this information.

[0068] In an optional embodiment of the present invention, step S14 may include: Step S141: Extract the metadata of the target account and the cross-platform query seed set from the published content through the first intelligent agent.

[0069] Specifically, LLM first parses the target account's metadata and content, extracts potential cross-platform identifiers, and performs normalization processing to form a cross-platform query seed set. The normalization processing includes removing special characters, standardizing case, and extracting the main domain name.

[0070] Step S142: Use a multimodal matching model to perform text matching, visual matching, and behavioral verification on the target platform to confirm cross-platform associated identity.

[0071] Specifically, this step utilizes a pre-built cross-platform account mapping knowledge base and a multimodal matching model to perform candidate retrieval on the target platform. The cross-platform account mapping knowledge base contains historically confirmed fan-created account pairs.

[0072] In an optional embodiment of the present invention, the specific execution process of step S142 includes: (1) Text matching: Matching is performed based on the semantic similarity of nicknames and profiles. Specifically, large model embedding or a dedicated cross-platform encoder can be used.

[0073] (2) Visual matching: Perceptual hashing or CLIP feature comparison is performed on the headshot and cover image.

[0074] (3) Behavioral evidence: If the candidate account publishes highly similar content on multiple platforms, such as the same text and pictures, the matching confidence will be increased.

[0075] In an optional embodiment of the present invention, in order to reduce the risk of false matching, this step sets a multi-evidence fusion threshold, and cross-platform association is confirmed only when at least two independent clues reach a high confidence level.

[0076] Step S143: Obtain the metadata and published content of the associated identity on other platforms to generate cross-platform features.

[0077] Specifically, once the cross-platform identity is confirmed, the system automatically retrieves the account's metadata, content publishing records, and interaction behaviors on other platforms, ultimately generating a unified cross-platform profile.

[0078] In an optional embodiment of the present invention, the cross-platform features include: identity consistency score, used to assess whether the information on each platform is self-consistent; content strategy difference, used to identify whether there is a disconnect between the persona and the content; and cross-platform traffic redirection behavior index, such as the behavior of frequently redirecting traffic from platform A to private messages on platform B.

[0079] Step S2: Use the second intelligent agent to perform malice identification and reasoning analysis on the enhanced feature set to obtain account classification results and judgment reports.

[0080] Specifically, after completing the information enhancement in step S1, this step integrates the target account's metadata, content information, behavioral information, as well as information from distant social media accounts related to the target account and cross-platform account information obtained by the target account on other platforms, to predict whether the account is malicious based on malicious account characteristic rules. Meta-information includes nickname, bio, and number of followers; content information includes the account's posted content and comments; and behavioral information includes liking and commenting behaviors.

[0081] In an optional embodiment of the present invention, the second intelligent agent is constructed based on LLM and acts as the final judge to make a comprehensive, context-aware determination of the malice level of the target account. This stage is no longer limited to single-point content or isolated behavior, but utilizes the reasoning ability of the large language model to organically integrate multi-dimensional heterogeneous evidence such as historical behavior trajectories, long-distance social topologies, and cross-platform identity consistency to construct a high-dimensional, dynamic, and interpretable malice determination framework.

[0082] In an optional embodiment of the present invention, step S2 may include: Step S21: Perform a scan of each historical tweet sequence in the enhanced feature set and score its suspicion.

[0083] Specifically, this step is used to perform a fine-grained evaluation of individual content. To prevent malicious accounts from using content dilution strategies—that is, mixing in a large amount of normal life content to cover up occasional malicious behavior—this step scans and scores each complete historical tweet sequence collected in the first phase, using the higher-scoring content as the basis for a comprehensive assessment of the account. The historical tweet sequence includes comments, reposts, and original posts.

[0084] In an optional embodiment of the present invention, this step utilizes a large language model to evaluate each piece of content based on malicious content assessment feature dimensions. Specific assessment dimension categories include: (1) Extreme emotions and language: for example, frequent use of exclamation marks, question marks or ellipses; use of extreme emotional words; dense use of rhetorical questions or interrogative sentences; use of conspiracy theory rhetoric.

[0085] (2) Information authenticity risk: such as containing unverified assertions; missing or falsified sources; taking things out of context or misrepresenting information; spreading information that has been debunked.

[0086] (3) Aggression and personal denigration: such as insulting or mocking a specific group or individual; using labeling language; threatening or cursing expressions.

[0087] (4) Templated and mechanical repetition: For example, fixed wording structure; multiple accounts posting highly similar texts; text embedding similarity greater than a preset threshold.

[0088] (5) Inducing and manipulating intentions: For example, explicitly or implicitly guiding forwarding, liking, or reporting; creating a sense of urgency; and secretly diverting traffic.

[0089] (6) Machine-generated traces: such as low vocabulary diversity; grammatically correct but semantically empty; high-probability sampling features, such as repeated n-grams; LLM watermark detection signals.

[0090] (7) Logic breaks with common sense: for example, self-contradictory statements; violation of basic facts or scientific common sense; confusion of causal relationships.

[0091] In an optional embodiment of the present invention, each piece of content of the target account is scored for suspicion by a large language model for each of the above dimensions. The scoring criteria are shown in Table 1. The scores for each dimension are divided into five levels: 1, 2, 3, 4, and 5, which are used to describe completely inconsistent, somewhat inconsistent, neutral, somewhat consistent, and completely consistent, respectively.

[0092] Table 1. Malicious Content Assessment Feature Dimensions This step calculates the maliciousness level of the content by summing the scores from all dimensions. The top K accounts with the highest maliciousness level will be used as content information features for subsequent steps for comprehensive judgment.

[0093] Step S22: For the accounts with the highest scores, the second agent is used to make a comprehensive judgment by combining the features of various dimensions in the enhanced feature set, and the account classification results and judgment report are output.

[0094] Specifically, this step, as the expert rule injection and thought chain comprehensive judgment stage, receives the enhanced feature set from the first stage. The enhanced feature set includes: target account metadata, including nickname and profile; target account content information, including highly malicious content information and its score; historical behavior features, including content evolution features, behavior rhythm features, and interaction authenticity features; long-distance associated account features, including associated account content information, topological distance, and behavior consistency score; and cross-platform account features, including supplementary information such as the account's metadata on other platforms and content posting records.

[0095] In an optional embodiment of the present invention, content evolution characteristics include topic drift; behavioral rhythm characteristics include the proportion of nighttime activity; and interaction authenticity characteristics include comment reply rate. To ensure the interpretability, robustness, and resistance to adversarial forces in the judgment process, a multi-dimensional discrimination dimension constructed by experts is explicitly injected into the system prompts. The multi-dimensional discrimination dimensions are shown in Table 2 and include: emotional or extreme content, such as a large number of exclamation marks, rhetorical questions, and conspiracy theories; spreading unverified information, such as rumors and taking things out of context; attacking dissidents, such as personal attacks and labeling; using fixed language templates, i.e., repetitive expressions; suspicious identity, such as short registration time and abnormal fan-to-follow ratio; abnormal behavior, such as high-frequency posting in a short period of time; and guiding group behavior, such as inciting forwarding and reporting.

[0096] Table 2. Evaluation Characteristics Dimensions of Malicious Social Media Accounts In an optional embodiment of the present invention, step S22 may include: Step S221: The second intelligent agent is used to sort out the enhanced features and hit the objective anomalies in the expert rule base.

[0097] Specifically, this step corresponds to the observation stage in the deep analysis logic of the thought chain. The second agent first observes and sorts out which objective anomalies in the expert rule base were hit by the enhanced features.

[0098] Step S222: Utilize the second intelligent agent to analyze the systematic malice level of the anomaly in conjunction with the social context, and perform fault tolerance verification to find a reasonable explanation that refutes the malicious hypothesis.

[0099] Specifically, this step corresponds to the analysis and reflection phases in the deep analysis logic of the thought chain. In the analysis phase, the second agent, considering the social context, analyzes whether the objective anomalies are due to occasional individual expression habits or systemic malicious manipulation. Subsequently, in the reflection phase, fault tolerance verification is performed to attempt to find reasonable explanations to refute the malicious assumptions.

[0100] Step S223: Weigh the positive and negative evidence to reach a judgment conclusion and generate an analysis report that includes the analysis process.

[0101] Specifically, this step corresponds to the judgment stage in the deep analysis logic of the thought chain. The second intelligent agent weighs positive and negative evidence, outputs the final binary classification label, and generates a structured, logically consistent reasoning report. The binary classification label is either malicious or normal; the reasoning report is the judgment report, containing the hit feature dimensions and the analysis process.

[0102] Step S3: Based on the account classification results, analysis report, and expert annotation data, fine-tune and train the second intelligent agent to obtain an updated second intelligent agent, and apply the updated second intelligent agent to subsequent malicious social account identification.

[0103] Specifically, this section aims to address the issues of false positives and false negatives in complex business scenarios using a second intelligent agent. To this end, this step collects malicious social media accounts identified during the operation of the above-mentioned scheme, verifies genuine malicious social media accounts through human expert evaluation, and examines and corrects the reasoning process for predicting erroneous malicious social media accounts. This data is used to construct preference alignment data, which aligns the judgment results of the second intelligent agent with human expectations.

[0104] In an optional embodiment of the present invention, this step fine-tunes the second agent based on Direct Preference Optimization (DPO). Based on the execution results of the aforementioned two stages, large-scale prediction data of malicious social media accounts is accumulated. Labeled data of malicious and non-malicious social media accounts are constructed through confirmation by human experts. Through DPO fine-tuning, the large model's ability to understand and predict malicious content and ambiguous features is enhanced, reducing the false positive and false negative rates of malicious social media account identification, and enabling continuous updating of the second agent for malicious social media accounts.

[0105] In an optional embodiment of the present invention, step S3 may include: Step S31: Based on the account classification results and the judgment report, extract the account feature information of the misjudged samples as input samples, and use the correct reasoning process rewritten by the expert and the incorrect reasoning report generated by the second agent as the winning response and the losing response, respectively, to construct a triplet preference dataset.

[0106] Specifically, this step is responsible for capturing the bias in the model's judgment results. The specific method involves collecting false positive samples (i.e., misjudged samples) generated by the system during early training or testing, extracting the feature information of the target account, and denoting it as... Subsequently, the logic of human experts is introduced: the correct reasoning process, rewritten by experts based on real-world contexts, is obtained and used as the winning response, i.e., a positive preference, denoted as... Simultaneously, erroneous reasoning reports generated by the original large model that contain overreactions or logical illusions based on surface features are retained and treated as failed responses, i.e., negative preferences, denoted as... These three elements combined constitute a highly targeted triplet preference dataset. .

[0107] In an optional embodiment of the invention, the expert-rewritten correct reasoning report can accurately identify the subtle boundary between emotional venting and malicious attacks.

[0108] Step S32: Update the network weights of the second agent by minimizing the direct preference optimization loss function using the triplet preference dataset.

[0109] Specifically, this step utilizes the constructed triplet preference dataset to perform lightweight fine-tuning on the second agent undertaking the comprehensive decision-making task. In terms of algorithm implementation, this step abandons the cumbersome steps of training independent reward models in traditional reinforcement learning, instead directly minimizing the direct preference optimization loss function at the policy model level. .

[0110] In an optional embodiment of the present invention, the optimization objective formula of the above-mentioned loss function is as follows: in, This represents the network parameters of the policy model that are currently undergoing fine-tuning and optimization. This represents the initial reference model that was frozen before fine-tuning; These are hyperparameters used to control the penalty for deviations from the reference model; This represents the Sigmoid activation function, which is used to smoothly map the generated logarithmic term differences to the probability space; This represents the account feature information of the input sample; Indicates a winning response; Indicates a failure response; This represents a triplet preference dataset.

[0111] Specifically, the initial reference model serves as a baseline in computation to prevent excessive bias or forgetting of basic common sense during fine-tuning. The logarithmic ratio difference term within the formula essentially reflects the difference between the implicit reward value for correct reasoning by the model-generated expert and the implicit reward value for incorrect reasoning. By continuously minimizing this loss function through gradient descent, this step forcibly maximizes the logarithmic probability difference between the model-generated expert's winning response and the model's failing response, thereby substantially solidifying the discriminative intuition and fault-tolerance principles of human security experts into the network weights of the large model. After this alignment and optimization step, the second agent, while maintaining a high recall rate against stealthy bots, learns to protect normal, highly active users in ambiguous areas, completely overcoming the bottleneck of high false positive rates inherent in traditional second agents.

[0112] This invention also provides a malicious social media account identification system based on two-agent collaboration, such as... Figure 3 As shown, it includes: The feature expansion module is used to perform content analysis and feature expansion on the target account information using the first intelligent agent to obtain an enhanced feature set. The malicious identification module is used to perform maliciousness identification and reasoning analysis on the enhanced feature set using a second intelligent agent, and to obtain account classification results and judgment reports. The model optimization module is used to fine-tune the training of the second agent based on the account classification results, the judgment report, and expert annotation data to obtain an updated second agent, and then apply the updated second agent to subsequent malicious social account identification.

[0113] Method performance testing: This invention was tested on a malicious social media account identification task. By comparing it with existing mainstream methods on datasets of different dimensions, the effectiveness of the proposed two-stage agent identification method was verified. Specific experimental results are shown in Tables 3 and 4.

[0114] Table 3 shows a comparison of the performance of the method of the present invention with traditional machine learning models (such as random forests), pre-trained language models (such as RoBERTa-Large), graph neural network models (such as BotRGCN and RGT), and zero-shot inference of large language models.

[0115] Table 3. Comparison of the accuracy of the method of the present invention and existing methods on a self-built dataset for account recognition tasks. The experimental results in Table 3 show that the method of the present invention exhibits significantly better performance than existing benchmark methods on the self-built dataset, as detailed below.

[0116] (1) In terms of accuracy and F1 score: The accuracy of the method of this invention reaches 84.5% and the F1 score reaches 83.6%, which is significantly higher than the current best-performing Botsay model (accuracy 78.5%, F1 score 77.4%), and also far exceeds the zero-shot reasoning performance of general large language models such as Llama-3, GLM-4 and Qwen-3.

[0117] (2) False positive rate: This is a key indicator for evaluating the robustness of the identification system. The false positive rate of the method in this invention is only 3.4%, which is much lower than all other comparative models (such as Botsay's 8.5% and Qwen-3-14B's 11.5%). This proves that the present invention effectively protects normal, highly active users in the ambiguous zone and greatly reduces false positives through information enhancement in step S1 and direct preference optimization (DPO) training in step S3.

[0118] Table 4 verifies the generalization ability of the method of the present invention in different social platform environments. On the TwiBot-20 dataset, the accuracy and F1 score of the method of the present invention reached 89.1% and 89.5%, respectively; on the more difficult TwiBot-22 dataset, the method of the present invention still maintained an accuracy of 82.8% and an F1 score of 80.9%.

[0119] Table 4. Comparison of the accuracy of the method of the present invention and existing methods on public datasets for account recognition tasks. The experimental results in Table 4 show that the method of the present invention also achieves leading performance on the public datasets TwiBot-20 and TwiBot-22, as detailed below.

[0120] (1) Compared with graph neural network models that focus on social topology (such as BotRGCN and RGT), the method of this invention has an advantage in identifying malicious accounts with strong concealment by introducing cross-platform information and multi-dimensional thought chain reasoning.

[0121] (2) Compared with similar agent-based or large language model-based methods (such as Botsay and Qwen-3-14B), the method of this invention achieves a performance improvement of about 3% to 4% on both public datasets.

[0122] The above experimental results fully demonstrate that the technical solution proposed in this invention, through the collaborative work of two-stage large language models, not only improves the recognition accuracy of complex disguised accounts, but also achieves an extremely low false positive rate while ensuring a high recall rate through a closed-loop iterative fine-tuning mechanism, thus possessing strong industrial application value.

[0123] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Appropriate modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention should be covered within the protection scope of the present invention, which is defined by the claims.

Claims

1. A method for identifying malicious social account based on two-agent cooperation, characterized in that, Includes the following steps: The first intelligent agent is used to perform content analysis and feature expansion on the target account information to obtain an enhanced feature set; The second intelligent agent is used to perform malice identification and reasoning analysis on the enhanced feature set to obtain account classification results and an assessment report. The second agent is fine-tuned and trained based on the account classification results, analysis reports, and expert annotation data to obtain an updated second agent, which is then applied to subsequent malicious social media account identification.

2. The method of claim 1, wherein, The first intelligent agent performs content analysis and feature expansion on the target account information to obtain an enhanced feature set, including: The first intelligent agent is used to perform a quantitative evaluation of the information completeness of the target account information, and a comprehensive completeness score is obtained. Based on the comprehensive completeness score, the historical behavior trajectory of the target account is traced back using the first intelligent agent to obtain historical information features; Based on the comprehensive completeness score, the first agent traverses the multi-hop association network in the social behavior graph to obtain the features of long-distance associated accounts. Based on the comprehensive completeness score, the first intelligent agent is used to identify the corresponding entities of the target account on other social platforms through cross-domain identity association, thereby obtaining cross-platform account features.

3. The method as described in claim 2, characterized in that, The first intelligent agent performs a quantitative assessment of the information completeness of the target account information to obtain a comprehensive completeness score, including: Calculate the objective rule statistics score based on the total number of tweets published by the target account and the effective population ratio of metadata fields; A first intelligent agent based on a large language model is used to semantically score whether the existing information is sufficient to characterize the true personality of the account, and a semantic inference score is obtained. The objective rule statistical score and the semantic inference score are weighted and fused to obtain a comprehensive completeness score.

4. The method as described in claim 2, characterized in that, Based on the comprehensive completeness score, the historical behavior trajectory of the target account is traced back using the first intelligent agent to obtain historical information features, including: When the comprehensive completeness score is lower than a preset threshold, the first intelligent agent dynamically determines the historical time range and behavior type to be collected for the target account based on the current observation and thinking results. Based on the determined historical time range and behavior type, a stratified sampling strategy is used to obtain fine-grained historical content of the target account during abnormally active periods through social platform interfaces; The acquired fine-grained historical content is subjected to feature extraction and structuring to obtain the historical information features composed of content evolution features, behavioral rhythm features, and interaction authenticity features.

5. The method as described in claim 2, characterized in that, Based on the comprehensive completeness score, the first agent traverses the multi-hop association network in the social behavior graph to obtain long-distance associated account features, including: Using the target account as the root node, construct an initial adjacency graph based on interactive behaviors; The system uses a dual-constraint pruning strategy of semantics and structure to filter irrelevant nodes and identifies potential collaborative malicious groups through the first agent. The topological features, behavioral consistency features, and identity disguise clues of the coordinated malicious group are extracted to obtain the features of long-distance related accounts.

6. The method as described in claim 2, characterized in that, Based on the comprehensive completeness score, the first intelligent agent identifies the corresponding entities of the target account on other social platforms through cross-domain identity association, obtaining cross-platform account features, including: Extract the target account's metadata and cross-platform query seed set from the published content using the first intelligent agent; Multimodal matching models are used to perform text matching, visual matching, and behavioral verification on the target platform to confirm cross-platform identity association. Obtain the metadata and published content of the associated identity on other platforms to generate cross-platform features.

7. The method as described in claim 1, characterized in that, The enhanced feature set is analyzed for malice level using a second intelligent agent to obtain account classification results and an assessment report, including: The historical tweet sequences in the enhanced feature set are scanned one by one, and a suspicion score is assigned. For accounts with high scores, a second intelligent agent is used to make a comprehensive judgment by combining features from various dimensions in the enhanced feature set, and the account classification results and judgment report are output.

8. The method as described in claim 7, characterized in that, For accounts with high scores, a second intelligent agent is used to make a comprehensive judgment based on features from various dimensions in the enhanced feature set, outputting account classification results and an analysis report, including: By using a second intelligent agent to analyze and enhance features, objective outliers in the expert rule base were accurately identified. The systemic malice of the anomalies is analyzed by using a second intelligent agent in conjunction with social context, and fault tolerance verification is performed to find reasonable explanations to refute the malicious hypothesis. Weighing the positive and negative evidence, we arrive at a judgment and generate an analysis report that includes the analytical process.

9. The method as described in claim 1, characterized in that, The second agent is fine-tuned and trained based on the account classification results, analysis report, and expert annotation data, including: Based on the account classification results and the analysis report, the account feature information of the misjudged samples is extracted as input samples, and the correct reasoning process rewritten by the expert and the incorrect reasoning report generated by the second agent are used as winning response and losing response, respectively, to construct a triplet preference dataset. The network weights of the second agent are updated by minimizing the direct preference optimization loss function using the triplet preference dataset.

10. A malicious social media account identification system based on two-agent collaboration, characterized in that, include: The feature expansion module is used to perform content analysis and feature expansion on the target account information using the first intelligent agent to obtain an enhanced feature set. The malicious identification module is used to perform maliciousness identification and reasoning analysis on the enhanced feature set using a second intelligent agent, and to obtain account classification results and judgment reports. The model optimization module is used to fine-tune the training of the second agent based on the account classification results, the judgment report, and expert annotation data to obtain an updated second agent, and then apply the updated second agent to subsequent malicious social account identification.

Images