An Automated Data Security Assessment Method and System Based on Intelligent Agents

Abstract

This invention discloses an automated data security assessment method and system based on intelligent agents, belonging to the field of data security technology. It includes defining and constructing multiple intelligent agents, determining agent attributes, and generating capability profiles for the agents based on these attributes, providing a reliable foundation for subsequent modal analysis and resource scheduling. Multimodal analysis is performed on data assets, and resource scheduling and allocation are performed based on the agent capability profiles. Output conflicts are handled through a confidence-flow arbitration mechanism. Traditional automated tools often generate numerous false positives or false negatives due to rigid rules. This system no longer simply takes the union or intersection of values, but uses a dynamically flowing confidence algorithm for automatic adjudication. The risks of data assets are quantified based on attack graphs and Bayesian networks. Reasonable scheduling and coordination of multiple intelligent agents resolve conflicts between them, improving the efficiency and reliability of automated data security assessment and meeting complex and ever-changing security assessment needs.

Description

Technical Field

[0001] This invention relates to the field of data security technology, and in particular to an automated data security assessment method and system based on intelligent agents. Background Technology

[0002] In the deep waters of digital transformation, data has become a core production factor, but the traditional passive defense architecture of "blacklist + rule engine" is completely outdated. Faced with the explosive growth of massive heterogeneous data in cloud-native environments, the stealth of APT attacks, and the dynamic drift of business systems, the "human wave tactics" relying on human experts not only face the predicament of a shortage of high-end talent, but also fall into a security dilemma of "invisible, unmanageable, and undefended" due to slow response and high false alarm rates. An automated data security assessment solution based on intelligent agents has emerged to address this need. Its core technology lies in utilizing the semantic understanding and reasoning capabilities of Large Language Models (LLM) to endow different intelligent agents with different capabilities. Combined with knowledge graphs and reinforcement learning mechanisms, it endows the security system with a complete intelligent closed loop of "perception-cognition-decision-response".

[0003] In existing technologies, when there are many intelligent agents, the coordination between them is low and conflicts occur frequently, resulting in poor efficiency and reliability of automated data security assessment, which cannot meet the complex and ever-changing security assessment needs.

[0004] Therefore, how to improve the efficiency and reliability of automated data security assessment is a technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to address the problem of poor efficiency and reliability in automated data security assessment due to low coordination and frequent conflicts among intelligent agents in existing technologies. The proposed method is an automated data security assessment method based on intelligent agents, comprising: Define and construct multiple agents, determine agent attributes, and generate agent capability profiles based on agent attributes; Scan data assets and perform multimodal analysis on them to construct a holographic digital twin of the data assets. Based on the capability profile of the intelligent agent, allocate and schedule resources. Detect output conflicts between multiple agents and handle output conflicts through a confidence stream arbitration mechanism; Based on the output of the intelligent agent, the risks of data assets are quantified using attack graphs and Bayesian networks to generate a security report.

[0006] In some embodiments of this application, multiple intelligent agents are defined and constructed, and the attributes of the intelligent agents are determined, including... The entire process of data security assessment is obtained and broken down into multiple assessment stages. The assessment objectives for each assessment stage are determined. Multiple agent types are defined based on the assessment stages. Multiple agents are constructed with the assessment objectives as the output. Boundary processing is performed between the multiple agents. Agent attributes include static attributes and dynamic attributes; Static attributes of an intelligent agent are determined based on identity and permission information, while dynamic attributes are determined based on state, performance, and historical service information.

[0007] In some embodiments of this application, a capability profile of the intelligent agent is generated based on the agent's attributes, including... The historical service information is divided into units of each task to obtain multiple single-task service information. The single-task service information includes task requirement information and task completion quality information. The task requirement information and task completion information are analyzed separately to obtain task difficulty and task completion quality. Based on task difficulty and task completion quality, the baseline capability level of the agent under multiple tasks is calculated. A comprehensive performance index is defined by combining state information and performance information. A curve showing the change of the comprehensive performance index over a period of time is plotted. The mode interval on the curve is determined. The curve is then divided into multiple segments in chronological order. The coefficient of variation and median of each segment are calculated. An instability penalty factor is assigned to each segment based on the coefficient of variation. The overall median of the comprehensive performance index is calculated by combining the instability penalty factor and the median of each segment. The stable value of the comprehensive performance index is determined based on the overall median and the mode interval. The baseline capability level is adjusted by the stable value of the comprehensive performance index to obtain the current capability level of the agent. The static attributes, dynamic attributes, and current capability level of the agent are used as the content of the agent's capability profile to describe the agent's current capability level.

[0008] In some embodiments of this application, multimodal analysis is performed on data assets to construct a holographic digital twin of the data assets, including: The system classifies multiple modalities of data assets, extracts local and global features for each modality, defines the basic scale for different modalities, completes upsampling and downsampling of the feature pyramid through hierarchical aggregation, sets the scale quantization standard, and outputs the feature pyramid set to achieve feature scale alignment. Define the granularity of semantic units in each modality, generate candidate units, and perform cross-modal interaction to achieve cross-modal semantic granularity alignment; During the training of the multimodal mapping function, dynamic weight coefficients are calculated based on the normalized loss rate and gradient signal-to-noise ratio. The decoupling loss and variance preservation loss of each modality are calculated. The total loss is determined based on the dynamic weight coefficients, modal decoupling loss and variance preservation loss, and backpropagation is performed to achieve multimodal training balance. We construct holographic digital twins of data assets by aligning feature scales, aligning cross-modal semantic granularity, and balancing multimodal training.

[0009] In some embodiments of this application, resource scheduling and allocation are performed based on the capability profile of the intelligent agent, including... Identify the agent's overall task and break it down into multiple subtasks. Allocate subtasks according to the agent's current capability level in its capability profile, quantify the task requirement level of each subtask, and determine the matching margin based on the current capability level and the task requirement level. The basic resource quantity required by each agent is determined in a segmented manner according to the matching margin. A cost-effectiveness score is defined, and the basic resource quantity required by each agent is adjusted based on the cost-effectiveness score. The target resource quantity required by each agent is generated, and resource scheduling and allocation are performed based on the target resource quantity required by each agent.

[0010] In some embodiments of this application, output conflicts between multiple intelligent agents are detected, including: The output of the agent is standardized and normalized. The output is generated into semantic vectors through the Embedding model. The agents are divided into groups of agents to be tested according to the data asset objects processed by the agents and the categories of the agents' outputs. The cosine similarity between the semantic vectors of different agents in the group of agents to be tested is calculated. The cosine similarity is used to determine whether there is output conflict between agents.

[0011] In some embodiments of this application, output conflicts are handled through a confidence-based arbitration mechanism, including... Each agent with output conflicts undergoes atomic decomposition of its reasoning process to obtain a structured reasoning chain graph containing multiple reasoning fragments. A multi-dimensional quantitative evaluation model for the reasoning fragments is established to obtain fragment quality scores. Fragment flow is calculated using fragment quality scores and transfer functions. Cross-agent structured reasoning chain graph matching is performed to align reasoning fragments between different agents. The reasoning fragments of different agents are then merged into a fused reasoning graph for confidence flow simulation. Output conflicts are then addressed using the results of the confidence flow simulation.

[0012] In some embodiments of this application, based on the agent's output, the risk of data assets is quantified using attack graphs and Bayesian networks, including: A probabilistic attack graph is constructed based on the agent's output. A Bayesian network is then mapped based on the probabilistic attack graph to construct a conditional probability table. Evidence fusion is used to output the final probability. The attack process and state transitions are modeled using Markov chains to determine the final risk.

[0013] Correspondingly, this application also provides an automated data security evaluation system based on intelligent agents, including, The first module is used to define and build multiple intelligent agents, determine the attributes of the intelligent agents, and generate capability profiles of the intelligent agents based on the attributes of the intelligent agents. The second module is used to scan data assets, perform multimodal analysis on the data assets, construct a holographic digital twin of the data assets, and allocate resources based on the capability profile of the intelligent agent. The third module is used to detect output conflicts between multiple agents and handle output conflicts through a confidence stream arbitration mechanism. The fourth module is used to quantify the risks of data assets based on attack graphs and Bayesian networks, and generate a security report, based on the output of the intelligent agent.

[0014] Compared with the prior art, the beneficial effects of this invention are as follows: 1. Define and construct multiple intelligent agents, determine their attributes, and generate capability profiles based on these attributes. Considering the different functions and capabilities of different agents, construct their corresponding capability profiles to provide a reliable foundation for subsequent modal analysis and resource scheduling. Perform multimodal analysis on data assets and allocate resources based on the capability profiles of the agents. The system can accurately match the most suitable agent for processing based on asset characteristics. This "expert consultation" and "precise scheduling" model greatly improves the depth and efficiency of evaluation, avoids wasting computing power, and ensures that high-value assets are protected by the highest level of intelligent agents.

[0015] 2. Output conflicts are handled through a confidence-based arbitration mechanism. Traditional automated tools often generate numerous false positives or false negatives due to rigid rules. Instead of simply taking unions or intersections, the system uses a dynamically flowing confidence algorithm for automatic adjudication. This effectively filters out noisy alarms, significantly reducing the false positive rate and making the assessment results closer to real-world business scenarios. The system quantifies the risks of data assets based on attack graphs and Bayesian networks. Unlike traditional scans that only output binary results of "vulnerability presence or absence," this system uses attack graphs to deduce attack paths and combines Bayesian networks to perform probabilistic reasoning on the agent's output. This not only identifies existing vulnerabilities but also quantifies relevant risk parameters. The generated security report is no longer a cold data compilation but a dynamic risk assessment with causal logic, helping managers intuitively understand the risk transmission path and make scientific decisions on "remediation priorities." Reasonable scheduling and coordination of multiple agents resolve conflicts between them, improving the efficiency and reliability of automated data security assessments and meeting complex and ever-changing security assessment needs. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the automated data security assessment method based on intelligent agents proposed in this invention. Figure 2 This is a schematic diagram of the structure of the automated data security evaluation system based on intelligent agents proposed in this invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Reference Figure 1 An automated data security assessment method based on intelligent agents includes the following steps: Step S101: Define and construct multiple intelligent agents, determine the attributes of the intelligent agents, and generate capability profiles of the intelligent agents based on the attributes of the intelligent agents.

[0019] In this embodiment, the agent is not a traditional script program, but an AI entity with autonomous perception, cognitive decision-making, and dynamic execution. It is the smallest execution unit and responsible entity for the data security assessment task. It includes the following three dimensions: Functional dimension: It is a collection of dedicated AI models encapsulated for a specific stage in the entire data security assessment process (such as "sensitive data identification", "vulnerability scanning", "compliance audit"), with clear input and output boundaries and assessment objectives.

[0020] Attribute dimensions: It has a "digital ID card" which includes static attributes (identity ID, permission level, domain, professional field tag) and dynamic attributes (current load status, inference latency, historical task success rate, real-time confidence).

[0021] Profile Dimension: It possesses a "capability value," which is the current capability level calculated based on historical service performance and real-time performance. This capability value is dynamic and determines the level of difficulty of tasks it can undertake.

[0022] Understandably, after defining and building multiple agents, in order to experiment with the agents' performance, the agents will be made to perform multiple tasks to obtain the following service information. After certification, they will be added to the automated safety testing system.

[0023] In some embodiments of this application, multiple intelligent agents are defined and constructed, and the attributes of the intelligent agents are determined, including... The entire process of data security assessment is obtained and broken down into multiple assessment stages. The assessment objectives for each assessment stage are determined. Multiple agent types are defined based on the assessment stages. Multiple agents are constructed with the assessment objectives as the output. Boundary processing is performed between the multiple agents. Agent attributes include static attributes and dynamic attributes; Static attributes of an intelligent agent are determined based on identity and permission information, while dynamic attributes are determined based on state, performance, and historical service information.

[0024] In this embodiment, the entire process is acquired. The system first ingests the standard operating procedures (SOPs) for data security assessment, such as the national standard GB / T37988 (Data Security Capability Maturity Model) or industry-specific specifications, to obtain the complete assessment process, including asset inventory, vulnerability scanning, compliance assessment, and risk analysis. The process is then atomized, using process mining techniques or expert rules to break down the complex assessment process into indivisible "assessment steps" (e.g., sensitive data identification, database vulnerability scanning, privacy policy comparison). Each assessment step often involves multiple tasks of different natures, requiring collaboration between agents with varying capabilities. For each step, its assessment objectives are identified.

[0025] Example (Step 2: Vulnerability Scan): Target A: Identify open ports -> Define as NetScan_Agent Target B: Detect Web vulnerabilities (SQLi / XSS) -> Define as WebVuln_Agent Target C: Check the configuration baseline -> Define as Baseline_Agent Define agent types based on evaluation stages: abstract the above objectives into types, such as "scanning agent" and "auditing agent".

[0026] Construct intelligent agents with evaluation targets as outputs: instantiate specific agents, each agent is responsible for only one specific target.

[0027] To prevent data pollution caused by unauthorized actions or "illusions" by intelligent agents, the principle of least privilege must be ensured.

[0028] Implementation technology: Containerization isolation (Docker / K8sNamespace): Each agent runs in an independent container, with file system and network space isolated.

[0029] API Gateway Authentication: Issues a unique API Token to each intelligent agent, limiting the range of interfaces it can call.

[0030] For example, WebVuln_Agent can only call the scanning interface of the web application and is prohibited from accessing the write interface of the underlying database.

[0031] Data Access Control (RBAC): Write the data scope in the agent configuration, such as "Access to the 192.168.1.0 / 24 network segment is restricted" or "Read the table structure is restricted, and reading the data in the table is prohibited".

[0032] An agent's attributes are its "digital ID card" and "health record," divided into static attributes (innate) and dynamic attributes (that change later). Static attributes are written to a configuration center (such as Etcd or Nacos) when the agent is created and usually do not change during runtime.

[0033] Identity Information (IdentityInfo): Agent_ID: Globally Unique Identifier (UUID).

[0034] Agent_Name: A human-readable name (e.g., "Web Vulnerability Scanner-01").

[0035] Agent_Type: Type (scan / audit / analysis).

[0036] Version: Model version number (e.g., Llama3-70B-v1.2).

[0037] Permission information (PermissionInfo): Data_Scope: The scope of access to data assets (e.g., financial domain only).

[0038] API_Whitelist: A list of APIs that are allowed to be called.

[0039] Resource_Quota: Maximum resource limit (maximum CPU / GPU memory).

[0040] Security_Level: The security level (Top Secret / Confidential / Public) determines what level of data it can handle.

[0041] Dynamic attributes are collected in real time by the monitoring agent (AgentSidecar) and reported to the dispatch center. These include, but are not limited to: Status Information (StatusInfo): Runtime_Status: Current status (running / idle / blocked / error).

[0042] Task_Queue_Len: The number of tasks currently pending.

[0043] Last_Heartbeat: The time of the last heartbeat (used to determine whether the person is alive).

[0044] Performance Information (PerformanceInfo): Inference_Latency: Time taken for a single inference attempt (ms).

[0045] Token_Usage: LLM token consumption rate.

[0046] GPU_Util: Percentage of GPU utilization.

[0047] Error_Rate: The error rate of tasks in the last 5 minutes.

[0048] In some embodiments of this application, a capability profile of the intelligent agent is generated based on the agent's attributes, including... The historical service information is divided into units of each task to obtain multiple single-task service information. The single-task service information includes task requirement information and task completion quality information. The task requirement information and task completion information are analyzed separately to obtain task difficulty and task completion quality. Based on task difficulty and task completion quality, the baseline capability level of the agent under multiple tasks is calculated. A comprehensive performance index is defined by combining state information and performance information. A curve showing the change of the comprehensive performance index over a period of time is plotted. The mode interval on the curve is determined. The curve is then divided into multiple segments in chronological order. The coefficient of variation and median of each segment are calculated. An instability penalty factor is assigned to each segment based on the coefficient of variation. The overall median of the comprehensive performance index is calculated by combining the instability penalty factor and the median of each segment. The stable value of the comprehensive performance index is determined based on the overall median and the mode interval. The baseline capability level is adjusted by the stable value of the comprehensive performance index to obtain the current capability level of the agent. The static attributes, dynamic attributes, and current capability level of the agent are used as the content of the agent's capability profile to describe the agent's current capability level.

[0049] In this embodiment, the agent's capabilities will vary depending on the tasks it handles and the state it is in. Therefore, a current capability level is defined here based on the agent's past task performance and overall performance stability.

[0050] Task requirements (Input): Includes task type, asset size, assessment depth, number of compliance clauses, etc.

[0051] Task completion quality information (Output): includes vulnerability detection rate, false alarm rate, compliance matching degree, and manual review score.

[0052] The task requirement information and task completion quality information are quantified to obtain the task difficulty coefficient and task completion quality score. The baseline capability level is calculated by combining the task difficulty coefficient and task completion quality score (here, the exponential smoothing method is used (emphasizing recent performance) to give higher weight to recent tasks, reflecting the "timeliness" of the agent). The baseline capability level describes the baseline capability level of the agent determined by analyzing the past task completion situation.

[0053] The Comprehensive Performance Index (CPI) is constructed by combining state information (CPU / memory / queue) and performance information (latency / throughput). A real-time CPI index (0-100 points, higher is better) is generated through Principal Component Analysis (PCA) or simple weighting. The CPI data is discrete and noisy; curve analysis reveals the agent's operational patterns and resilience. The mode interval represents the agent's most frequently occurring performance state (comfort zone). The coefficient of variation describes the stability of the curve segment; a larger coefficient of variation results in a larger instability penalty factor and a greater penalty. The overall median of the CPI is calculated by combining the instability penalty factor of each curve segment with the median. Different weights are mapped based on the instability penalty factor, and a weighted median is calculated to output the overall median. The stable value of the comprehensive performance index is determined based on the overall median and mode interval. If the overall median falls within the mode interval, it is used as the stable value of the comprehensive performance index. Otherwise, the overall median is adjusted by the distance between it and the mode interval (different distances correspond to different adjustment coefficients, which are multiplied by the overall median). The baseline capability level is adjusted using the stable value of the comprehensive performance index to obtain the agent's current capability level.

[0054] ; in, This represents the agent's current (adjusted) capability level. As a baseline capability level, This represents a stable value for the overall performance index. This represents the maximum value of the overall performance index.

[0055] It should be noted that the introduction of "variance coefficient" and "penalty factor" can accurately detect performance fluctuations of the agent (such as slowed inference due to insufficient video memory), preventing the assignment of critical tasks to agents that are "seemingly idle but actually lagging". The capability profile is not static, but is updated in real time with each task execution and performance monitoring every second, realizing true "state-aware scheduling".

[0056] Step S102: Scan the data assets and perform multimodal analysis on the data assets to construct a holographic digital twin of the data assets, and allocate resources according to the capability profile of the intelligent agent.

[0057] In this embodiment, the data has multiple modalities, including text, images, traffic data, and tables. Therefore, it is necessary to perform unified feature space mapping on the modal data. When mapping heterogeneous data such as text, images, and time-series signals to the same high-dimensional semantic space (Common Embedding Space), alignment is usually required. High-dimensional data (such as the pixel matrix of a 4K image or long-time sensor waveforms) is typically compressed into fixed-length vectors (such as 768-dimensional or 1024-dimensional vectors). In this process, it is easy to lose key high-frequency details or sparse features, necessitating feature scale alignment. Moreover, different modalities may express the same concept differently, even leading to polysemy or polyillusory representations of the same image. Furthermore, during the training of the mapping function, if the data volume of a certain modality is extremely large or the feature extractor is better pre-trained (such as CLIP trained with large-scale image-text pairs), it will dominate the shared space, causing other modalities (such as numerical time-series data) to be "squeezed" to the edge regions and unable to effectively contribute information. Multimodal analysis mainly aims to solve the above three problems.

[0058] In some embodiments of this application, multimodal analysis is performed on data assets to construct a holographic digital twin of the data assets, including: The system classifies multiple modalities of data assets, extracts local and global features for each modality, defines the basic scale for different modalities, completes upsampling and downsampling of the feature pyramid through hierarchical aggregation, sets the scale quantization standard, and outputs the feature pyramid set to achieve feature scale alignment. Define the granularity of semantic units in each modality, generate candidate units, and perform cross-modal interaction to achieve cross-modal semantic granularity alignment; During the training of the multimodal mapping function, dynamic weight coefficients are calculated based on the normalized loss rate and gradient signal-to-noise ratio. The decoupling loss and variance preservation loss of each modality are calculated. The total loss is determined based on the dynamic weight coefficients, modal decoupling loss and variance preservation loss, and backpropagation is performed to achieve multimodal training balance. We construct holographic digital twins of data assets by aligning feature scales, aligning cross-modal semantic granularity, and balancing multimodal training.

[0059] In this embodiment, to address the problem of key feature loss, feature scale alignment is achieved by preserving a multi-scale feature pyramid through hierarchical mapping.

[0060] Local features: refer to the most basic and lowest-level signal units in the data.

[0061] Image: pixel blocks, edges, textures, corners (such as the edge of a screw).

[0062] Text: Word embeddings, n-gram phrases (e.g., "temperature too high").

[0063] Time series: Abrupt values ​​at a single time point, short-time frequency domain characteristics (such as vibration peak value at a certain millisecond).

[0064] Global Features: These refer to semantic concepts that have been abstracted and aggregated.

[0065] Image: semantic tags for the whole image (e.g., "This is a motor"), overall layout.

[0066] Text: sentence vectors, document topic (e.g., "This is a fault report").

[0067] Time series: long-term trends and statistical characteristics (such as "average rotation speed over the past hour").

[0068] The feature scale alignment process is as follows: 1. Definition of basic scale The finest granularity is defined based on the resolution of the original data.

[0069] Image: Patch size (e.g., 16×16 pixels).

[0070] Text: Token length (e.g., per word).

[0071] Timing: Sampling window (e.g., 10ms).

[0072] 2. Hierarchical aggregation Downsampling (compression): By using convolution or pooling, adjacent local features are merged, the scale is doubled, and the semantics become more abstract.

[0073] Upsampling (reduction): In order to enable high-level semantics to guide low-level details, high-level features are amplified and laterally connected with low-level features (Lateral Connection).

[0074] 3. Scale Quantification Standards Based on receptive field: the receptive field of low-level features is small (only looking at the local area), while the receptive field of high-level features is large (looking at the global area).

[0075] The output is no longer a 1×768 vector, but a feature pyramid set {P0,P1,P2,P3}, where P0 retains high-frequency details (high resolution, low semantics) and P3 retains macroscopic semantics (low resolution, high semantics).

[0076] To address the issue of different modalities expressing the same concept in different ways, and even the problem of "one word having multiple meanings" or "one image having multiple meanings," a fine-grained alignment method is adopted to achieve semantic alignment between different modalities.

[0077] Finer granularity is not always better. There are diminishing marginal effects of "computational explosion" and "semantic noise", so it is important to control the granularity appropriately.

[0078] The process of cross-modal semantic alignment is as follows: 1. Define semantic units Image side: Not pixels, nor the entire image, but "Region Proposals" or "Semantic Segmentation Patches" (such as meaningful blocks like "tire" or "steering wheel"). Text side: Not characters, nor entire paragraphs, but "Named Entities" or "Core Words Extracted by Dependency Parsing" (such as "pressure" or "overload"). Temporal side: "Event Segments" (such as "voltage drop interval").

[0079] Visual (image) modality: A region unit based on "perceptual grouping," abstractly defined, possessing similar textures, colors, and closed contours, typically corresponding to a physical object. Image segmentation can be performed using superpixel / graph cut methods.

[0080] Text modality: A lexical unit based on "syntax and entity," an abstract definition, and the smallest grammatical unit carrying core semantics, typically a noun, verb, or proper noun. Text segmentation is performed through dependency syntax core word extraction.

[0081] Temporal modality: Based on "statistically significant points," the event unit is an abstract definition of a segment where a signal undergoes a significant change, rather than a single sampling point. It can be segmented using change point detection; for example, when the mean / variance of the signal abruptly changes within a window W, and the magnitude of the change exceeds 3σ (standard deviation), it is segmented into an event segment.

[0082] 2. Coarse sieve First, calculate the global similarity using the highest-level features (P3, W2, S2). If the global similarity is below the threshold, it is directly determined to be irrelevant, saving computational resources.

[0083] 3. Conduct a detailed investigation Construct a unified query space: Use key entities in the text (such as "Bank") as queries.

[0084] Multi-scale scanning: Let this query "sweep through the multi-scale feature pyramid of the image".

[0085] For example, in P0 (detail layer): is there an edge texture like "riverbank"? In P1 (component layer): Is there a shape that looks like a "riverbank"? In P2 (scene layer): Is there a background that looks like a "river"? Cross-attention fusion: Calculate the attention weights of the query and each layer and region of the image.

[0086] 4. Feature enhancement and remapping The aligned information is "written back" to the original feature.

[0087] Example: The text feature "Bank" absorbs the visual features of "riverbank" in the image, becoming a new vector V. bank_river It is no longer a simple text vector. At the same time, the "riverbank" area in the image has also been labeled with the semantic tag "Bank".

[0088] When performing multimodal analysis, if the amount of data for a certain modality is extremely large or the feature extractor is better pre-trained (such as CLIP trained with large-scale graph pairs), it will dominate in the shared space, causing other modalities (such as numerical time series data) to be "squeezed" to the marginal region and unable to contribute information effectively.

[0089] Gradient balancing: In multi-task learning, the weights of gradient backpropagation are dynamically adjusted based on the rate of loss descent of each modality (such as the GradNorm algorithm) to prevent gradient explosion in one modality from suppressing other modalities.

[0090] An optimized and improved gradient balancing approach is used to address the problem of "strong mode suppressing weak mode." While the traditional GradNorm is effective, it primarily focuses on the "magnitude of the gradient," neglecting the "distribution of features" and the "dynamic stability of training." Therefore, an improved dynamic balancing strategy, Adaptive Gradient Control with Orthogonality (AGCO), is designed based on signal-to-noise ratio and feature orthogonality.

[0091] AGCO's logic is: 1. Check the signal-to-noise ratio (SNR): If the gradient of a certain mode fluctuates greatly (is unstable), simply increasing the weight will cause the model to oscillate, so it needs to be stabilized first.

[0092] 2. Consider Orthogonality: If visual features already fill the feature space, forcing temporal features to fit the visual features will only turn the temporal features into noise. The two must be forced to be "orthogonal" (not interfering with each other, each managing its own space).

[0093] Phase 1: Calculating Dynamic Task Weighting Instead of directly looking at the loss value, we look at the "relative learning rate" and the "gradient signal-to-noise ratio". We calculate the normalized loss rate, assuming there are two modalities: visual (V) and temporal (T).

[0094] ; in, , These represent the current loss values ​​for the visual (V) and temporal (T) modalities, respectively. , These are the reference losses for the two modes (usually the average loss of the first few steps is taken as the baseline). , Let be the relative descent velocities of the two modes, if... =0.1 indicates that the visual loss is only 10% of the baseline, meaning it is learning very quickly; if =0.8 indicates that the timing is still 80% complete, meaning it is learning slowly.

[0095] Introducing Gradient Signal-to-Noise Ratio (GNR) Strong mode gradients (such as CLIP) are often large but stable; weak mode gradients are small and fluctuate greatly. Calculate the mean of the gradients over the most recent K steps. and standard deviation .

[0096] ; in, For signal-to-noise ratio, To avoid a compensation constant with a denominator of 0, a high SNR indicates consistent gradient direction and stable training; a low SNR indicates erratic gradient movement and unstable training.

[0097] Calculate dynamic weight coefficients The traditional GradNorm formula is quite complex; we have simplified it and added an SNR correction: ; Then, normalization (Softmax) is performed to make... + =1. (Sensitivity factor): Typically set between 0.5 and 1.0. If For modes with a SNR greater than 0 (instable), the denominator increases, and the weights are suppressed. This is to prevent weak modes from skewing the model due to gradient erratic movements. Once the training stabilizes (SNR increases), the weights will automatically recover.

[0098] Phase 2: Feature Space Anti-Squeezing Simply adjusting the loss weights is not enough, because the feature vectors will still clash in the space. We need to add an "orthogonality penalty term" to the loss function.

[0099] 1. Modality Disentanglement Loss Extract the feature vectors for each modality: fv (visual) and ft (temporal).

[0100] Calculate their cosine similarity: ; Orthogonal penalty term: ; We hope The value is close to 0. If visual and temporal features are highly correlated (e.g., both are 0.9), it means they express the same information. The weak mode is "swallowed" by the strong mode, and this loss will increase, forcing the encoder to look for temporally unique features that are not present in the visual data.

[0101] 2. Variance Preservation Loss Strong modes tend to concentrate the variance of the feature space on a few principal components, causing the features of weak modes to be squeezed to the edge (close to 0).

[0102] Calculate the variance of each modality feature within the batch. .

[0103] Penalties: ; Among them, if the variance of the time series features Less than the threshold (This means the features have collapsed and are all crammed together), which results in a loss, forcing the model to "spread out" the temporal features and occupy more space.

[0104] Phase 3: Total Loss Function and Backpropagation The above components are combined into the final training objective: ; in, , The task weights are dynamic. They are calculated in real time using the "signal-to-noise ratio weighting method" from Phase 1 and updated every N steps. , The loss is the main task loss for each modality (e.g., classification loss, regression loss). , For hyperparameters, (Orthogonality coefficient): It is recommended to start with 0.1 and increase it as the number of training rounds increases (a little overlap is allowed at the beginning, but orthogonality is required in the later stages).

[0105] (Variance coefficient): Usually set to a small constant (e.g., 0.01), which only takes effect during feature collapse.

[0106] In some embodiments of this application, resource scheduling and allocation are performed based on the capability profile of the intelligent agent, including... Identify the agent's overall task and break it down into multiple subtasks. Allocate subtasks according to the agent's current capability level in its capability profile, quantify the task requirement level of each subtask, and determine the matching margin based on the current capability level and the task requirement level. The basic resource quantity required by each agent is determined in a segmented manner according to the matching margin. A cost-effectiveness score is defined, and the basic resource quantity required by each agent is adjusted based on the cost-effectiveness score. The target resource quantity required by each agent is generated, and resource scheduling and allocation are performed based on the target resource quantity required by each agent.

[0107] In this embodiment, a directed acyclic graph (DAG) is used to decompose the total test task into multiple sub-tasks. Task matching is performed according to preset rules, quantifying the task requirement level (including computational complexity, data throughput, and accuracy requirements, etc., obtained through comprehensive analysis), and calculating the matching margin (the ratio M between the current capability level and the task requirement level, generally greater than 1). Basic resource quantity is described by the product of the basic resource coefficient k and the predetermined resource quantity, using a statistically based dynamic segmentation, typically 4-5 segments are most suitable (balancing accuracy and scheduling complexity). Thresholds are determined based on the distribution (quantiles) of historical scheduling data combined with the business SLA. For example, if M < 0.8, the resource status is insufficient, k is 0 or allocation is rejected, strictly prohibiting allocation to avoid task failure. A "task transfer" or "manual intervention" process must be triggered. If M ≥ 1.6, the resource status is severely redundant, k is 0.8, and downsizing is implemented to proactively reduce resource allocation and save computing power for other tasks. The same process applies when M is between 0.8 and 1.6.

[0108] Cost-Performance Ratio (CPR) is not just about being "cheap," but rather about "the effective value generated per unit of resource."

[0109] ; in, Rate it for value for money. Assign a task value score (determined by business importance, such as core database assessment = 100, log auditing = 10). For resource costs, This represents the expected success rate (based on historical success rates of the capability profile). For the target resource quantity, For the reserved resource quantity, Based on the basic resource coefficient, The threshold for the cost-effectiveness rating. The correction function represents the adjustment of the basic resource quantity through cost-effectiveness scoring. The cost-effectiveness scoring threshold is used to balance the size of the correction function and control it within a reasonable range.

[0110] Step S103: Detect output conflicts between multiple agents and handle output conflicts through a confidence stream arbitration mechanism.

[0111] In this embodiment, output conflict between agents refers to situations where, for the same specific data asset object (such as an IP address or a database table) and under the same evaluation dimension (such as vulnerability scanning or compliance assessment), the evaluation conclusions generated by different agents (or the same agent at different times) are logically mutually exclusive or semantically severely divergent. Specifically, this manifests as two agents giving completely opposite judgments on the security of the same asset (e.g., Agent A determines "no risk," while Agent B determines "high-risk vulnerability"), or the difference in the risk level assessment of the same vulnerability exceeds a preset threshold (e.g., Agent A rates it "medium risk," while Agent B rates it "severe"), causing the system to be unable to directly accept a single result and necessitating an arbitration mechanism for verification.

[0112] In some embodiments of this application, output conflicts between multiple intelligent agents are detected, including: The output of the agent is standardized and normalized. The output is generated into semantic vectors through the Embedding model. The agents are divided into groups of agents to be tested according to the data asset objects processed by the agents and the categories of the agents' outputs. The cosine similarity between the semantic vectors of different agents in the group of agents to be tested is calculated. The cosine similarity is used to determine whether there is output conflict between agents.

[0113] In this embodiment, the collision detection process is as follows: Phase 1: Output Standardization Objective: To eliminate the differences in output formats among different agents and to unify the language.

[0114] Establish a standard schema: Define a unified JSON format for assessment results, including core fields: Asset_ID (Unique Asset Identifier) Check_Item (Check item, such as "SQL injection") Result_Type (Result type: Pass / Fail / Vulnerability exists) Risk_Level (Risk Level: Low / Medium / High / Severe) Description (Natural Language Description) Confidence Rule mapping and LLM normalization: For structured output (such as JSON), the field mapping is done directly through the ETL script.

[0115] For unstructured output (such as natural language text: "SQL injection risk detected" vs "SQLiVulnerabilityDetected"), invoke a lightweight LLM (such as Llama3-8B) or a rules engine to force it to be rewritten into a standard schema.

[0116] Example: Input "This server has a weak password" -> Output {"Result_Type":"Fail","Risk_Level":"High","Description":"WeakPassword"}.

[0117] Phase 2: Semantic Vectorization (Embedding) Objective: To transform textual semantics into mathematical vectors to facilitate the calculation of similarity.

[0118] Choose an Embedding model: Select a pre-trained model that has been fine-tuned in the security field (such as bge-m3 or sentinel-bert) that can understand the semantics of professional terms such as "vulnerability", "risk", and "compliance".

[0119] Generate semantic vectors: Concatenate the normalized Description field with the Risk_Level field (e.g., "HighRisk-SQLInjectionVulnerabilityDetected").

[0120] Input the Embedding model and generate a fixed-dimensional floating-point vector (e.g., 768-dimensional), denoted as Voutput.

[0121] Key technical point: To emphasize the risk level, the risk level terms can be weighted (WeightedPooling) to ensure that "high risk" and "low risk" are far enough apart in the vector space.

[0122] Phase 3: Grouping of Intelligent Agents Objective: To avoid invalid comparisons across assets and dimensions, and reduce computational load.

[0123] Construct the group key (GroupKey): Key-value pair: {Asset_ID, Check_Item_Category}.

[0124] Asset_ID: Ensures that the comparison is of the same asset.

[0125] Check_Item_Category: Ensure that the comparison is of the same type of task (e.g., compare the "Vulnerability Scan" group vs. the "Configuration Baseline" group separately, because "Port Open" and "Compliance Defects" do not conflict).

[0126] Dynamic group creation: The system maintains a real-time hash table Map <GroupKey,List<Agent_Output> >

[0127] Whenever an agent reports a result, it is added to the corresponding list of agents to be tested.

[0128] Phase 4: Cosine Similarity Calculation and Conflict Detection Objective: To identify conflicts through mathematical calculations.

[0129] Cosine similarity calculation: For any two output vectors in the same group of agents under test, the value range is [-1, 1]. In security evaluation, positive correlation is usually the focus. The closer the value is to 1, the more consistent the semantics; the closer it is to 0, the more unrelated the semantics; and the closer it is to -1, the opposite the semantics (very rare).

[0130] Conflict determination logic (double threshold method): Semantic conflict threshold (Thsem): Set a low threshold (e.g., 0.6). If the similarity is <0.6, it indicates a large semantic difference, and caution is advised.

[0131] Threshold for risk level difference: Sets the maximum allowable difference in risk level (e.g., level 2).

[0132] Phase 5: Conflict Output and Marking The system generates conflict events, which include: conflicting party IDs, asset IDs, comparison of conflict content, similarity score, and timestamp.

[0133] The event will be pushed into the "Confidence Stream Arbitration Mechanism" module (i.e., the inference chain alignment step mentioned in the previous question) for further processing.

[0134] In some embodiments of this application, output conflicts are handled through a confidence-based arbitration mechanism, including... Each agent with output conflicts undergoes atomic decomposition of its reasoning process to obtain a structured reasoning chain graph containing multiple reasoning fragments. A multi-dimensional quantitative evaluation model for the reasoning fragments is established to obtain fragment quality scores. Fragment flow is calculated using fragment quality scores and transfer functions. Cross-agent structured reasoning chain graph matching is performed to align reasoning fragments between different agents. The reasoning fragments of different agents are then merged into a fused reasoning graph for confidence flow simulation. Output conflicts are then addressed using the results of the confidence flow simulation.

[0135] In this embodiment, it no longer simply follows the "majority rule" principle of traditional voting mechanisms, but instead delves into the reasoning logic of the agent, using graph structure matching and flow simulation to determine "whose reasoning process is more credible".

[0136] I. Atomization of the Reasoning Process Objective: To open up the "black box" reasoning process of intelligent agents and break it down into observable and comparable "logical fragments".

[0137] Inference chain extraction: The agent is required to output the chain of thought (CoT) along with the conclusion.

[0138] Example: Agent A outputs "high risk". Its reasoning process is: "Port 3306 is detected to be open -> Identified as MySQL service -> Weak password detected -> Determined to be high risk".

[0139] Atomized disassembly: Using NLP parsing techniques (such as dependency parsing), the long text above is broken down into reasoning fragments.

[0140] Each fragment contains three elements: Premise: such as "port 3306 is open".

[0141] Operators / rules: such as "IF port=3306 THEN service=MySQL".

[0142] Conclusion sub-node: such as "Service=MySQL".

[0143] Constructing a structured reasoning chain graph (DAG): By treating fragments as nodes and logical flow as edges, a directed acyclic graph Gagent=(V,E) is generated.

[0144] Node attributes include fragmented text, source agent ID, timestamp, and initial confidence level.

[0145] II. Multidimensional Quantitative Evaluation of Reasoning Fragments (Quality Scoring) Objective: To assign a "quality score" to each piece of reasoning to determine whether the logical step is reliable.

[0146] Establish a multidimensional evaluation model: weighted summation of evaluation indicators in four dimensions to obtain fragment quality scores.

[0147] Logical rigor (Slogic): Whether the operator application conforms to the logical specifications (e.g., "port open" cannot directly lead to "controlled by hackers"; missing intermediate steps will result in a low score).

[0148] Strength of evidence: Whether the source of the premise information is reliable (e.g., real-time data from the "asset holographic twin" scores higher than data from the "historical cache").

[0149] Historical accuracy: The success rate of the agent in using similar operators / rules in the past (from historical service information of the capability profile).

[0150] Context relevance: Whether the fragment is applicable to the current asset scenario (e.g., "Web vulnerability rules" used on "database assets" have a low relevance score).

[0151] The higher the score, the more "hardcore" and credible the reasoning step is.

[0152] III. Constructing the Transfer Function If we consider inference fragments as nodes in a pipe network, and confidence as water flowing through the pipes, then the transfer function is the rule that defines "how the water flows from one node to the next node".

[0153] Confidence-based flow: Water pipes vary in diameter. The transfer function determines the "conduction coefficient" of the water pipe.

[0154] If the previous fragment is of high quality and its logical connection to the next fragment is tight, the transfer function will give a large coefficient (thick pipe), allowing a large amount of confidence to pass through.

[0155] If the previous fragment is of poor quality or the logical jump is too large, the transfer function will give a small coefficient (a thin pipe), or even cut off the flow. The core purpose is to allow the confidence level (the flow of information) to automatically flow to the most logically rigorous and well-supported reasoning path, rather than distributing it evenly.

[0156] A transfer function is constructed by combining the quality scores of individual fragments and the passage conditions between fragments; this is the "road condition" of the connecting edges. It measures whether the premises of v derived from the conclusions of u are natural and coherent. The passage conditions between fragments include semantic / logical similarity and structural / topological dependency. Semantic / logical similarity calculates the degree of matching between the output conclusion of u and the input premises of v. Structural / topological dependency measures the degree of dependence of v on u (implemented by introducing a gating mechanism or attention weights).

[0157] The transfer function is essentially the transfer coefficient, with a range of 0-1, and is defined as follows: ; in, For fragments To fragments The transmission coefficient between them The Sigmoid function is used to compress all numerical values ​​to the range of 0-1. For fragments Quality score, , These are the weights for semantic / logical similarity and structural / topological dependency, respectively. , Fragments To fragments The magnitudes of the parameters for semantic / logical similarity and structural / topological dependency between them. This is the bias term, and it represents the threshold for passing the basic threshold.

[0158] IV. Cross-Agent Inference Chain Matching and Alignment Structured reasoning chain graph matching: The graph edit distance (GED) or maximum common subgraph (MCS) algorithm is used.

[0159] In the group of intelligent agents to be tested, find semantically equivalent fragment nodes.

[0160] Example: Agent A's fragment "Port 3306 is Open" successfully matches Agent B's fragment "Port3306isOpen".

[0161] Inference Fragment Alignment and Fusion: Merge Node Generation: Merge successfully matched fragments into a single "Merge Inference Node".

[0162] Retaining Difference Nodes: Unmatched fragments (i.e., sources of conflict) are retained as independent "competitive nodes".

[0163] Generate a fusion inference graph: Reconnect all fusion nodes and competing nodes to form a global fusion inference graph G that includes multi-agent logic. fusion .

[0164] V. Confidence Flow Simulation Objective: To simulate the flow of "confidence" on the fusion graph and see which conclusion the final flow converges to.

[0165] Initialize traffic: Use the initial confidence of each agent (from the Capcurrent of the capability profile or the Confidence of this output) as the initial flow Fsource of the source node.

[0166] Traffic propagation rules: Multiple iterations based on PageRank or message passing mechanisms: Competition mechanism of conflicting nodes: If a node has multiple inputs (representing different conclusions from different agents), traffic will converge and compete here.

[0167] Introducing a saturation function: If a conclusion node receives too much traffic, its marginal gain decreases, preventing a single high-confidence agent from monopolizing the market (unless its reasoning quality is extremely high).

[0168] Convergence criterion: The process stops when the total graph flow change is less than the threshold ϵ or when the maximum number of iterations is reached.

[0169] Ultimately, the traffic distribution on the sink node (i.e., the final evaluation conclusion) is the arbitration result.

[0170] Result: If a "high-risk" node accounts for 70% of the traffic and a "medium-risk" node accounts for 30%, the system will determine it as "high-risk" but will need to mark it as "disputed".

[0171] VI. Conflict Handling and Output Correction Objective: To generate a final report based on the simulation results.

[0172] Result and verdict: Set a traffic advantage threshold (e.g., 60%). If a conclusion's traffic share exceeds the threshold, it is adopted directly.

[0173] If no conclusion is reached (e.g., 50% vs 50%), it is marked as "high-risk uncertainty", triggering manual review or calling a higher-level "expert agent" for retesting.

[0174] Step S104: Based on the output of the intelligent agent, the risk of the data asset is quantified using attack graphs and Bayesian networks to generate a security report.

[0175] In some embodiments of this application, based on the agent's output, the risk of data assets is quantified using attack graphs and Bayesian networks, including: A probabilistic attack graph is constructed based on the agent's output. A Bayesian network is then mapped based on the probabilistic attack graph to construct a conditional probability table. Evidence fusion is used to output the final probability. The attack process and state transitions are modeled using Markov chains to determine the final risk.

[0176] In this embodiment, risk quantification based on attack graphs and Bayesian networks is the "brain" of the entire security assessment process. It connects the discrete "vulnerabilities" detected by the agent in the previous steps into an "attack chain" and calculates the actual risk probability faced by the data assets.

[0177] I. Specific Implementation Process This process is divided into four core stages: graph construction -> network mapping -> probabilistic reasoning -> dynamic evolution.

[0178] Phase 1: Constructing the Probabilistic Attack Graph (PAG) Objective: To transform the static vulnerabilities output by the agent into dynamic, probabilistic attack path graphs.

[0179] Node definition: Asset Node: Represents data assets (e.g., database D, API interface A).

[0180] Status nodes: Represent the security status of assets (e.g., Web server_intruded=True / False, administrator privileges_acquired=True / False).

[0181] Vulnerability node: Represents the specific vulnerability detected by the agent (e.g., CVE-2023-XXXX).

[0182] Edges are defined as follows: This represents an attack action or causal relationship (e.g., using SQL injection -> gaining database privileges).

[0183] Probability assignment (core): Prior probability: Confidence derived from the agent's output.

[0184] For example: If the Agent detects an "SQL injection vulnerability" with a confidence level of 90%, then node P(Vuln_SQLi=True)=0.9.

[0185] Exploit Difficulty Probability: The probability of successful attack P(Exploit) is calculated by combining CVSS score and asset exposure surface.

[0186] Formula: P(Exploit) = BaseScore × ExposureFactor.

[0187] Output: Generate a directed graph G attack Each edge and each node in the graph carries a probability value.

[0188] Phase 2: Bayesian Network (BN) Mapping and CPT Construction Objective: To transform the attack graph into a Bayesian network and use conditional probability to handle uncertainty.

[0189] Topology mapping: PAG is directly mapped to a Bayesian network structure (DAG).

[0190] Parent node: Prerequisites for the attack (e.g., the existence of vulnerabilities, the availability of attack tools).

[0191] Child nodes: the result of the attack (e.g., gaining privileges, data leakage).

[0192] Construct a conditional probability table (CPT): This is the core of Batch Normalization (BN), which describes "the probability of a child node occurring given that the parent node has occurred".

[0193] Noisy-OR gate model (commonly simplified model): Assume that each parent node (attack path) has an independent probability of causing a change in the state of its child nodes.

[0194] Example: Child node: Obtain root privileges.

[0195] Parent node A: Has a privilege escalation vulnerability (P=0.8), Parent node B: Weak password (P=0.6).

[0196] CPT calculates the overall probability of obtaining root privileges when either A or B exists.

[0197] Phase 3: Evidence Fusion and Probability Update (Bayesian Inference) Objective: To input the outputs of multiple agents as "evidence" into the network and update the posterior probability.

[0198] Evidence node injection: The results detected by the agent (such as "port 3306 is open" or "weak password found") are injected into the leaf nodes of BN as observation evidence E.

[0199] Bayes' theorem calculation: Update the probabilities of intermediate and root nodes using Bayes' theorem: Message passing algorithms (such as JunctionTreeAlgorithm): In BN, the probability of propagation is determined by both bottom-up (prediction) and top-down (diagnosis).

[0200] When multiple agents have conflicting outputs about the same node (e.g., Agent A says there is a vulnerability, while Agent B says there isn't), Batch Normalization (BN) will merge them based on their respective confidence weights.

[0201] Output: Obtain the posterior risk probability P(Risk) for each asset node.

[0202] Phase 4: Markov Chain Modeling and State Transitions (Dynamic Risks) Objective: Not only to observe the current state, but also to predict how the attack will evolve over time.

[0203] State space definition: Define the discrete state set of the system as S = {S0, S1, ..., Sn}.

[0204] S0: Safe state (no intrusion).

[0205] S1: Initial penetration (e.g., obtaining ordinary user privileges).

[0206] S2: Privilege escalation successful (administrator privileges acquired).

[0207] Sn: Objective achieved (data breach / ransom).

[0208] Transition Matrix: Construct a matrix T, where the elements T ij This represents the probability of transitioning from state i to state j.

[0209] T ij It is determined by the attack success probability calculated by BN.

[0210] For example: the probability of going from S0 to S1 = P (successful exploitation).

[0211] Risk Calculation: Absorption state probability: The probability that the system will eventually reach the absorption state of "data leakage".

[0212] Expected loss: Risk = P(leakage) × Asset_Value × Impact_Factor.

[0213] MTTC (Mean Attack Time): Calculates the expected attack time using the step size of the Markov chain.

[0214] II. Generating a Security Report Based on the above quantitative results, the system automatically generates a structured security assessment report, including: Risk Heatmap: Based on the posterior probability of each asset calculated by BN, the risk level is visualized using red / yellow / green.

[0215] Attack path simulation: Showing the top 3 most probable attack paths (e.g., external web -> SQL injection -> internal database).

[0216] Quantified risk value: Provide a specific numerical value (e.g., core database leakage risk value = 85.2 points, leakage probability = 12%).

[0217] Sensitivity analysis: Identify which vulnerabilities, when patched, will minimize the overall risk (causal analysis based on BN).

[0218] Confidence level explanation: The confidence interval of the risk value is indicated (e.g., 95% CI [10%, 14%]), which reflects the scientific nature of the results.

[0219] Correspondingly, this application also provides an automated data security evaluation system based on intelligent agents, such as... Figure 2 As shown, including, The first module is used to define and build multiple intelligent agents, determine the attributes of the intelligent agents, and generate capability profiles of the intelligent agents based on the attributes of the intelligent agents. The second module is used to scan data assets, perform multimodal analysis on the data assets, construct a holographic digital twin of the data assets, and allocate resources based on the capability profile of the intelligent agent. The third module is used to detect output conflicts between multiple agents and handle output conflicts through a confidence stream arbitration mechanism. The fourth module is used to quantify the risks of data assets based on attack graphs and Bayesian networks, and generate a security report, based on the output of the intelligent agent.

[0220] Compared with the prior art, the beneficial effects of this invention are as follows: 1. Define and construct multiple intelligent agents, determine their attributes, and generate capability profiles based on these attributes. Considering the different functions and capabilities of different agents, construct their corresponding capability profiles to provide a reliable foundation for subsequent modal analysis and resource scheduling. Perform multimodal analysis on data assets and allocate resources based on the capability profiles of the agents. The system can accurately match the most suitable agent for processing based on asset characteristics. This "expert consultation" and "precise scheduling" model greatly improves the depth and efficiency of evaluation, avoids wasting computing power, and ensures that high-value assets are protected by the highest level of intelligent agents.

[0221] 2. Output conflicts are handled through a confidence-based arbitration mechanism. Traditional automated tools often generate numerous false positives or false negatives due to rigid rules. Instead of simply taking unions or intersections, the system uses a dynamically flowing confidence algorithm for automatic adjudication. This effectively filters out noisy alarms, significantly reducing the false positive rate and making the assessment results closer to real-world business scenarios. The system quantifies the risks of data assets based on attack graphs and Bayesian networks. Unlike traditional scans that only output binary results of "vulnerability presence or absence," this system uses attack graphs to deduce attack paths and combines Bayesian networks to perform probabilistic reasoning on the agent's output. This not only identifies existing vulnerabilities but also quantifies relevant risk parameters. The generated security report is no longer a cold data compilation but a dynamic risk assessment with causal logic, helping managers intuitively understand the risk transmission path and make scientific decisions on "remediation priorities." Reasonable scheduling and coordination of multiple agents resolve conflicts between them, improving the efficiency and reliability of automated data security assessments and meeting complex and ever-changing security assessment needs.

[0222] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented in hardware or by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0223] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0224] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0225] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automated data security assessment method based on intelligent agents, characterized in that, include, Define and construct multiple agents, determine agent attributes, and generate agent capability profiles based on agent attributes; Scan data assets and perform multimodal analysis on them to construct a holographic digital twin of the data assets. Based on the capability profile of the intelligent agent, allocate and schedule resources. Detect output conflicts between multiple agents and handle output conflicts through a confidence stream arbitration mechanism; Based on the output of the intelligent agent, the risks of data assets are quantified using attack graphs and Bayesian networks to generate a security report.

2. The automated data security assessment method based on intelligent agents according to claim 1, characterized in that, Define and construct multiple agents, and determine agent attributes, including: The entire process of data security assessment is obtained and broken down into multiple assessment stages. The assessment objectives for each assessment stage are determined. Multiple agent types are defined based on the assessment stages. Multiple agents are constructed with the assessment objectives as the output. Boundary processing is performed between the multiple agents. Agent attributes include static attributes and dynamic attributes; Static attributes of an intelligent agent are determined based on identity and permission information, while dynamic attributes are determined based on state, performance, and historical service information.

3. The automated data security assessment method based on intelligent agents according to claim 2, characterized in that, Generate a capability profile of the agent based on its attributes. include, The historical service information is divided into units of each task to obtain multiple single-task service information. The single-task service information includes task requirement information and task completion quality information. The task requirement information and task completion information are analyzed separately to obtain task difficulty and task completion quality. Based on task difficulty and task completion quality, the baseline capability level of the agent under multiple tasks is calculated. A comprehensive performance index is defined by combining state information and performance information. A curve showing the change of the comprehensive performance index over a period of time is plotted. The mode interval on the curve is determined. The curve is then divided into multiple segments in chronological order. The coefficient of variation and median of each segment are calculated. An instability penalty factor is assigned to each segment based on the coefficient of variation. The overall median of the comprehensive performance index is calculated by combining the instability penalty factor and the median of each segment. The stable value of the comprehensive performance index is determined based on the overall median and the mode interval. The baseline capability level is adjusted by the stable value of the comprehensive performance index to obtain the current capability level of the agent. The static attributes, dynamic attributes, and current capability level of the agent are used as the content of the agent's capability profile to describe the agent's current capability level.

4. The automated data security assessment method based on intelligent agents according to claim 1, characterized in that, Perform multimodal analysis on data assets to construct holographic digital twins of the data assets, including: The system classifies multiple modalities of data assets, extracts local and global features for each modality, defines the basic scale for different modalities, completes upsampling and downsampling of the feature pyramid through hierarchical aggregation, sets the scale quantization standard, and outputs the feature pyramid set to achieve feature scale alignment. Define the granularity of semantic units in each modality, generate candidate units, and perform cross-modal interaction to achieve cross-modal semantic granularity alignment; During the training of the multimodal mapping function, dynamic weight coefficients are calculated based on the normalized loss rate and gradient signal-to-noise ratio. The decoupling loss and variance preservation loss of each modality are calculated. The total loss is determined based on the dynamic weight coefficients, modal decoupling loss and variance preservation loss, and backpropagation is performed to achieve multimodal training balance. We construct holographic digital twins of data assets by aligning feature scales, aligning cross-modal semantic granularity, and balancing multimodal training.

5. The automated data security assessment method based on intelligent agents according to claim 3, characterized in that, Resource scheduling and allocation are performed based on the capabilities of the intelligent agent, including: Identify the agent's overall task and break it down into multiple subtasks. Allocate subtasks according to the agent's current capability level in its capability profile, quantify the task requirement level of each subtask, and determine the matching margin based on the current capability level and the task requirement level. The basic resource quantity required by each agent is determined in a segmented manner according to the matching margin. A cost-effectiveness score is defined, and the basic resource quantity required by each agent is adjusted based on the cost-effectiveness score. The target resource quantity required by each agent is generated, and resource scheduling and allocation are performed based on the target resource quantity required by each agent.

6. The automated data security assessment method based on intelligent agents according to claim 1, characterized in that, Detecting output conflicts between multiple agents, including, The output of the agent is standardized and normalized. The output is generated into semantic vectors through the Embedding model. The agents are divided into groups of agents to be tested according to the data asset objects processed by the agents and the categories of the agents' outputs. The cosine similarity between the semantic vectors of different agents in the group of agents to be tested is calculated. The cosine similarity is used to determine whether there is output conflict between agents.

7. The automated data security assessment method based on intelligent agents according to claim 1, characterized in that, Output conflicts are handled through a confidence-based arbitration mechanism, including... Each agent with output conflicts undergoes atomic decomposition of its reasoning process to obtain a structured reasoning chain graph containing multiple reasoning fragments. A multi-dimensional quantitative evaluation model for the reasoning fragments is established to obtain fragment quality scores. Fragment flow is calculated using fragment quality scores and transfer functions. Cross-agent structured reasoning chain graph matching is performed to align reasoning fragments between different agents. The reasoning fragments of different agents are then merged into a fused reasoning graph for confidence flow simulation. Output conflicts are then addressed using the results of the confidence flow simulation.

8. The automated data security assessment method based on intelligent agents according to claim 3, characterized in that, Based on the agent's output, the risk of data assets is quantified using attack graphs and Bayesian networks, including... A probabilistic attack graph is constructed based on the agent's output. A Bayesian network is then mapped based on the probabilistic attack graph to construct a conditional probability table. Evidence fusion is used to output the final probability. The attack process and state transitions are modeled using Markov chains to determine the final risk.

9. An automated data security assessment system based on intelligent agents, characterized in that, include, The first module is used to define and build multiple intelligent agents, determine the attributes of the intelligent agents, and generate capability profiles of the intelligent agents based on the attributes of the intelligent agents. The second module is used to scan data assets, perform multimodal analysis on the data assets, construct a holographic digital twin of the data assets, and allocate resources based on the capability profile of the intelligent agent. The third module is used to detect output conflicts between multiple agents and handle output conflicts through a confidence stream arbitration mechanism. The fourth module is used to quantify the risks of data assets based on attack graphs and Bayesian networks, and generate a security report, based on the output of the intelligent agent.

Images