An ai ethics risk monitoring and governance system based on robust artificial intelligence

Abstract

The present application relates to the technical field of AI ethical risk governance, in particular to an AI ethical risk monitoring and governance system based on robust artificial intelligence, which comprises identifying abnormal jump behavior, extracting semantic conflict instructions and cross-execution events, aggregating high-frequency risk statements, tracking response misplacement paths, classifying offset response content, and forming an ethical risk governance result set based on user identity and function call records. The present application compares the coherence and time difference of function call jumps, extracts low-difference abnormal paths, identifies behavior conflicts by combining semantic opposition and time cross, classifies guiding, ambiguous and permission mismatch statements to form a high-frequency risk set, fuses multi-modal signal start order to extract response misplacement paths, constructs a collaborative review chain, and enhances the breadth and level of risk identification.

Description

Technical Field

[0001] This invention relates to the field of AI ethical risk governance technology, and in particular to an AI ethical risk monitoring and governance system based on robust artificial intelligence. Background Technology

[0002] The field of AI ethical risk governance technology involves identifying and managing the ethical risks that may arise from the practical application of artificial intelligence systems. With the widespread application of AI technology, effectively addressing the ethical issues that may arise in AI data processing, algorithm design, and application scenarios has become an important research direction in this field. The core content of this field includes four main types of risks: data ethical risks, algorithmic ethical risks, application ethical risks, and governance ethical risks. Its goal is to ensure the fairness, transparency, and security of AI systems, avoiding issues such as unfair algorithmic decisions, data leaks, and privacy violations, while ensuring that the application of AI technology complies with legal, ethical, and social responsibility requirements.

[0003] The AI ​​Ethical Risk Monitoring and Governance System based on Robust Artificial Intelligence refers to a system that comprehensively identifies and effectively governs AI ethical risks through robust AI technology. It meticulously categorizes core AI ethical risks, covering four main areas: data ethics, algorithm ethics, application ethics, and governance ethics, and further breaks down each risk category into specific risk points. Through robust technology, the system can handle data changes in real time and address dynamic issues such as algorithm updates and data distribution drift, effectively resisting noise interference and malicious attacks. The system learns from historical risk cases and combines them with real-time data to construct a dynamically updated risk feature database, and uses data fusion and anomaly detection methods to accurately identify various ethical risks. Finally, the system prioritizes risks based on their impact and probability of occurrence, providing a scientific basis for risk prevention and control.

[0004] Existing technologies lack comparative analysis of jump rhythm and time difference in the functional module operation chain, making it difficult to distinguish continuous abnormal paths. In scenarios with multiple overlapping instructions, they lack the ability to screen conflict events based on time intersection relationships. They have not formed a systematic classification and frequency aggregation for risk content with ambiguity and permission mismatch characteristics, making it difficult to sort out risk information in a hierarchical manner. At the same time, they lack the means to verify the start sequence of device response and multimodal signals, making it difficult to trace the source when response paths and behavioral actions are disconnected, thus limiting the completeness and timeliness of risk management. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI ethical risk monitoring and governance system based on robust artificial intelligence.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an AI ethical risk monitoring and governance system based on robust artificial intelligence, the system comprising:

[0007] Call the difference recognition module to extract user identifiers, functional modules and jump records, calculate the time difference of frequent jump behaviors between functions, filter records with continuous differences less than the set lower limit, and integrate their interface tags to generate an ethical jump abnormal path set;

[0008] The conflict behavior localization module calls the set of abnormal ethical jump paths, searches for combinations of action tags with opposite semantics, filters conflicting operations with overlapping execution times under the same path, extracts abnormal execution lines according to time coverage, and generates a set of ethical action conflict events.

[0009] The risk semantic aggregation module calls up expression fragments in the set of ethical action conflict events, identifies leading words and vague liability terms, clusters semantically similar expressions and counts their occurrences, filters expressions with repetitions exceeding a threshold, marks their semantic risk level, and generates a high-frequency aggregation list of ethical expressions.

[0010] The modal output tracking module calls the time period in the high-frequency aggregation list of ethical expressions, matches image frames, voice and device response, compares the time difference of the signal start order, filters records whose response is earlier than the input signal and lacks image matching, and generates an abnormal trajectory set of ethical output response.

[0011] As a further aspect of the present invention, the set of abnormal ethical transition paths includes combinations of abnormal transition labels, markers of continuous low-difference time periods, and transition offset structures between functional units; the set of ethical action conflict events includes semantic conflict instruction groups, cross-execution time coverage segments, and semantic labels of conflict events; the list of high-frequency aggregations of ethical expressions includes high-frequency sensitive expression statements, guiding and ambiguous semantic categories, and permission mismatch intent markers; and the set of abnormal ethical output response trajectories includes abnormal response timing segments, missing image action records, and prematurely triggered device response paths.

[0012] As a further aspect of the present invention, the invocation of the difference recognition module includes:

[0013] The identity information extraction submodule extracts the user's identity number, functional module identifier, and jump path label based on the user's operation log in the system. It then combines the timestamp information of each record to construct a corresponding dataset. By sorting the dataset by timestamp, it obtains the sequential structure of functional unit calls and generates a functional call time series dataset.

[0014] The jump behavior sequence construction submodule calls the function to call the continuous time points in the time series dataset and their corresponding unit labels, establishes a path mapping relationship for adjacent unit labels, calculates their corresponding time difference, constitutes the mapping entries between jump paths and time intervals, and forms a jump path time difference sequence table.

[0015] The time difference judgment submodule sets a minimum threshold standard for jump response based on the difference information recorded in the jump path time difference sequence table. It filters and merges the tags of path records that are continuously lower than the time difference lower limit, extracts the corresponding paths to form a set, and performs deduplication and sequence structure recombination on the path tags to obtain the set of ethical jump abnormal paths.

[0016] As a further aspect of the present invention, the conflict behavior localization module includes:

[0017] The tag sequence parsing submodule extracts the action instructions and their timestamp information corresponding to each node in the path based on the path tags in the set of abnormal ethical jump paths, constructs the correspondence structure between action tags and time fields, establishes the number mapping and time sequence arrangement relationship according to the path order, and generates an instruction tag time sequence matrix.

[0018] The semantic conflict identification submodule calls the action instruction tags in the instruction tag time series matrix, performs logical direction judgment on the instruction meaning between tags according to the established semantic judgment rules, filters the tag combinations with semantic opposition, records the marking results in the form of key values, and generates a set of semantic conflict tag pairs.

[0019] The cross-execution judgment submodule extracts the execution start and end times of the corresponding instruction label time series matrix based on the semantic conflict label pair set that has been determined to be opposing, calculates the proportion of the time period intersection in the total path duration, and marks and extracts records with coverage exceeding the cross threshold standard to obtain the set of ethical action conflict events.

[0020] As a further aspect of the present invention, the risk semantic aggregation module includes:

[0021] The expression content extraction submodule extracts corresponding text fragments, voice data and instruction description information based on the event items in the set of ethical action conflict events, unifies the format and reorganizes the order of the multi-source content, encodes it in combination with execution time tags, and performs standardized conversion according to the channel to generate a standardized corpus of expression content.

[0022] The risk semantic recognition submodule calls the text and speech expression data in the standardized corpus of the expression content, performs semantic unit decomposition and feature annotation on each content according to the risk dictionary entries and intent identification rules, filters expression units with related guidance, semantic ambiguity or permission mismatch features, divides the structure according to the classification labels, and obtains the risk expression semantic classification results.

[0023] The high-frequency content aggregation submodule, based on the set of semantic tags marked in the risk expression semantic classification results, groups and counts the expression units that appear repeatedly, calculates the frequency value of each expression in the corpus, and performs summary and marking operations on the statement units whose frequency exceeds the set occurrence threshold to obtain a high-frequency aggregation list of ethical expressions.

[0024] As a further aspect of the present invention, the modal output tracking module includes:

[0025] The time-segment signal extraction submodule collects the corresponding time location data based on the risk expression statements marked in the high-frequency aggregation list of ethical expressions, retrieves visual image frame sequences, audio action sequences and device response logs within the same time period, integrates the three types of signals according to timestamps and establishes corresponding index relationships, and generates a joint record set of modal signals in the same segment.

[0026] The multi-mode signal timing submodule calls the visual frames, audio actions, and device response signals in the joint record set of the same modal signal, establishes a sequence based on the start time field of each type of signal, compares the signal start order one by one according to the time difference standard, filters the records where the device response is earlier than the audio or image trigger, and obtains a list of abnormal signal start order.

[0027] The response mismatch identification submodule detects the correspondence between the device response signal and the visual frame sequence action based on the records in the signal start-up sequence anomaly list, removes items with image matching actions, and extracts records that simultaneously meet the conditions of early response start and no corresponding image action to obtain the ethical output response anomaly trajectory set.

[0028] As a further aspect of the present invention, the system further includes:

[0029] The monitoring and governance classification module calls the set of abnormal ethical output responses, sorts out the differences between the expressed content and the response number, archives and numbers the abnormal response groups for risk, and constructs a governance data master table after labeling and classifying them to generate a set of ethical risk monitoring and governance results.

[0030] The ethical risk monitoring and governance result set includes response diversity offset samples, expression classification numbers to be governed, and ethical risk content governance codes.

[0031] As a further aspect of the present invention, the monitoring and management classification module includes:

[0032] The response offset extraction submodule, based on the identified response numbers, expression content and behavioral instruction records in the ethical output response anomaly trajectory set, retrieves and matches the response behaviors corresponding to the same instruction by number, identifies expression instances with different response performances, and extracts record units with behavioral state offset features by number as index, generating a response diversity offset instance set.

[0033] The governance fragment labeling submodule calls the expression data in the response diversity offset instance set, performs structural parsing based on the role call order and permission expression field, screens content items with subject-object configuration conflicts and permission boundary offsets, assigns a tag to be governed to the expression instances that meet the conditions, and binds the number to establish a correspondence with the semantic fragment to obtain the expression fragment identifier set to be governed.

[0034] The content structure classification submodule, based on the semantic classification information and expression fields recorded in the set of expression fragment identifiers to be governed, encodes each expression instance according to behavioral tags and semantic patterns, constructs a unified field mapping structure, and archives all marked expressions after merging and reconstructing them according to field groups to obtain the ethical risk monitoring and governance result set.

[0035] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0036] In this invention, by comparing the jump coherence and time difference sequence between function calls, a set of low time difference abnormal paths is formed. By combining the cross-coverage ratio of semantically contradictory instructions and execution time, potential behavioral conflict events are extracted. Semantic classification and repetition statistics are carried out using guiding, ambiguous, and permission mismatch features to form a set of high-frequency risk content. Furthermore, by integrating the differences in the start-up order of visual frames, audio actions, and device response signals, response misalignment paths are extracted. Finally, a collaborative verification chain among multimodal signals is constructed to improve the comprehensiveness, dynamism, and hierarchical management capabilities of risk identification. Attached Figure Description

[0037] Figure 1 This is a system flowchart of the present invention;

[0038] Figure 2 This is a flowchart illustrating the process of calling the difference recognition module in this invention.

[0039] Figure 3 This is a flowchart illustrating the acquisition process of the conflict behavior localization module of the present invention.

[0040] Figure 4 This is a flowchart illustrating the acquisition process of the risk semantic aggregation module of the present invention.

[0041] Figure 5 This is a flowchart illustrating the acquisition process of the modal output tracking module of the present invention.

[0042] Figure 6 This is a flowchart illustrating the acquisition process of the monitoring and management classification module of this invention. Detailed Implementation

[0043] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0044] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0045] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0046] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0047] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0048] Please see Figure 1 This invention provides a technical solution: an AI ethical risk monitoring and governance system based on robust artificial intelligence, the system comprising:

[0049] The difference identification module is invoked to retrieve the user's identity number in the system, the functional unit used, and the jump record. The continuity of the rapid jump behavior between different functional units is checked, the time difference sequence in continuous calls is analyzed, and operation records with differences continuously lower than the standard lower limit are marked as abnormal. The call tags are summarized to form a jump deviation set, and an ethical jump abnormal path set is generated.

[0050] The conflict behavior localization module, based on the tag sequence of the centralized summary of abnormal ethical jump paths, searches for semantically contradictory instruction combinations between actions within the path, analyzes whether semantically conflicting instruction groups have cross-execution in the time dimension, extracts cases where the cross-execution period coverage exceeds the limit as conflict events, and generates a set of ethical action conflict events.

[0051] The risk semantic aggregation module extracts text fragments, voice content and instruction descriptions from the set of ethical action conflict events, identifies sensitive words and classifies risk intentions in the content, categorizes and counts expressions with guiding, ambiguous or permission mismatch characteristics according to semantics, summarizes and marks risk content that appears more than a set threshold, and generates a high-frequency aggregation list of ethical expressions.

[0052] The modal output tracking module calls the time location of risk expressions in the high-frequency aggregation list of ethical expressions, retrieves visual frames, audio actions and device responses within the same segment, performs time-synchronized analysis on the start order of the three types of signals, screens records where device responses precede behavioral instructions and lack corresponding image actions, extracts them as response mismatch paths, and generates an abnormal trajectory set of ethical output responses.

[0053] The monitoring and governance classification module, by combining the response numbers, content, and instruction execution status of abnormal ethical output response trajectories, identifies expression instances with response diversity deviations, marks them as fragments to be governed, and categorizes and codes them according to content characteristics, summarizing them into an ethical risk monitoring and governance result set that can be used for review and implementation governance.

[0054] The set of abnormal ethical transition paths includes combinations of abnormal transition labels, continuous low-difference time period markers, and transition offset structures between functional units. The set of ethical action conflict events includes semantic conflict instruction groups, cross-execution time coverage segments, and semantic labels of conflict events. The list of high-frequency ethical expressions includes high-frequency sensitive expression statements, guiding and ambiguous semantic categories, and permission mismatch type intent markers. The set of abnormal ethical output response trajectories includes abnormal response timing segments, missing image action records, and premature device response triggering paths. The set of ethical risk monitoring and governance results includes response diversity offset samples, expression classification numbers to be governed, and ethical risk content governance codes.

[0055] Please see Figure 2 Calling the difference recognition module includes:

[0056] The identity information extraction submodule extracts the user's identity number, functional unit identifier, and jump path label based on the user's operation log in the system. It also constructs a corresponding dataset by combining the timestamp information of each record. By sorting the set by timestamp, it obtains the sequential structure of functional unit calls and generates a functional call time series dataset.

[0057] The system processes all user operation logs recorded by the backend server (as shown in Table 1). It extracts the user ID, functional unit identifier, redirection path label, and informed consent status label from the data source. Each operation record, combined with its timestamp information, is transformed into a structured five-element data set. For example, the log record "Timestamp: 1677628801, User ID: U001, Operation Unit: M05, Path: / user / login, Agreement Status: Agreed" is transformed into (U001, M05, / user / login, Agreed). The time stamp 1677628801 is used to simultaneously perform preliminary verification of data collection compliance. If the agreement status label is not 'Agreed', the record will be marked as a potential data ethics risk. All transformed five-element data sets are sorted in ascending order by timestamp to form a functional unit call sequence that accurately reproduces the user's behavior trajectory. For example, the call sequence for user U001 is recorded as [(1677628801,M05,Agreed),(1677628802,M12,Agreed),(1677628804,M09,Agreed)]. Integrating the sequences of all users, a functional call time series dataset is finally generated.

[0058] Table 1: Example of User Operation Log Data

[0059]

[0060] As shown in Table 1, this table displays the key fields extracted from the system operation log and their example values, providing raw data input for subsequent behavioral sequence and data ethics risk analysis.

[0061] The jump behavior sequence construction submodule calls the function to call the continuous time points in the time series dataset and their corresponding unit labels, establishes a path mapping relationship for adjacent unit labels, calculates their corresponding time difference, and forms a mapping entry between jump path and time interval, forming a jump path time difference sequence table.

[0062] Using the generated function call time series dataset, jump paths are constructed by traversing consecutive elements in the sequence. For each consecutive time point, such as (1677628801, M05, Ageed) and (1677628802, M12, Ageed), a path mapping relationship M05->M12 is established, and the timestamp difference is calculated. Seconds are recorded, and the consent status at both ends of the path is also recorded. This process pays special attention to changes in the consent status. If a jump from 'Agreed' to 'Expired' occurs, the entry will be marked as a potential data sharing ethical risk. For example, the mapping entry generated by user U002's jump (M05,Agreed)->(M15,Expired) is (M05->M15,1s,Agreed->Expired,RiskFlag=1). After processing all user sequences, all generated mapping entries are aggregated to form a jump path time difference sequence table.

[0063] The time difference judgment submodule sets a minimum threshold standard for jump response based on the difference information recorded in the jump path time difference sequence table. It filters and merges the tags of path records that are continuously lower than the time difference lower limit, extracts the corresponding paths to form a set, and performs deduplication and sequence structure recombination on the path tags to obtain the set of ethical jump abnormal paths.

[0064] To determine the redirection response speed, a robust minimum threshold was set. This threshold (450 milliseconds in this example) was determined based on 30,000 response time data points under different loads, and adversarial training was conducted using a time feature library containing 50 known attack patterns. The threshold was determined after analysis using robust anomaly detection models such as Isolation Forest, which can effectively distinguish between normal user operations and automated abnormal behavior. The system filters records in the redirection path time difference sequence list whose time interval is lower than this threshold for three consecutive times or more. For example, consecutive records (M01->M02, 410ms), (M02->M07, 430ms), (M07->M11, 425ms) are selected. The selected records are assigned a common merge label, and their corresponding paths are deduplicated and recombined to form a complete continuous abnormal redirection path, such as M01->M02->M07->M11. All such paths ultimately constitute the ethical redirection abnormal path set.

[0065] Please see Figure 3 The conflict behavior localization module includes:

[0066] The tag sequence parsing submodule extracts the action instructions and their timestamp information corresponding to each node in the path based on the path tags in the set of abnormal ethical jump paths, constructs the correspondence structure between action tags and time fields, establishes the number mapping and time sequence arrangement relationship according to the path order, and generates the instruction tag time sequence matrix.

[0067] Analyzing specific paths within the set of abnormal ethical transition paths, such as M01->M02->M07->M11, we extract the corresponding action commands, timestamps, and algorithmic ethical risk classifications for each node in the path within that time period from the system command log and the algorithm fairness rule base. For example, M01 corresponds to the command "ApproveLoan," with a timestamp of 1677628810.100 and a risk classification of "high fairness risk"; M11 corresponds to the command "DenyLoan," with a timestamp of 1677628811.495. The risk classification is "high fairness risk". The above information is constructed into a structured list, such as [('ApproveLoan', 1677628810.100, 'HighRisk'), ..., ('DenyLoan', 1677628811.495, 'HighRisk')]. After establishing a numbering mapping starting from 1 for each instruction according to the path order, an N-row, 3-column instruction label time series matrix is ​​generated, with each column representing the action instruction label, timestamp, and algorithm ethics risk classification, respectively.

[0068] The semantic conflict identification submodule calls the action instruction tags in the instruction tag time series matrix, performs logical direction judgment on the instruction meaning between tags according to the established semantic judgment rules, filters the tag combinations with semantic opposition, records the marking results in the form of key values, and generates a set of semantic conflict tag pairs.

[0069] The action instruction column in the instruction label time series matrix is ​​invoked, such as ['ApproveLoan', ..., 'DenyLoan']. Based on the semantic judgment rule table (as shown in Table 2) based on the algorithmic fairness criterion, a dual judgment is made on both logical and ethical dimensions. This rule table predefines 85 groups of instructions with logically contradictory or ethically conflicting relationships. For example, the rule table not only defines "ApproveLoan" and "DenyLoan" as semantically contradictory, but also defines that if "QueryUserCreditScore" is followed by "DenyLoan" without any other decision node directly following it in a single process, it constitutes a potential algorithmic opaque conflict. By comparing the instruction combinations in the sequence with the rule table, all label combinations with contradictions or conflicts are filtered out, and their numbers in the sequence are recorded as key-value pairs, such as {('ApproveLoan', 'DenyLoan'): (1, 4)}, generating a set of semantically conflicting label pairs.

[0070] Table 2: Examples of Semantic Judgment Rules for Algorithm Fairness

[0071]

[0072] As shown in Table 2, this form defines the pre-defined semantic logic and algorithmic ethical conflict relationships between some instructions.

[0073] The cross-execution judgment submodule extracts the execution start and end times of the corresponding instruction label time series matrix based on the semantic conflict label pairs that have been determined to be opposing in the set, calculates the proportion of the time interval intersection in the total path duration, and marks and extracts records with coverage exceeding the cross threshold standard to obtain the set of ethical action conflict events.

[0074] Process the set of semantically conflicting tag pairs, such as {('FilterResumeBySchool','RankCandidateBySkill'):(2,3)}. Extract the start and end times of the execution of conflicting instructions from the instruction tag time series matrix. For example, 'FilterResumeBySchool' is [1677628820.200,1677628820.800], and 'RankCandidateBySkill' is [1677628820.500,1677628821.100]. Calculate the percentage of intersection between the two time periods in the total path duration (0.9 seconds). In this example, the intersection duration is 0.3 seconds, and the percentage is... This percentage will be compared to a dynamic cross-threshold associated with the risk level, with a threshold benchmark value of [value missing]. However, the threshold will be adjusted according to the ethical risk classification of the instruction. The adjustment factor for high-risk instructions is 0.5, so the adjusted threshold is... ,because Exceeded Once the threshold is reached, the record is extracted, and all records that meet the conditions are aggregated to obtain a set of ethical action conflict events.

[0075] Please see Figure 4 The risk semantic aggregation module includes:

[0076] The expression content extraction submodule extracts corresponding text fragments, voice data and instruction description information based on event items in the set of ethical action conflict events. It unifies the format and reorganizes the order of multi-source content, encodes it in combination with execution time sequence tags, and performs standardized conversion by channel to generate a standardized corpus of expression content.

[0077] For ethical conflict events, such as those related to ('FilterResumeBySchool', 'RankCandidateBySkill'), relevant expressions can be extracted from multimodal interaction logs and application ethical risk feature libraries. Assuming that during the event period, the text configuration record of operator U003, "Prioritize screening resumes from 985 universities and sort by skill matching degree," a screenshot of the operation interface, and the text description of related instructions can be extracted. At the same time, relevant features such as "recruitment discrimination" can be retrieved from the risk feature library. The retrieved heterogeneous content is then formatted in a unified manner and reorganized chronologically. For example, the screenshot image is converted into text using OCR technology. After encoding, all information is integrated into a JSON object containing fields such as content, timestamp, source channel, and associated risk features, generating a standardized corpus of expressions.

[0078] The risk semantic recognition submodule calls the text and speech expression data in the standardized corpus of expression content, performs semantic unit decomposition and feature annotation on each piece of content according to the risk dictionary entries and intent identification rules, filters expression units with related guidance, semantic ambiguity or permission mismatch features, divides the structure according to the classification labels, and obtains the risk expression semantic classification results.

[0079] The system calls upon a standardized corpus of expressed content and analyzes it based on an AI ethical risk feature library constructed using robust machine learning. This feature library was obtained through adversarial training on 1,000 known ethical risk cases, effectively resisting noisy data. For input text data, such as "prioritize screening resumes from 985 universities...", the system decomposes it into semantic units and calculates the similarity between each unit and the feature vector in the risk feature library. For example, the similarity between the unit "prioritize screening resumes from 985 universities" and the feature "academic discrimination" is 0.92, exceeding the set threshold of 0.85. Therefore, this unit is labeled as "algorithmic fairness risk" and "exacerbates social inequality". All expression units labeled with any of the four types of ethical risks are classified and summarized to obtain the semantic classification results of risk expressions.

[0080] The high-frequency content aggregation submodule, based on the set of semantic tags marked in the semantic classification results of risk expressions, groups and counts the expression units that appear repeatedly, calculates the frequency value of each expression in the corpus, and performs summary and marking operations on the statement units whose frequency exceeds the set occurrence threshold to obtain a high-frequency aggregation list of ethical expressions.

[0081] The analysis of semantic classification results for risk expressions revealed recurring risk expression units. For example, after analyzing 5000 corpora, the expression unit "prioritize screening XX universities" appeared 150 times, of which 140 were labeled as "algorithm fairness risk," with a frequency of [missing information]. This frequency value will be compared to an adaptive occurrence threshold, which is based on a reference value of [value missing]. It also dynamically adjusts based on data distribution drift in AI application scenarios. For example, if the variance of education level distribution in resume data increases by 20%, the threshold will be adjusted accordingly. ,because If the adjusted threshold is exceeded, the statement unit "prioritize screening XX institutions" will be identified, and its correlation will be analyzed using a graph neural network (GNN) to give priority warnings. All statement units that exceed the limit and have high correlation will be aggregated to obtain a high-frequency aggregation list of ethical expressions.

[0082] Please see Figure 5 The modal output tracking module includes:

[0083] The time-segment signal extraction submodule collects the corresponding time location data based on the risk expression statements marked in the high-frequency aggregation list of ethical expressions, retrieves visual image frame sequences, audio action sequences and device response logs within the same time period, integrates the three types of signals according to timestamps and establishes corresponding index relationships, and generates a joint record set of modal signals in the same segment.

[0084] Based on risk expression statements in the high-frequency aggregation list of ethical expressions, such as "prioritize screening 985 universities", the system collects the timestamp of each occurrence of such statements in the logs, such as 1677628820.300. A 5-second time window ([1677628817.800, 1677628822.800]) is set around this timestamp. The width of this window is set according to the maximum traceability time for addressing ethical risks. Within this time window, the system retrieves and extracts all relevant records from visual monitoring logs, audio action sequences (such as keyboard typing acoustic features), and device response logs. These records include screenshots of the operation interface, keyboard typing rhythms, and logs of operations such as "FilterResumeBySchool". The system integrates these three types of heterogeneous signals by timestamp and establishes an index pointing to the central risk statement, generating a joint record set of modal signals in the same segment.

[0085] The multi-mode signal timing submodule calls the visual frames, audio actions, and device response signals in the joint record set of the same modal signal, establishes a sequence based on the start time field of each type of signal, compares the signal start order one by one according to the time difference standard, filters the records where the device response is earlier than the audio or image trigger, and obtains a list of abnormal signal start order.

[0086] The system analyzes visual, audio, and device response signals from the joint record set of signals in the same modality. A time sequence is established based on the start time field of each signal, and the start order is compared one by one. If the start timestamp of the device response signal is earlier than the timestamp of any attributable human interaction signal (such as keyboard keystroke or mouse click) by more than a preset threshold (200 milliseconds in this example), it is considered an abnormal sequence. This threshold is determined based on statistics (mean plus three standard deviations) of response delays in 1000 normal human-computer interactions. For example, if the device response time 1677628819.900 is 250 milliseconds earlier than the earliest user keyboard keystroke signal 1677628820.150, exceeding the threshold, then this record is filtered out. After summarizing all such records, a list of abnormal signal start order is obtained.

[0087] The response mismatch identification submodule detects the correspondence between the device response signal and the visual frame sequence action based on the records in the signal start sequence anomaly list, removes items with image matching actions, and extracts records that simultaneously meet the conditions of early response start and no corresponding image action to obtain the ethical output response anomaly trajectory set.

[0088] Each record in the signal initiation sequence anomaly list is examined to check whether there is a predefined correspondence between the device response signal and the visual image frame sequence action. The correspondence is stored in a matching rule base generated by analyzing historical operation logs. For example, the rule base defines the interface action of "clicking the 'Filter' button" as corresponding to the response "FilterResumeBySchool". In the anomaly record of this example, the device response is "FilterResumeBySchool", but the visual frame sequence only shows "mouse pointer hovering over the input box". No direct mapping is found in the rule base, so the record is retained. Through this process, all records that simultaneously meet the conditions of "device response started early" and "no clear interface trigger action" are extracted to obtain the ethical output response anomaly trajectory set.

[0089] Please see Figure 6 The monitoring and governance classification module includes:

[0090] The response offset extraction submodule, based on the identified response numbers, expression content and behavioral instruction records in the ethical output response anomaly trajectory set, retrieves and matches the response behaviors corresponding to the same instruction by number, identifies expression instances with different response performances, and extracts record units with behavioral state offset features by number as index, generating a response diversity offset instance set;

[0091] Based on the abnormal trajectory set of ethical output responses, records are grouped according to behavioral instructions (such as "RankCandidateBySkill"). Within each group, the system retrieves and matches multiple response behaviors with the same instruction to identify performance differences. For example, for 100 resumes with the same input, if the response results of two "RankCandidateBySkill" instructions show a statistically significant difference in the gender distribution of candidates, with the difference rate exceeding the preset algorithm fairness risk threshold of 5%, then a behavioral state shift is identified between these two responses. Using the response number as an index, the corresponding record unit is extracted, including the expression content, instruction, timestamp, and complete device response log. After all such instances are aggregated, a response diversity shift instance set is generated.

[0092] The governance fragment labeling submodule calls the expression data in the response diversity offset instance set, performs structural parsing based on the role call order and permission expression field, screens content items with subject-object configuration conflicts and permission boundary offsets, assigns the expression instance to be governed with a label, binds the number to establish a correspondence with the semantic fragment, and obtains the expression fragment identifier set to be governed.

[0093] The system retrieves the expression data from the response diversity offset instance set and performs structured parsing based on the governance ethics risk rule base. The parsing process focuses on subject-object configuration conflicts and permission boundary offsets. For example, if a gender difference offset is found in a response, it can be traced back to the algorithm assigning certain implicit features a weight limit that exceeds the maximum limit set by the governance criteria (e.g., the maximum is set to 0.3, but it is actually 0.8). This is then determined to be a permission boundary offset, and the instance will be labeled "to be governed". Its event number is then bound to the specific algorithm parameters that led to the judgment (e.g., "implicit feature weight") to establish a correspondence. All instances that meet the conditions and their corresponding relationships are collected to obtain the set of expression fragment identifiers to be governed.

[0094] The content structure classification submodule, based on the semantic classification information and expression fields recorded in the identifier set of the expression fragments to be governed, encodes each expression instance according to behavioral tags and semantic patterns, constructs a unified field mapping structure, and archives all marked expressions after merging and reconstructing them according to field groups to obtain the ethical risk monitoring and governance result set;

[0095] Based on the information in the set of expression fragment identifiers to be governed, each instance is encoded in multiple dimensions according to its behavioral tags, semantic patterns, and the core AI ethical risk category it belongs to, such as {Category:'Algorithm',Action:'Rank',Mode:'Bias.Gender'}. The system constructs a unified mapping structure containing fields such as "event number", "user identity", and "risk classification" (as shown in Table 3). All labeled expression instances are merged, reconstructed, and archived. At the same time, the result automatically triggers a verification and feedback mechanism, which compares the results with those of manual review. Confirmed cases will be used as new data to adaptively update the "AI ethical risk feature library" and the identification model, forming a closed loop that continuously improves robustness, and finally obtaining the ethical risk monitoring and governance result set.

[0096] Table 3: Classification Structure of Ethical Risk Events

[0097]

[0098] As shown in Table 3, this table presents the structured classification results of the final output, which standardizes and archives the identified ethical risk events and their core elements, and clarifies the core risk categories to which they belong.

[0099] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A robust artificial intelligence-based AI ethical risk monitoring and governance system, characterized in that, The system includes: Call the difference recognition module to extract user identifiers, functional units and jump records, calculate the time difference of frequent jump behaviors between functions, filter records with continuous differences less than the set lower limit, and integrate their interface tags to generate an ethical jump abnormal path set; The set of abnormal ethical jump paths includes abnormal jump label combinations, continuous low difference time period markers, and jump offset structures between functional units. The conflict behavior localization module calls the set of abnormal ethical jump paths, searches for combinations of action tags with opposite semantics, filters conflicting operations with overlapping execution times under the same path, extracts abnormal execution lines according to time coverage, and generates a set of ethical action conflict events. The set of ethical action conflict events includes semantic conflict instruction groups, cross-execution time coverage segments, and conflict event semantic tags; The risk semantic aggregation module calls up expression fragments in the set of ethical action conflict events, identifies leading words and vague liability terms, clusters semantically similar expressions and counts their occurrences, filters expressions with repetitions exceeding a threshold, marks their semantic risk level, and generates a high-frequency aggregation list of ethical expressions. The high-frequency aggregation list of ethical expressions includes high-frequency sensitive expressions, guiding and ambiguous semantic categories, and permission mismatch intent markers; The modal output tracking module calls the time period in the high-frequency aggregation list of ethical expressions, matches image frames, voice and device response, compares the time difference of the signal start order, filters records whose response is earlier than the input signal and lacks image matching, and generates an abnormal trajectory set of ethical output response. The set of abnormal trajectory for ethical output response includes abnormal response timing segments, missing image action records, and premature device response paths. The monitoring and governance classification module calls the set of abnormal ethical output responses, sorts out the differences between the expressed content and the response number, archives and numbers the abnormal response groups for risk, and constructs a governance data master table after labeling and classifying them to generate a set of ethical risk monitoring and governance results. The ethical risk monitoring and governance result set includes response diversity offset samples, expression classification numbers to be governed, and ethical risk content governance codes. The monitoring and governance classification module includes: The response offset extraction submodule, based on the identified response numbers, expression content and behavioral instruction records in the ethical output response anomaly trajectory set, retrieves and matches the response behaviors corresponding to the same instruction by number, identifies expression instances with different response performances, and extracts record units with behavioral state offset features by number as index, generating a response diversity offset instance set. The governance fragment labeling submodule calls the expression data in the response diversity offset instance set, performs structural parsing based on the role call order and permission expression field, screens content items with subject-object configuration conflicts and permission boundary offsets, assigns a tag to be governed to the expression instances that meet the conditions, and binds the number to establish a correspondence with the semantic fragment to obtain the expression fragment identifier set to be governed. The content structure classification submodule, based on the semantic classification information and expression fields recorded in the set of expression fragment identifiers to be governed, encodes each expression instance according to behavioral tags and semantic patterns, constructs a unified field mapping structure, and archives all marked expressions after merging and reconstructing them according to field groups to obtain the ethical risk monitoring and governance result set.

2. The robust artificial intelligence-based AI ethical risk monitoring and governance system according to claim 1, wherein, The invocation of the difference recognition module includes: The identity information extraction submodule extracts the user's identity number, functional unit identifier, and jump path label based on the user's operation log in the system. It also constructs a corresponding dataset by combining the timestamp information of each record. By sorting the set by timestamp, it obtains the sequential structure of functional unit calls and generates a functional call time series dataset. The jump behavior sequence construction submodule calls the function to call the continuous time points in the time series dataset and their corresponding unit labels, establishes a path mapping relationship for adjacent unit labels, calculates their corresponding time difference, constitutes the mapping entries between jump paths and time intervals, and forms a jump path time difference sequence table. The time difference judgment submodule sets a minimum threshold standard for jump response based on the difference information recorded in the jump path time difference sequence table. It filters and merges the tags of path records that are continuously lower than the time difference lower limit, extracts the corresponding paths to form a set, and performs deduplication and sequence structure recombination on the path tags to obtain the set of ethical jump abnormal paths. 3.The robust artificial intelligence based AI ethical risk monitoring and governance system according to claim 1, wherein, The conflict behavior localization module includes: The tag sequence parsing submodule extracts the action instructions and their timestamp information corresponding to each node in the path based on the path tags in the set of abnormal ethical jump paths, constructs the correspondence structure between action tags and time fields, establishes the number mapping and time sequence arrangement relationship according to the path order, and generates an instruction tag time sequence matrix. The semantic conflict identification submodule calls the action instruction tags in the instruction tag time series matrix, performs logical direction judgment on the instruction meaning between tags according to the established semantic judgment rules, filters the tag combinations with semantic opposition, records the marking results in the form of key values, and generates a set of semantic conflict tag pairs. The cross-execution judgment submodule extracts the execution start and end times of the corresponding instruction label time series matrix based on the semantic conflict label pair set that has been determined to be opposing, calculates the proportion of the time period intersection in the total path duration, and marks and extracts records with coverage exceeding the cross threshold standard to obtain the set of ethical action conflict events.

4. The robust artificial intelligence-based AI ethical risk monitoring and governance system according to claim 1, wherein, The risk semantic aggregation module includes: The expression content extraction submodule extracts corresponding text fragments, voice data and instruction description information based on the event items in the set of ethical action conflict events, unifies the format and reorganizes the order of the multi-source content, encodes it in combination with execution time tags, and performs standardized conversion according to the channel to generate a standardized corpus of expression content. The risk semantic recognition submodule calls the text and speech expression data in the standardized corpus of the expression content, performs semantic unit decomposition and feature annotation on each content according to the risk dictionary entries and intent identification rules, filters expression units with related guidance, semantic ambiguity or permission mismatch features, divides the structure according to the classification labels, and obtains the risk expression semantic classification results. The high-frequency content aggregation submodule, based on the set of semantic tags marked in the risk expression semantic classification results, groups and counts the expression units that appear repeatedly, calculates the frequency value of each expression in the corpus, and performs summary and marking operations on the statement units whose frequency exceeds the set occurrence threshold to obtain a high-frequency aggregation list of ethical expressions. 5.The robust artificial intelligence based AI ethical risk monitoring and governance system according to claim 1, wherein, The modal output tracking module includes: The time-segment signal extraction submodule collects the corresponding time location data based on the risk expression statements marked in the high-frequency aggregation list of ethical expressions, retrieves visual image frame sequences, audio action sequences and device response logs within the same time period, integrates the three types of signals according to timestamps and establishes corresponding index relationships, and generates a joint record set of modal signals in the same segment. The multi-mode signal timing submodule calls the visual frames, audio actions, and device response signals in the joint record set of the same modal signal, establishes a sequence based on the start time field of each type of signal, compares the signal start order one by one according to the time difference standard, filters the records where the device response is earlier than the audio or image trigger, and obtains a list of abnormal signal start order. The response mismatch identification submodule detects the correspondence between the device response signal and the visual frame sequence action based on the records in the signal start-up sequence anomaly list, removes items with image matching actions, and extracts records that simultaneously meet the conditions of early response start and no corresponding image action to obtain the ethical output response anomaly trajectory set.

Images