AI ethical risk monitoring and treatment system based on robustness artificial intelligence

Abstract

The invention relates to the technical field of AI ethical risk management, in particular to an AI ethical risk monitoring and management system based on robustness artificial intelligence, which comprises the steps of identifying abnormal jump behaviors, extracting semantic conflict instructions and cross execution events, aggregating high-frequency risk statements and tracking response dislocation paths based on user identities and function calling records. And classifying the offset response content to form an ethical risk management result set. The method comprises the following steps: comparing function call jump continuity with a time difference value, extracting a low-difference abnormal path, combining semantic opposition and time cross to identify behavior conflicts, classification guidance, ambiguity and permission mismatch statements to form a high-frequency risk set, fusing a multi-modal signal starting sequence to extract a response dislocation path, constructing a collaborative check chain, and performing collaborative check. And the breadth and level of risk identification are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of AI ethical risk governance, and particularly relates to an AI ethical risk monitoring and governance system based on robust artificial intelligence. BACKGROUND

[0002] The technical field of AI ethical risk governance involves identifying and managing the ethical risks that may arise from the practical application of artificial intelligence systems. With the widespread application of artificial intelligence technology, how to effectively address the ethical issues that may arise in 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 risks: data ethical risk, algorithm ethical risk, application ethical risk and governance ethical risk. The goal is to ensure the fairness, transparency and security of artificial intelligence systems, avoid unfair algorithmic decisions, data breaches, privacy violations and other problems, and ensure that the application of AI technology complies with legal, ethical and social responsibility requirements.

[0003] Among them, the AI ethical risk monitoring and governance system based on robust artificial intelligence refers to a system that uses robust AI technology to comprehensively identify and effectively manage AI ethical risks. By thoroughly analyzing the core categories of AI ethical risks, it covers the four main areas of data ethics, algorithm ethics, application ethics and governance ethics, and further breaks down each risk into specific risk points. Through robust technology, the system can handle data changes in real time and deal with dynamic problems such as algorithm updates and data distribution drift, effectively resisting noise interference and malicious attacks. The system learns from historical risk cases and combines real-time data to build a dynamically updated risk feature library, 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, providing a scientific basis for risk prevention and control.

[0004] The prior art lacks comparative analysis of jump rhythm and time difference in the functional module operation link, making it difficult to distinguish continuous abnormal paths, lacking conflict event screening capabilities based on time cross relationships in multi-instruction overlapping scenarios, and not forming a systematic classification and frequency aggregation of ambiguous and permission mismatch characteristics of risk content, making it difficult to layer the risk information, and lacking means to check the starting order of device responses and multi-modal signals, making it difficult to trace the source when the response path and behavior action are disconnected, limiting the completeness and timeliness of risk management. SUMMARY

[0005] The purpose of the present application is to solve the shortcomings in the prior art and to propose an AI ethical risk monitoring and governance system based on robust artificial intelligence.

[0006] In order to achieve the above object, the application adopts the following technical scheme: An AI ethics risk monitoring and management system based on robust artificial intelligence, comprising:

[0007] A call difference identification module is called to extract user identification, function modules and jump records, perform time difference calculation on frequent jump behavior between functions, filter records with continuous difference less than the set lower limit, integrate interface tags to generate an ethics jump abnormal path set;

[0008] A conflict behavior positioning module is called to call the ethics jump abnormal path set, find action tag combinations with opposite semantics, filter conflict operations with overlapping execution time in the same path, extract abnormal execution lines according to time coverage, and generate an ethics action conflict event set;

[0009] A risk semantic aggregation module is called to call expression fragments in the ethics action conflict event set, identify inductive words and ambiguous responsibility language, cluster expressions with similar semantics and count the number of occurrences, filter expressions with a repetition number exceeding a threshold, mark the semantic risk level, and generate an ethics expression high-frequency aggregation list;

[0010] A modal output tracking module is called to call the time period in the ethics expression high-frequency aggregation list, match image frames, voice and device responses, perform time difference comparison on signal start order, filter records with responses earlier than input signals and lack of image matching, and generate an ethics output response abnormal trajectory set.

[0011] As a further scheme of the application, the ethics jump abnormal path set includes abnormal jump tag combinations, continuous low difference time period markers, and jump offset structures between function units, the ethics action conflict event set includes semantic conflict instruction groups, cross execution time coverage fragments, and conflict event semantic tags, the ethics expression high-frequency aggregation list includes high-frequency sensitive expression sentences, guiding and ambiguous semantic categories, and permission mismatch type intent markers, and the ethics output response abnormal trajectory set includes response time sequence abnormal fragments, image action missing records, and device response early trigger paths.

[0012] As a further scheme of the application, the call difference identification module comprises:

[0013] An identity information extraction sub-module extracts identity numbers, function module identifiers and jump path tags based on user operation logs in the system, constructs corresponding data sets in combination with time stamp information of each record, sorts the sets by time stamp to obtain the order structure of function unit calls, and generates a function call time sequence data set;

[0014] The jump behavior sequence construction submodule constructs a mapping entry between a jump path and a time interval by calling a continuous time point in the function call time sequence dataset and a corresponding unit label, establishing a path mapping relationship between adjacent unit labels, and calculating a corresponding time difference value, and forms a jump path time difference value sequence list;

[0015] The time difference value judgment submodule sets a minimum threshold standard for a jump response according to the difference value information recorded in the jump path time difference value sequence list, filters and labels the path records successively below the lower limit of the time difference value, extracts a corresponding path to form a set, and removes duplicates and reorganizes the sequence structure of the path labels to obtain an ethical jump abnormal path set.

[0016] As a further scheme of the present application, the conflict behavior positioning module comprises:

[0017] The label sequence analysis submodule extracts action instructions and timestamp information corresponding to each node in the path based on the path labels in the ethical jump abnormal path set, constructs a corresponding structure of action labels and time fields, establishes a numbering mapping and time sequence arrangement relationship according to the path order, and generates an instruction label time sequence matrix;

[0018] The semantic conflict identification submodule calls the action instruction labels in the instruction label time sequence matrix, judges the logical direction of the instruction meaning between the labels according to the predetermined semantic judgment rules, filters the label combinations with semantic opposite relationship, records the marking results in the form of key-value, and generates a semantic conflict label pair set;

[0019] The cross-execution judgment submodule extracts the execution start and end time of the label combination determined to be opposite in the semantic conflict label pair set in the instruction label time sequence matrix, calculates the time period intersection ratio in the total path duration, and extracts the records marked with a coverage rate exceeding a cross threshold standard to obtain an ethical action conflict event set.

[0020] As a further scheme of the present application, the risk semantic aggregation module comprises:

[0021] The expression content extraction submodule extracts corresponding text segments, voice data and instruction description information based on the event items in the ethical action conflict event set, uniformly formats and sequentially reorganizes the multi-source content, encodes the execution time sequence labels, and standardizes the conversion according to the channels to generate an expression content standardized corpus set;

[0022] The risk semantic recognition submodule calls the text and voice expression data in the expression content normalized corpus, performs semantic unit decomposition and feature labeling on each content according to the risk dictionary entries and intention identification rules, screens expression units with features of associated guidance, semantic ambiguity or permission mismatch, and obtains a risk expression semantic classification result according to a classification label division structure.

[0023] The high-frequency content aggregation submodule groups and counts expression units that repeatedly appear in the semantic label set in the risk expression semantic classification result, calculates the frequency values of each expression content in the corpus, and performs aggregation and identification operations on sentence units with frequency values exceeding a set appearance threshold, to obtain an ethical expression high-frequency aggregation list.

[0024] As a further scheme of the present application, the modal output tracking module comprises:

[0025] The time period signal extraction submodule collects the corresponding time positioning data of the marked risk expression sentences in the ethical expression high-frequency aggregation list, retrieves visual image frame sequences, audio action sequences and device response logs in the same time period, integrates and establishes a corresponding index relationship among the three types of signals according to timestamps, and generates a same-period modal signal joint record set;

[0026] The multi-modal signal time alignment submodule calls the visual frames, audio actions and device response signals in the same-period modal signal joint record set, establishes a sequential sequence according to the start time field of each type of signal, compares the start order of signals according to the time difference value standard, screens records in which the device response is earlier than the audio or image trigger, and obtains an abnormal signal start order list;

[0027] The response mismatch recognition submodule detects the corresponding relationship between the device response signal and the visual frame sequence action according to the records in the abnormal signal start order list, eliminates items with image matching actions, extracts records that simultaneously satisfy the conditions of early start of response and absence of image corresponding actions, and obtains an ethical output response abnormal trajectory set.

[0028] As a further scheme of the present application, the system further comprises:

[0029] The monitoring and management classification module calls the ethical output response abnormal trajectory set, analyzes the differences between expression content and response numbers, performs risk archiving and numbering processing on abnormal response groups, labels and classifies to construct a management data summary table, and generates an ethical risk monitoring and management result set;

[0030] The ethical risk monitoring and management result set comprises response diversity offset samples, expression classification numbers to be managed, and ethical risk content management codes.

[0031] As a further scheme of the present application, the monitoring governance classification module comprises:

[0032] The response offset extraction submodule extracts, based on the identified response number, expression content, and behavior instruction record in the ethical output response abnormal trajectory set, the response behaviors corresponding to the same instruction according to the number, identifies expression instances with response performance differences, and extracts record units with behavior state offset characteristics as indexes to generate a response diversity offset instance set;

[0033] The governance segment labeling submodule calls the expression data in the response diversity offset instance set, performs structural analysis according to the role calling sequence and permission expression field, screens content items with subject-object configuration conflicts and permission boundary offset, assigns a governance-to-be-labeled label to expression instances meeting the conditions, and binds the number and semantic segments to establish a corresponding relationship to obtain a governance-to-be-labeled expression segment identification set;

[0034] The content structure classification submodule encodes each expression instance according to the behavior label and semantic mode based on the semantic classification information and expression field structure recorded in the governance-to-be-labeled expression segment identification set, constructs a unified field mapping structure, refiles all labeled expressions after grouping and reconstruction according to the field group, and obtains an ethical risk monitoring governance result set.

[0035] Compared with the prior art, the present application has the following advantages and positive effects:

[0036] In the present application, the low-time-difference abnormal path set is formed by comparing the jump coherence and time difference value sequence between function calls, the potential behavior conflict events are extracted by combining the semantic opposite instruction and the cross-coverage proportion of execution time, the semantic classification and repeated statistics are carried out by using the guiding, ambiguity, and permission mismatch characteristics, the high-frequency risk content set is formed, the response misplacement path is extracted by fusing the start sequence difference of visual frames, audio actions, and device response signals, and finally the collaborative review chain between multi-modal signals is constructed to improve the comprehensiveness, dynamics, and hierarchical management capability of risk identification. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The system flowchart of the present application;

[0038] Figure 2 The acquisition flowchart of the calling difference identification module of the present application;

[0039] Figure 3 The acquisition flowchart of the conflict behavior positioning module of the present application;

[0040] Figure 4 The acquisition flowchart of the risk semantic aggregation module of the present application;

[0041] Figure 5 Acquisition flowchart of the modal output tracking module of the present application;

[0042] Figure 6 Acquisition flowchart of the monitoring and management classification module of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the present application will be described below with reference to the drawings.

[0044] In the embodiments of the present application, the words such as "example", "for example" are used to represent an example, illustration or description. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the word "example" is intended to present the concept in a specific manner. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0045] In the embodiments of the present application, "image" and "picture" can be used interchangeably at times. It should be pointed out that when the distinction is not emphasized, the meanings expressed are consistent. "Of", "corresponding" and "corresponding" can be used interchangeably at times. It should be pointed out that when the distinction is not emphasized, the meanings expressed are consistent.

[0046] In the embodiments of the present application, sometimes the subscript such as W1 can be written in the form of non-subscript such as W1. When the distinction is not emphasized, the meanings expressed are consistent.

[0047] To make the technical problems, technical solutions and advantages to be solved by the present application clearer, specific embodiments will be described in detail below with reference to the drawings.

[0048] Please refer to Figure 1 The present application provides a technical solution: an AI ethical risk monitoring and management system based on robust artificial intelligence, which comprises:

[0049] The difference identification module is called to retrieve the identity number of the user in the system, the function unit used and the jump record, to perform a coherence check on the rapid jump behavior between different function units, analyze the time difference value sequence in continuous calling, mark the operation record as abnormal when the difference is continuously lower than the lower limit of the standard, and summarize the calling labels to form a jump deviation set, to generate an ethical jump abnormal path set;

[0050] The conflict behavior positioning module, based on the label sequence collected in the ethical jump abnormal path set, finds out whether there is a semantic conflict instruction combination between actions in the path, analyzes whether the instruction group with semantic conflict exists cross-execution in the time dimension, extracts the conflict event when the cross-execution period coverage exceeds the limited range, and generates an ethical action conflict event set;

[0051] The risk semantic aggregation module extracts the text segments, voice content and instruction descriptions involved from the ethical action conflict event set, identifies sensitive words and divides risk intentions, classifies and counts expressions with guiding, ambiguous or permission mismatch characteristics according to semantics, aggregates and marks risk content that appears more than a set threshold, and generates an ethical expression high-frequency aggregation list;

[0052] The modal output tracking module calls the time positioning of the risk expression in the ethical expression high-frequency aggregation list, retrieves the visual frames, audio actions and device responses in the same segment, analyzes the starting order of the three types of signals, screens the device responses that are earlier than the behavior instructions and lack of image corresponding action records, extracts them as response mismatch paths, and generates an ethical output response abnormal trajectory set;

[0053] The monitoring and governance classification module, combined with the response number, expression content and instruction execution in the ethical output response abnormal trajectory set, combs out expression instances with response diversity deviation, marks them as governance segments, and classifies and codes them according to content characteristics, and aggregates them into an ethical risk monitoring and governance result set that can be used for review and governance.

[0054] The ethical jump abnormal path set includes abnormal jump label combination, continuous low difference time period mark, function unit jump deviation structure, the ethical action conflict event set includes semantic conflict instruction group, cross-execution time coverage segment, conflict event semantic label, the ethical expression high-frequency aggregation list includes high-frequency sensitive expression sentence, guiding and ambiguous semantic category, permission mismatch type intention mark, the ethical output response abnormal trajectory set includes response time sequence abnormal segment, image action missing record, device response early trigger path, and the ethical risk monitoring and governance result set includes response diversity deviation sample, governance expression classification number, and ethical risk content governance code.

[0055] Please refer to Figure 2 , the difference identification module includes:

[0056] The identity information extraction submodule extracts the identity number, function unit identifier and jump path label based on the operation log of the user in the system, and constructs the corresponding data set combined with the timestamp information of each record, obtains the function call order structure by sorting the set by timestamp, and generates a function call time sequence data set;

[0057] The processing system records all user operation logs (as shown in Table 1) in the background server, extracts the identity number, functional unit identifier, jump path label, and data source informed consent status label, and each operation record is combined with its timestamp information to be converted into a structured five-element data group, for example, the log record "timestamp: 1677628801, user ID: U001, operation unit: M05, path: / user / login, consent status: Agreed" is converted into (U001, M05, / user / login, Agreed) and timestamp 1677628801. This process synchronously performs preliminary verification of data collection compliance. If the consent status label is not 'Agreed', the record will be marked as a potential data ethics risk. All converted five-element data groups are sorted in ascending order of timestamp to form a functional unit call sequence that accurately reproduces the user behavior track, for example, the call sequence of user U001 is recorded as [(1677628801, M05, Agreed), (1677628802, M12, Agreed), (1677628804, M09, Agreed)]. Integrating all user sequences, a functional call time series dataset is finally generated.

[0058] Table 1: User operation log data example

[0059] As shown in Table 1, the table shows the key fields extracted from the system operation log and their example values, providing raw data input for subsequent behavior sequence and data ethics risk analysis.

[0060] The jump behavior sequence construction submodule calls the continuous time points in the functional call time series dataset and their corresponding unit labels, establishes a path mapping relationship between adjacent unit labels, and calculates the corresponding time difference value to form a mapping entry between the jump path and the time interval, forming a jump path time difference value sequence table;

[0061] Using the generated functional call time series dataset, the jump path is constructed by traversing the continuous elements in the sequence. For each group of consecutive time points, such as (1677628801, M05, Agreed) and (1677628802, M12, Agreed), a path mapping relationship M05->M12 is established, and the timestamp difference is calculated, i.e. seconds, and the agreement status at both ends of the path is also recorded. The process pays special attention to the change in the agreement status. If there is a jump from 'Agreed' to 'Expired', the entry will be marked as a potential ethical risk of data sharing. For example, the jump of user U002 (M05, Agreed) -> (M15, Expired) generates a mapping entry (M05->M15, 1s, Agreed->Expired, RiskFlag=1). After processing all user sequences, all generated mapping entries are collected to form a jump path time difference value sequence list.

[0062] A time difference value judgment submodule sets a minimum threshold value for jump response based on the difference value information recorded in the jump path time difference value sequence list. The path records that are continuously below the lower limit of the time difference value are screened and labeled, the corresponding path set is extracted, the path labels are de-duplicated and the sequence structure is reorganized, and an ethical jump abnormal path set is obtained.

[0063] To determine the jump response speed, a robust minimum threshold value is set. The threshold value (450 milliseconds in this example) is determined based on 30000 response time data of the system under different loads, and a time feature library containing 50 known attack patterns is introduced for counter-training. The threshold value is determined by robust anomaly detection models such as Isolation Forest, which can effectively distinguish between normal user operations and automated abnormal behaviors. The system screens records in the jump path time difference value sequence list that have a time interval of less than the threshold value for more than three times in a row, for example, (M01->M02, 410ms), (M02->M07, 430ms), and (M07->M11, 425ms). The screened records are assigned a common merging label, and their corresponding paths are de-duplicated and reorganized to form a complete continuous abnormal jump path, such as M01->M02->M07->M11. All such paths ultimately form an ethical jump abnormal path set.

[0064] Please refer to Figure 3 The conflict behavior positioning module includes:

[0065] A label sequence analysis submodule extracts the action instructions and their timestamp information corresponding to each node in the path based on the path labels in the ethical jump abnormal path set, constructs a corresponding structure of action labels and time fields, establishes a numbering mapping and time sequence arrangement relationship according to the path order, and generates an instruction label time sequence matrix.

[0066] The specific paths in the ethical jump abnormal path set, such as M01->M02->M07->M11, are analyzed. From the system instruction log and the algorithm fairness rule library, the corresponding action instructions, timestamps and algorithm ethical risk classifications of each node in the path within the time period are extracted, for example, M01 corresponds to the instruction "ApproveLoan", the timestamp is 1677628810.100, and the risk classification is "high fairness risk"; M11 corresponds to the instruction "DenyLoan", the timestamp is 1677628811.495, and the risk classification is "high fairness risk". The above information is structured 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-by-3-column instruction label time sequence matrix is generated, and each column is an action instruction label, a timestamp and an algorithm ethical risk classification.

[0067] The semantic conflict identification submodule calls the action instruction labels in the instruction label time sequence matrix, makes logical direction judgments on the instruction meanings between the labels according to the established semantic judgment rules, filters the label combinations with semantic opposite relationships, and records the marking results in the form of key-value pairs to generate a semantic conflict label pair set;

[0068] The action instruction column in the instruction label time sequence matrix, such as [‘ApproveLoan’, …, ‘DenyLoan’], is called, and a semantic judgment rule table based on the algorithm fairness criteria (as shown in Table 2) is used for logical and ethical dimension double judgment. The rule table predefines 85 groups of instructions with logical opposite or ethical conflict relationships, for example, the rule table not only defines that “ApproveLoan” and “DenyLoan” are semantically opposite, but also defines that if “QueryUserCreditScore” is not followed by “DenyLoan” directly without other decision nodes in a single process, it constitutes a potential algorithm opacity conflict. By comparing the instruction combinations in the sequence with the rule table, all label combinations with opposite or conflicting relationships are filtered, and their numbers in the sequence are recorded in the form of key-value pairs, such as {(‘ApproveLoan’, ‘DenyLoan’):(1,4)}, to generate a semantic conflict label pair set.

[0069] Table 2: Algorithm fairness semantic judgment rule example

[0070] As shown in Table 2, the table defines the pre-set semantic logic and algorithm ethical conflict relationships between some instructions.

[0071] The cross-execution judgment submodule extracts the execution start and end time of the label combination determined as opposite in the semantic conflict label pair set according to the semantic conflict label pair set, calculates the time period intersection ratio in the total path duration, and extracts the record label extraction with the coverage exceeding the cross threshold standard, to obtain the ethical action conflict event set;

[0072] The processing semantic conflict label pair set, such as {(‘FilterResumeBySchool’,‘RankCandidateBySkill’):(2,3)}, extracts the execution start and end time of the conflict instruction from the instruction label time series matrix, such as ‘FilterResumeBySchool’ for [1677628820.200,1677628820.800] and ‘RankCandidateBySkill’ for [1677628820.500,1677628821.100], calculates the intersection ratio of the two time periods in the total path duration (0.9 seconds), and the intersection time in this example is 0.3 seconds, accounting for The ratio will be compared with a dynamic cross threshold value associated with the risk level, and the threshold reference value is But it will be adjusted according to the ethical risk classification of the instruction, and the adjustment coefficient of high-risk instructions is 0.5, so the adjusted threshold is Since exceeds the threshold value, the record is extracted, and after all the records that meet the conditions are summarized, the ethical action conflict event set is obtained.

[0073] Please refer to Figure 4 The risk semantic aggregation module includes:

[0074] The expression content extraction submodule extracts the corresponding text segment, voice data and instruction description information based on the event items in the ethical action conflict event set, formats and reorganizes the multi-source content, encodes the execution time sequence label, and standardizes the conversion according to the channel, to generate the expression content standardized corpus set;

[0075] For ethical action conflict events such as the event associated with ('FilterResumeBySchool', 'RankCandidateBySkill'), relevant expression content can be extracted from the multi-modal interaction log and the application ethical risk feature library. Assuming that during the event occurrence period, the text configuration record of operator U003 is extracted, "Prioritize 985 university resumes and sort by skill matching degree", the screenshot image of the operation interface, and the text description of the related instructions, at the same time, relevant features such as "recruitment discrimination" are retrieved from the risk feature library, so that the retrieved heterogeneous content is unified in format and reorganized in chronological order. For example, the screenshot image is converted into text through OCR technology, and after encoding processing, all information is integrated into a JSON object containing content, timestamp, source channel, and associated risk features, generating an expression content standardized corpus set.

[0076] The risk semantic recognition submodule calls the text and voice expression data in the expression content standardized corpus set, performs semantic unit decomposition and feature labeling on each piece of content according to risk dictionary entries and intent identification rules, filters expression units with associated guidance, semantic ambiguity, or permission mismatch features, and obtains risk expression semantic classification results according to the classification label division structure.

[0077] The expression content standardized corpus set is called and analyzed according to the AI ethical risk feature library constructed based on robust machine learning. This feature library is obtained through adversarial training on 1000 known ethical risk cases and can effectively resist noise data. For input text data such as "Prioritize 985 university resumes…", the system will decompose it into semantic units and calculate the similarity between each unit and the feature vector in the risk feature library. For example, the similarity between the unit "Prioritize 985 university resumes" and the "educational discrimination" feature is 0.92, which exceeds the set threshold of 0.85. Therefore, this unit is labeled as "algorithm fairness risk" and "exacerbate social injustice". After classification and summarization, the risk expression semantic classification results are obtained.

[0078] The high-frequency content aggregation submodule groups and counts the expression units that repeatedly appear in the risk expression semantic classification results according to the semantic label set marked in the results, calculates the frequency of each expression content in the corpus set, and performs summarization and identification operations on the sentence units whose frequency exceeds the set appearance threshold, obtaining an ethical expression high-frequency aggregation list.

[0079] The risk expression semantic classification results are counted for repeated risk expression units. For example, after analyzing 5000 pieces of corpus, it is found that the expression unit "Prioritize XX university" appears 150 times, of which 140 times are labeled as "algorithm fairness risk", with a frequency of The frequency value will be compared with an adaptive occurrence threshold, which is based on and dynamically adjusted according to the data distribution drift of the AI application scenario, for example, if the variance of the education background distribution in the resume data increases by 20%, the threshold will be adjusted to Since exceeds the adjusted threshold, the statement unit "Priority Filter XX University" will be identified, and the relevance will be analyzed using a graph neural network (GNN), and a priority warning will be given. All statement units with frequency exceeding the threshold and high relevance are aggregated to obtain a high-frequency aggregation list of ethical expressions.

[0080] Please refer to Figure 5 , the modal output tracking module includes:

[0081] The time period signal extraction submodule collects the corresponding time positioning data of the marked risk expression statements in the high-frequency aggregation list of ethical expressions, retrieves the visual image frame sequence, audio action sequence and device response log within the same time period, integrates the three types of signals according to the timestamp and establishes a corresponding index relationship, and generates a joint record set of the same segment modal signals.

[0082] According to the risk expression statements in the high-frequency aggregation list of ethical expressions, such as "Priority Filter 985 University", collect the timestamp of each occurrence in the log, such as 1677628820.300. Set a 5-second time window ([1677628817.800, 1677628822.800]) centered on this timestamp. The width is set according to the maximum traceability period of governance ethical risk. Within this time window, the system retrieves and extracts all related records from the visual monitoring log, audio action sequence (such as keyboard click acoustic features) and device response log, including operation interface screenshots, keyboard click rhythm and "FilterResumeBySchool" operation logs. Integrate the three types of heterogeneous signals according to the timestamp, and establish an index pointing to the central risk statement to generate a joint record set of the same segment modal signals.

[0083] The multi-modal signal alignment submodule calls the visual frames, audio actions and device response signals in the same segment modal signal joint record set, establishes a sequential sequence according to the start time field of each type of signal, compares the signal start order according to the time difference value standard, filters the records where the device response is earlier than the audio or image trigger, and obtains an abnormal signal start order list.

[0084] The visual, audio, and device response signals in the same segment modality signal joint record set are analyzed, a time sequence is established according to the start time field of each signal, and the start order is compared piece by piece. If the start time stamp of the device response signal is earlier than the time stamp of any attributable human interaction signal (such as keyboard typing, mouse clicking) by more than a preset threshold (in this case, 200 milliseconds), it is determined that the order is abnormal. The threshold is determined based on the statistics of 1000 normal human-computer interaction response delays (mean plus three times the standard deviation). For example, the device response time 1677628819.900 is 250 milliseconds earlier than the earliest user keyboard typing signal 1677628820.150, which exceeds the threshold. Therefore, the record is filtered out. After all such records are aggregated, a signal start order abnormality list is obtained.

[0085] 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 order abnormality list, removes the items with image matching actions, extracts the records that simultaneously satisfy the early start of the response and the absence of image corresponding actions, and obtains an ethical output response abnormality track set;

[0086] Each record in the signal start order abnormality list is checked to detect 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 library generated by learning from historical operation logs. For example, the rule library defines that the interface action of “clicking the ‘Filter’ button” corresponds to the response of “FilterResumeBySchool”. In the abnormal record in this example, the device response is “FilterResumeBySchool”, but the visual frame sequence only displays “mouse pointer hovering over the input box”. No direct mapping is found in the rule library, so this record is retained. Through this process, all records that simultaneously satisfy “early start of device response” and “absence of explicit interface trigger action” are extracted to obtain an ethical output response abnormality track set.

[0087] Please refer to Figure 6 The monitoring and governance classification module includes:

[0088] The response offset extraction submodule retrieves and matches the response behaviors corresponding to the same instruction based on the response numbers, expression contents, and behavior instruction records identified in the ethical output response abnormality track set, identifies expression instances with response performance differences, and extracts record units with behavior state offset characteristics indexed by numbers to generate a response diversity offset instance set;

[0089] Based on the ethical output response abnormal trajectory set, the records are grouped according to the behavior instructions (such as "RankCandidateBySkill"), and in the group, the system retrieves and matches multiple response behaviors of the same instruction to identify performance differences. For example, for the same 100 resumes, the response results of the "RankCandidateBySkill" instruction twice have a statistically significant difference in the gender distribution of candidates, and the difference rate exceeds the preset algorithm fairness risk threshold of 5%. Then it is identified that there is a behavior state deviation between the two responses, and the corresponding record unit is extracted with the response number as the index, including the expression content, instruction, timestamp, and complete device response log. After all such instances are aggregated, a response diversity deviation instance set is generated.

[0090] The governance fragment labeling submodule calls the expression data in the response diversity deviation instance set, performs structural analysis based on the role calling order and permission expression field, screens content items with subject-object configuration conflicts and permission boundary deviation, assigns a governance label to expression instances that meet the conditions, and binds the number and semantic fragments to establish a corresponding relationship to obtain a set of governance expression fragment identifiers.

[0091] The expression data in the response diversity deviation instance set is called, and structural analysis is performed based on the governance ethical risk rule library. The analysis process focuses on subject-object configuration conflicts and permission boundary deviations. For example, if a gender difference deviation is found in a response, it can be traced back to the algorithm that assigns a maximum weight limit (such as a maximum of 0.3, actually 0.8) to certain implicit characteristics beyond the set limit in the governance guidelines. It is determined that there is a permission boundary deviation, and this instance will be assigned a "governance" label. The event number and specific algorithm parameters (such as "implicit feature weight") that cause the determination are bound to establish a corresponding relationship. All instances that meet the conditions and their corresponding relationships are collected to obtain a set of governance expression fragment identifiers.

[0092] The content structure classification submodule encodes each expression instance according to the behavior label and semantic pattern based on the semantic classification information and expression field structure of the governance expression fragment identifier set, constructs a unified field mapping structure, and archives all labeled expressions after grouping and reconstruction according to the field group to obtain an ethical risk monitoring and governance result set.

[0093] According to the information of the expression fragment identification set to be governed, each instance is encoded in multiple dimensions according to its behavior label, semantic mode and AI ethical risk core classification, such as {Category: 'Algorithm', Action: 'Rank', Mode: 'Bias.Gender'}, the system constructs a unified mapping structure containing fields such as "event number", "user identity", "risk classification" (as shown in Table 3), merges and restructures all labeled expression instances and archives them, at the same time, the result automatically triggers the verification and feedback mechanism, compares with the artificial audit result, the confirmed cases will be used as new data to update the "AI ethical risk feature library" and the identification model, forming a closed loop of continuous improvement of robustness, and finally obtaining the ethical risk monitoring and governance result set.

[0094] Table 3: Ethical risk event classification structure table

[0095] As shown in Table 3, the table shows the structured classification results of the final output, which standardizes the identified ethical risk events and their core elements, and clearly identifies the risk core classification to which they belong.

[0096] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An AI ethical risk monitoring and governance system based on robust artificial intelligence, 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 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 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 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.

2. The AI ​​ethical risk monitoring and governance system based on robust artificial intelligence according to claim 1, characterized in that: 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 ethical expressions includes high-frequency sensitive expression statements, guiding and ambiguous semantic categories, and permission mismatch intent markers. The set of abnormal ethical output response trajectories includes abnormal response timing segments, missing image action records, and prematurely triggered device response paths.

3. The AI ​​ethical risk monitoring and governance system based on robust artificial intelligence according to claim 1, characterized in that, 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.

4. The AI ​​ethical risk monitoring and governance system based on robust artificial intelligence according to claim 1, characterized in that, 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.

5. The AI ​​ethical risk monitoring and governance system based on robust artificial intelligence according to claim 1, characterized in that, 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.

6. The AI ​​ethical risk monitoring and governance system based on robust artificial intelligence according to claim 1, characterized in that, 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.

7. The AI ​​ethical risk monitoring and governance system based on robust artificial intelligence according to claim 1, characterized in that, The system also includes: 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.

8. The AI ​​ethical risk monitoring and governance system based on robust artificial intelligence according to claim 1, characterized in that, 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.

Images