Artificial intelligence information auditing system and method based on distributed micro-service architecture

Abstract

The application relates to the technical field of intelligent auditing, in particular to an artificial intelligence information auditing system and method based on a distributed micro-service architecture, which comprises the following steps: analyzing to-be-audited information at a gateway layer, extracting a lightweight decision context, the decision context comprising an information type, a semantic key frame and an emergency degree mark; binding a target micro-service instance according to the decision context, preloading the decision context to a memory work area of the target micro-service instance, triggering a distributed auditing engine to execute multi-modal analysis and generate an atomic auditing decision; and aggregating multiple atomic auditing decisions of the same information flow, generating a final global consistency decision through time window conflict resolution processing. The application improves the data volume compression ratio and reduces the bandwidth and storage pressure of the auditing system. Meanwhile, in combination with an information emergency degree dynamic calculation mechanism, the application realizes rapid and preferential response to emergent public opinion information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent auditing, in particular to an artificial intelligence information auditing system and method based on a distributed micro-service architecture. BACKGROUND

[0002] A large amount of information content submitted by users is involved in a bidding platform, such as bidding enterprise data, project performance proof, qualification documents, bidding clarification responses, online exchange comments, etc. These information types are complex and have various formats, including structured data and a large amount of unstructured information such as text, images, and attachments.

[0003] To ensure the compliance operation and transaction safety of the platform, information auditing mechanisms are generally introduced into bidding service platforms to identify and intercept sensitive information, false qualifications, malicious attack comments, commercial defamation, link induction, and other bad content. However, the existing information auditing schemes face the following prominent problems in the actual operation scenario of the bidding platform:

[0004] The text in the bidding information often has a mixed feature of "semi-official + business terminology", containing a large amount of industry abbreviations, policy references, document numbers, qualification descriptions, etc., and the standard language model is prone to misjudgment or omission. At the same time, some illegal information (such as external link induction and fake seal pictures) has a high degree of concealment and requires a combination of semantics and context judgment, which improves the requirement for the content understanding ability of the auditing system.

[0005] Before the centralized opening of the platform or the bidding deadline, the user operation traffic increases dramatically, and the traditional centralized auditing system is prone to problems such as auditing queue congestion and model calculation delay, resulting in delayed or even invalid auditing results, affecting the normal bidding process, and there is an obvious system bottleneck risk. Current platform information auditing mostly relies on static rule engines or fixed text classifiers, which cannot identify similar illegal behaviors published by the same bidding unit on multiple pages and multiple nodes, and the auditing results are scattered, judgment conflicts occur frequently, and there is a lack of global consistency judgment ability.

[0006] With the continuous changes in industry policies, compliance standards, and bidding strategies, illegal expression methods are also evolving. The existing auditing system often lacks online optimization and bias calibration mechanisms for the model, which is prone to the phenomenon of "outdated model running continuously", resulting in a gradual decline in auditing accuracy. SUMMARY

[0007] The present application provides an artificial intelligence information auditing system and method based on a distributed micro-service architecture, which has an information auditing method with the ability of lightweight context expression, dynamic semantic scheduling, global consistency judgment, and adaptive calibration.

[0008] The artificial intelligence information auditing method based on a distributed micro-service architecture comprises the following steps:

[0009] S1: parsing the information to be audited at the gateway layer, extracting a lightweight decision context, the decision context including information type, semantic key frame, and emergency degree label;

[0010] S2: binding a target microservice instance according to the decision context, preloading the decision context to the memory work area of the target microservice instance, triggering a distributed audit engine to perform multi-modal analysis and generate an atomic audit decision;

[0011] S3: aggregating multiple atomic audit decisions of the same information flow, generating a final globally consistent decision through time window conflict resolution processing.

[0012] Optionally, the S1 specifically includes performing the following at the gateway layer:

[0013] S11: parsing a text data stream to obtain a semantic key frame;

[0014] S12: calculating an initial emergency degree based on information source credibility and content sensitive word density;

[0015] S13: information type identification: according to the transmission protocol header characteristics and content structure characteristics, fusion determination is made as text, image, video or cross-modal combination type.

[0016] Optionally, the parsing of the text stream includes extracting a core predicate-argument structure through a dependency syntax analyzer to generate a semantic key frame, the semantic key frame including a subject-predicate-object triple.

[0017] Optionally, the S12 further introduces a time decay function: when the information is associated with a real-time hot event, the initial emergency degree is corrected by the time decay function to obtain a final emergency degree label.

[0018] Optionally, the binding of the target microservice instance in the S2 includes calculating a semantic key frame hash value in the decision context, and performing Hamming distance matching with a context signature set in the memory work area of each microservice instance; if there is a microservice instance with a matching distance less than a dynamic matching threshold, the current microservice instance is bound as the target microservice instance, otherwise a new microservice instance is created and the signature set is initialized.

[0019] Optionally, the S2 further includes encoding the decision context in a memory work area snapshot format, and directly writing to a memory work area snapshot area of the target microservice instance, and triggering atomic replication from the snapshot area to the execution area through a memory barrier instruction;

[0020] In the target microservice instance, multi-violation matching is performed based on the preloaded semantic key frame, and an atomic audit decision including decision type and confidence is output, the violation matching preferentially performing near real-time comparison using a historical decision mode cache in the memory work area.

[0021] Optionally, the dynamic matching threshold comprehensively considers the current system load and the urgency score of the information to be audited, and specifically includes setting a basic matching threshold, representing the maximum acceptable Hamming distance range of the 64-bit semantic hash signature, obtaining the current system CPU utilization, normalizing it into a load coefficient, using the urgency score corresponding to the information to be audited, setting a relaxation factor according to the urgency score, and calculating the dynamic matching threshold based on the basic matching threshold, the current system CPU utilization, and the relaxation factor.

[0022] Optionally, the S3 includes appending a decision flow fingerprint to all atomic audit decisions of the same information flow, and collecting decision sequences within a sliding time window based on the decision flow fingerprint, the length of the time window being dynamically adjusted according to the urgency.

[0023] Optionally, the conflict resolution process includes:

[0024] If there is a type conflict in the decision sequence within the time window, an implicit state backtracker is activated to perform:

[0025] i. inputting the atomic decision sequence as an observation value into an HMM;

[0026] ii. decoding the optimal hidden state sequence by using a Viterbi algorithm, and outputting a globally consistent decision.

[0027] An artificial intelligence information auditing system based on a distributed micro-service architecture, used to implement the above-mentioned auditing method, includes:

[0028] A gateway layer module: used to receive information to be audited and perform semantic analysis, extract lightweight decision context including information type, semantic key frame, and urgency label;

[0029] A micro-service instance cluster: used to bind target micro-service instances according to the decision context, and pre-load the context to the memory work area of the target micro-service instances;

[0030] An auditing engine module: deployed in each target micro-service instance, configured to perform multi-modal analysis based on the decision context, and generate atomic audit decisions;

[0031] A decision aggregation and conflict resolution module: used to perform time window aggregation and conflict determination on multiple atomic audit decisions of the same information flow, and output a final globally consistent audit result.

[0032] The beneficial effects of the present application are:

[0033] 1.The application introduces a dependency syntax driven semantic key frame extraction mechanism in the gateway layer of information review, compresses the information to be reviewed into predicate-argument triplets, and generates a 64-bit semantic hash signature. Compared with traditional BERT embedding or full modal feature extraction, this structure improves the data volume compression ratio and reduces the bandwidth and storage pressure of the review system. At the same time, combined with the dynamic calculation mechanism of information urgency, the rapid and preferential response to sudden public opinion information is realized. Through the completion of structure compression and urgency scoring in the gateway layer, the application realizes edge content pre-judgment and screening distribution, effectively supporting million-level concurrent review tasks.

[0034] 2.The traditional micro-service scheduling generally uses polling or static rules, which can easily lead to context fragmentation and high state migration cost. The application realizes dynamic instance binding driven by semantics by constructing a "semantic key frame hash signature-memory signature set" Hamming distance matching mechanism. Combined with the system load factor and information urgency, the matching threshold is dynamically adjusted to realize adaptive optimization of the scheduling strategy. In addition, the target instance preloads the context and executes decision generation in an atomic way, and caches historical decision patterns to realize redundant judgment reuse and control the overall response delay. This mechanism effectively improves the consistency and decision accuracy of similar information processing while maintaining high scalability, and is particularly suitable for attack variant intensive scenarios (such as ad-induced, borderline content, etc.).

[0035] 3.To solve the problems of asynchrony, error and conflict in atomic review decisions in a distributed environment, the application proposes a hidden state backtracking resolution mechanism based on the Viterbi algorithm, which maps the atomic decision sequence in a sliding time window to an observation sequence, and derives the optimal hidden state path through a hidden Markov model, thereby outputting a global review result with consistency and fault tolerance. It has stronger noise resistance and semantic consistency recognition ability. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only illustrate the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0037] Figure 1 The review method flowchart of the embodiment of the present application;

[0038] Figure 2 The review system composition schematic diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0039] The application will be described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art can also use other alternative ways to implement it. Moreover, the drawings are only used to more specifically describe the embodiments and are not intended to specifically limit the application.

[0040] As Figure 1 shown, the artificial intelligence information auditing method based on distributed micro-service architecture includes the following steps:

[0041] S1: parsing the information to be audited at the gateway layer, extracting a lightweight decision context, the decision context including information type, semantic key frame and urgency label.

[0042] S1 performs the following steps at the gateway layer:

[0043] S11, text stream processing: extracting core predicate-argument structure through a dependency syntax analyzer to generate semantic key frames. The semantic key frame is composed of triples, with the format: <action subject, object, modifier>; for example, <user, report, fraud link>, which is used to capture the core semantic relationship in the text and significantly compress the feature representation dimension.

[0044] The semantic key frame extraction process driven by dependency syntax is as follows:

[0045] a. Perform basic cleaning operations on the text to be audited, including full-width and half-width unification, removal of emoticons, URL and number normalization, etc.

[0046] b. Split the text into independent sentences and perform part-of-speech tagging for each sentence to provide word class prior information for the dependency syntax analyzer, such as verbs, nouns, adverbs.

[0047] c. Call the embedded lightweight dependency syntax analyzer to build a dependency relation tree for each sentence and extract the following three types of key dependency relations:

[0048] Subject-verb relationship (SBV): identify the action initiator (i.e. action subject);

[0049] Verb-object relationship (VOB): identify the action object (i.e. action object);

[0050] Adverbial relationship (ADV, ATT): identify the mood words, time words, place words, etc. used to modify the action.

[0051] d. Traverse the dependency relation tree, select all main verbs as semantic center predicate nodes, exclude function words and conjunction structures, and ensure that the semantic expression is clear and independent.

[0052] e Verb-centric, combine the subject, object and modifiers connected to it to build a triple structure, unified format is: <action subject, object, modifier>, if missing a certain item, can be set to empty or default value to ensure structural consistency, the final triple structure is packaged as a structured object, together with information type and urgency score to form a lightweight decision context for downstream microservices to call.

[0053] S12, dynamic urgency marking: urgency is dynamically generated based on content sensitivity and public opinion hotspots:

[0054] Where base urgency represents the base urgency, which is weighted by the sensitivity of sensitive words (such as "pornography" and "fraud") and the credibility of the information source, α represents the correlation coefficient of hot events, which scores the correlation between information content and hot events through a network public opinion platform (such as microblog hot search), β is the time decay factor, which controls the decay speed of the hotness of the emergency event over time, initially β = 0.1, then reduced to β = 0.01, t now represents the current time, t event represents the first time the associated event is monitored. Real-time calculation is performed in the FPGA acceleration module deployed in the gateway node, with a single operation delay of less than 2ms, meeting the large-scale concurrent demand.

[0055] base urgency = λ1·S content + λ2·(1-C source ); Where S content represents the content sensitivity score, which represents the density of sensitive words, high-risk phrases and offensive semantics in the information to be reviewed, with a value range of [0, 1], the higher the value, the more sensitive, C source is the information source credibility score, which represents the historical reputation level or platform certification weight of the information publisher, with a value range of [0, 1], the higher the value, the more credible, λ1, λ2 are weight coefficients, satisfying λ1 + λ2 = 1, set λ1 = 0.7, λ2 = 0.3, emphasizing the dominant role of content factors.

[0056] S13, information type identification: adopt a fusion decision mechanism, combining network protocol layer and content layer dual features:

[0057] According to the transmission protocol header, it is preliminarily determined that the information is text, image or video;

[0058] If the protocol header and the content features are detected to be contradictory, such as the header is marked as text / html, and the video frame signature is extracted in the content, the anomaly detector is activated to determine whether it is a camouflage attack; if it is not an attack behavior, it is marked as a cross-modal combined information type. This contradiction resolution mechanism can effectively prevent malicious information type camouflage, and improve the accuracy and security of the review.

[0059] The semantic key frame extraction, urgency calculation, and information type determination are all pre-processed in the gateway node, avoiding the bottleneck of large-scale information forwarding on the backend service. The predicate-argument triple and hash signature are used to replace high-dimensional semantic embedding, realizing lightweight review context packaging. The protocol layer and content layer information are combined for type confirmation, avoiding camouflage risks, and providing a more stable type basis for subsequent micro-service dynamic binding.

[0060] S2: According to the decision context, bind the target micro-service instance, preload the decision context to the memory work area of the target micro-service instance, trigger the distributed review engine to perform multi-modal analysis and generate atomic review decisions.

[0061] The target micro-service instance refers to the specific micro-service running unit that is bound based on semantic feature matching or scheduling strategy for the decision context of a piece of information to be reviewed in a distributed micro-service architecture. Each micro-service instance is a relatively independent review processing unit, integrating a review engine, which runs on a certain node in the micro-service cluster and has the following capabilities:

[0062] Receive and preload the decision context;

[0063] Maintain context signature cache (set) in memory;

[0064] Perform multi-modal review logic (pattern matching, confidence judgment);

[0065] Output atomic review decisions.

[0066] The target micro-service instance is the one that is most suitable for the current information flow selected from all available instances through the semantic hash signature matching algorithm, that is:

[0067] Either has processed similar content (semantically similar);

[0068] Or is newly created specifically for the current information (first occurrence, cold start);

[0069] For example: a comment text: "The explosion scene video is too real!", its semantic key frame hash is 0x8F21...E31A; there are 20 audit instances in the current cluster; among them, the signature set in instance #7 contains 0x8F21...E31B (Hamming distance 1); then this instance is selected as the target micro-service instance, that is, the binder of this piece of information.

[0070] S2 specifically comprises the following steps:

[0071] S21, dynamic instance binding: for the semantic key frame of the current information to be audited, calculate its hash signature value, and compare it with the context signature set maintained by the memory work area of each active target micro-service instance, using Hamming distance as the matching index. If there is a context signature set in some instance whose Hamming distance with the current hash signature value is less than the dynamic matching threshold d th , then immediately bind the instance; if there is no match, create a new instance dynamically and initialize its context signature set.

[0072] The dynamic matching threshold is calculated as:

[0073] d th =base threshold ·(1-load factor )+urgency·adaptive slack ;

[0074] Wherein, base threshold represents the maximum number of difference bits allowed for a 64-bit hash signature, base threshold = 8, load factor represents the current load level of the system, adaptive slack is a dynamic relaxation factor, defined as:

[0075]

[0076] The hash signature value of the semantic key frame is calculated as follows:

[0077] S211. Input semantic key frame from text stream, format is standardized triple: <action subject, action, object>; For example: <user, report, fraud link>.

[0078] S212. Encoding preparation: the three elements in the triple are processed as follows:

[0079] Perform Unicode standardization;

[0080] Remove stop words and spaces;

[0081] If the element is a phrase, concatenate the words in order into a single string.

[0082] For example: <user, report, fraud link> → "user reports fraud link".

[0083] S213. String normalization and bit string generation: Take the above spliced result as the input string, use the lightweight hash encoding algorithm xxHash to generate a 64-bit binary signature value as the hash signature of the current semantic key frame. Output format example: Semantic key frame: <user, report, fraud link> → Hash signature: 0x9A7F4D33B8021A7C; The signature is the context semantic fingerprint available for comparison.

[0084] The Hamming distance calculation method (for microservice instance binding matching) is as follows:

[0085] S214. Comparison object: The semantic key frame hash signature (binary string A) of the current information to be audited;

[0086] A set of historical signatures (binary string B) pre-stored in the memory work area of a microservice instance.

[0087] S215. Definition of Hamming distance: Hamming distance represents the number of different bits in two equal-length binary strings. Let two 64-bit signatures be:

[0088] A = a1a2...a 64 ;

[0089] B = b1b2...b 64 ;

[0090] The Hamming distance calculation method is: where is an indicator function that takes 1 when the condition is met, otherwise 0.

[0091] S216. Actual implementation: In the implementation layer, you can use bit operations to speed up:

[0092] Perform XOR operation (XOR) on two 64-bit integers:

[0093] Count the number of binary bits that are 1 (i.e., the number of bit differences), which can be obtained using the CPU built-in instruction popcount(diff). For example:

[0094] A: 1001101001111111...;

[0095] B: 1001101001111011...;

[0096] XOR: 0000000000000100...

[0097] Hamming distance: 1.

[0098] S217. Matching strategy: compare the above calculation results with dynamic matching threshold d th ;

[0099] If there is a Hamming(A, B i )<d th , the current instance is identified as a "semantic affinity" instance, and is bound as a target microservice instance; otherwise, trigger the cold start logic, create a new instance and initialize its context signature set.

[0100] S22, context preloading: after determining the target microservice instance, encode the current decision context in a structured snapshot format, and write it to the memory work area snapshot area of the instance through direct memory write operation, realizing context interruption-free transfer. Subsequently, complete atomic copy from snapshot area to execution area through memory barrier instruction (MOVDIR64B under x86 architecture), guaranteeing execution context consistency and supporting high concurrency.

[0101] Snapshot area: read-only data area, storing structured context copy;

[0102] Execution area: online decision execution unit real-time access area;

[0103] Atomic copy feature: each round of copy width is 512 bits, with a delay of less than 10 ns.

[0104] S23, atomic decision generation: after completing context preloading, the target microservice instance immediately performs multi-modal violation pattern matching analysis based on the semantic key frames in the context, generating atomic audit decision:

[0105] Fields include:

[0106] Audit type, including violation, normal, and suspected;

[0107] Decision confidence (floating point value 0-1);

[0108] Instance memory work area maintains decision pattern triangular matrix cache:

[0109] Pattern 1: <action subject, action, object> → decision type;

[0110] Pattern 2: <visual signature, text triple> → confidence correction coefficient;

[0111] If the Hamming distance between the current semantic key frame and a pattern in the cache is ≤ 2, the decision result is directly reused, and the confidence is decayed (default decay 20%). When a new instance is initialized, the Top100 high-frequency patterns can be pulled from the adjacent nodes as preheating cache to shorten the initial processing delay.

[0112] Application example scenario: rumor review in breaking news events.

[0113] Input information: semantic key frame: <unknown account, publish, chemical plant explosion in a certain place>

[0114] Urgency score: urgency = 0.95;

[0115] Signature calculation: current frame hash: 0x8A3D...C7F2;

[0116] Instance A signature set: [0x8A3D...C7F1, 0x91B2...E4A3] (Hamming distance = 1 from the first signature);

[0117] Dynamic matching threshold calculation (current CPU utilization is 70%): Matching result: 1 < 6 → successfully bind instance A;

[0118] Decision generation:

[0119] Preload context to A instance snapshot area, atomic copy to execution area;

[0120] Match cache pattern <*, publish, chemical plant accident> → preset as "fake information";

[0121] Original confidence: 0.92 → revised confidence after cache reuse:

[0122] Confidence final = 0.92 · 0.8 = 0.736;

[0123] Output atomic review decision: {"type": "violation", "confidence": 0.736}.

[0124] The foregoing atomic decision generation is as follows:

[0125] S231. After the target microservice instance receives the context data written from the gateway layer, it first extracts the semantic key frame of the current review task from the execution area memory, including:

[0126] Predicate-argument triple, hash signature value, urgency score;

[0127] S232. Start the violation pattern matching engine (pattern library call): The microservice instance maintains a violation pattern knowledge base internally, with a multimodal mapping table structure, mainly including the following two types of matching rules:

[0128] Pattern Type 1: Text Pattern (Structured Semantic Matching)

[0129] Format: <Main keywords, action keywords, object keywords □ → Decision type>

[0130] Matching methods: exact match or fuzzy match (supports partial match and semantic synonym match);

[0131] Each rule has a priority weight value (such as a preset weight of 0.9 for "<report, explosion, video>").

[0132] Mode Type 2: Multimodal Fusion Mode (Visual + Text)

[0133] Format: <Visual signature, semantic triple → confidence correction factor;

[0134] Matching method: The hash signature is compared with the existing visual signature by Hamming distance. If the distance is less than 2 bits, it is considered an "approximate match". The semantic keyframe information is combined for judgment, and a confidence correction coefficient of 0.8 is provided.

[0135] S233. Matching Process and Rule Selection:

[0136] a) Preliminary matching: Traverse the current semantic keyframe and the known violation patterns in the cache. If there is a complete match or a partial match, record the matching item of the pattern and the corresponding decision suggestion type (such as "false information" or "advertising inducement").

[0137] b) When multiple rules conflict: sort them according to priority weight. If there is a decision conflict (such as one rule matching "violation" and another matching "suspected"), the confidence fusion module is triggered to calculate the comprehensive judgment value.

[0138] S234. Confidence scoring and decay mechanism:

[0139] Original confidence score calculation: Confidence score is calculated based on matching pattern weights and urgency.

[0140] confidence raw =w match ·(1+γ·urgency); where w match The priority weight for rule matching is γ, which is an urgency correction coefficient ranging from 0.2 to 0.5. A cache reuse decay mechanism is used to avoid over-reliance on history; if a historical decision pattern in the cache is hit (hash distance ≤ 2), the confidence level is reduced by 20% from the original level.final = confidence raw · 0.8;

[0141] S235. Finally, the structured audit result is output by the decision generation unit within the instance:

[0142] Audit type (enumerated value: violation / normal / suspected);

[0143] Confidence score (floating point value 0~1).

[0144] S3: Aggregating multiple atomic audit decisions of the same information flow, generating a final global consistent decision through time window conflict resolution processing.

[0145] S3 specifically includes the following:

[0146] S31, Stream decision aggregation: Assign a unique decision stream fingerprint to each information stream to be audited, which is obtained by hashing the "information stream identifier and the first atomic decision timestamp", which is used to identify the aggregation of the information within the system:

[0147] fingerprint = SHA3(FlowID || T0); Where FlowID represents the unique identifier of the information stream, T0 represents the timestamp of the first atomic audit decision, || represents the string concatenation operation, and SHA3 outputs a 256-bit hash value.

[0148] All atomic decisions with the same fingerprint are collected within a sliding time window for subsequent consistency judgment, and the time window length is dynamically controlled by the information urgency:

[0149]

[0150] S32, Conflict resolution processing: If there is a type conflict (e.g. part of the "violation", part of the "compliance") in the atomic audit decisions collected within the time window, the system will start the hidden state backtracker for conflict resolution.

[0151] The atomic decision sequence is regarded as the observation value sequence, which is input into the preset Hidden Markov Model (HMM):

[0152] State space:

[0153] S1: Violation state; S2: Compliance state;

[0154] Initial probability: P(S1) = 0.6, P(S2) = 0.4;

[0155] State transition probability (example):

[0156] P(S1→S1) = 0.7, P(S1→S2) = 0.3;

[0157] P(S2→S2)=0.6, P(S2→S1)=0.4;

[0158] The Viterbi algorithm is used to decode the atomic decision sequence and output the optimal hidden state path as the final globally consistent decision, as follows:

[0159] S321: Problem Modeling and Input:

[0160] Problem modeling: Treat the aggregated atomic review decision sequence within the time window as an observation sequence O = [O1, o2, ..., o T ], each o t The "decision type" (e.g., "violation" or "compliance") of the t-th atomic decision is represented, and T represents the total length (time steps) of the atomic review decision sequence.

[0161] Modeling assumption: The actual review status of the information flow exhibits Markov property in the short term, that is, its hidden review state sequence Q = [q1, q2, ..., q T Satisfying the first-order Markov condition, each q t Let represent the hidden audit state at the t-th time step. The transition is controlled by the inter-state probability, and the observation is controlled by the state emission probability.

[0162] State space definition: q t ∈{S1: violation, S2: compliance}, that is:

[0163] S1 is a hidden state 1, indicating a "violation" state;

[0164] S2 hidden state 2 indicates "compliance" status;

[0165] S322, Algorithm Input (HMM Parameters): The Viterbi algorithm requires a Hidden Markov Model (HMM) containing the following parameters:

[0166] 1. Initial state probability vector π: π = [π1, π2] = [P(q1 = S1), P(q1 = S2)]; π1 = 0.6, π2 = 0.4;

[0167] 2. State transition probability matrix A:

[0168]

[0169] 3. Emission probability matrix B (i.e., the probability of a state generating an observation): The emission probability of each hidden state to an observation is modeled as follows:

[0170]

[0171] S323, Viterbi algorithm decoding steps:

[0172] Given an observation sequence O = [o1, o2,..., oT], perform the following steps: T

[0173] 1: Initialization (t = 1): For each state s e {S1, S2}, initialize the path probability matrix δ1(s) and the path pointer ψ1(s) as follows:

[0174]

[0175] ψ1(s) = 0;

[0176] π s represents the initial probability of being in state s, δ1(s) represents the probability value of the most probable path among all paths from the start point to state s at time t = 1, and ψ1(s) represents the previous state of the optimal path to state s at time t = 1 (used for path backtracking), represents the emission probability: the probability of hidden state s generating observation o1.

[0177] 2: Recursion (t = 2 to T): For each time step t e [2, T] and each current state s j , perform:

[0178]

[0179] represents the state transition probability from state s i to state s j , represents the emission probability: the probability of hidden state s j generating observation o t ;

[0180] 3: Termination: Find the state with the maximum termination probability:

[0181] Record the maximum path probability: represents the state where the optimal path terminates at time T.

[0182] 4: Path backtracking: Starting from , backtrack the hidden state path through the pointer matrix ψ t :

[0183]

[0184] Finally, obtain the most probable hidden state path: ​

[0185] S324, Global Consistency Decision Generation: Optimal Hidden State Path Q * Statistical analysis was conducted, and the state that appeared most frequently was selected as the global consistency audit result within that time window.

[0186] If S1 (violation) occurs the most times → output "violation";

[0187] If S2 (compliance) is the primary parameter, then the output will be "compliance".

[0188] like Figure 2 As shown, an AI-powered information review system based on a distributed microservice architecture is used to implement the aforementioned review methods, including:

[0189] Gateway layer module: Used to receive information to be reviewed and perform semantic parsing to extract lightweight decision context, including information type, semantic keyframes and urgency markers;

[0190] Microservice instance cluster: used to bind the target microservice instance according to the decision context, and preload the context into the memory workspace of the target microservice instance;

[0191] The audit engine module is deployed in each target microservice instance and configured to perform multimodal analysis based on the decision context and generate atomic audit decisions.

[0192] The decision aggregation and conflict resolution module is used to aggregate and determine conflicts among multiple atomic review decisions in the same information flow within a time window, and output the final globally consistent review result.

[0193] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0194] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An artificial intelligence information auditing method based on a distributed microservice architecture, characterized in that, The method comprises the following steps: S1: parsing the information to be audited at the gateway layer, extracting a lightweight decision context, the decision context comprising information type, semantic key frame and emergency degree label; specifically comprising the following steps performed at the gateway layer: S11: parsing the information to be audited to obtain a semantic key frame; S12: calculating an initial emergency degree based on information source credibility and content sensitive word density of the information to be audited; S13: identifying information type: according to transmission protocol header features and content structure features of the information to be audited, determining as text, image, video or cross-modal combination type; The parsing of the information to be audited comprises extracting core predicate-argument structure through a dependency syntax analyzer to generate a semantic key frame, the semantic key frame comprising action subject, object and modifier triplets; S2: binding a target microservice instance according to the decision context, preloading the decision context to a memory work area of the target microservice instance, triggering a distributed audit engine to perform multi-modal analysis and generate an atomic audit decision; The binding of the target microservice instance comprises calculating a semantic key frame hash value in the decision context, and performing Hamming distance matching with a context signature set in a memory work area of each microservice instance; if there is a microservice instance with a matching distance less than a dynamic matching threshold, the microservice instance is bound as the target microservice instance, otherwise, a new microservice instance is created and the signature set is initialized; The S2 further comprises encoding the decision context in a memory work area snapshot format, and directly writing to a memory work area snapshot area of the target microservice instance, and triggering atomic replication from the snapshot area to the execution area through a memory barrier instruction; In the target microservice instance, multi-violation matching is performed based on the preloaded semantic key frame, and an atomic audit decision comprising decision type and confidence is output, the multi-violation matching preferentially performing near real-time comparison using a historical decision mode cache in the memory work area; S3: aggregating multiple atomic audit decisions of the same information stream, and generating a final globally consistent decision through time window conflict resolution processing. 2.The artificial intelligence information auditing method based on the distributed micro-service architecture according to claim 1, characterized in that, The S12 further introduces a time decay function: when the information to be audited is associated with a real-time hot event, the initial emergency degree is corrected by the time decay function to obtain a final emergency degree label. 3.The artificial intelligence information auditing method based on the distributed micro-service architecture according to claim 1, characterized in that, The dynamic matching threshold considers the current system load and the emergency degree label of the information to be audited, specifically comprising setting a basic matching threshold, representing a maximum acceptable Hamming distance range of a 64-bit semantic hash signature, obtaining a current system CPU utilization, normalizing it into a load coefficient, using the emergency degree label corresponding to the information to be audited, setting a relaxation factor according to the emergency degree label, and calculating the dynamic matching threshold based on the basic matching threshold, the load coefficient and the relaxation factor. 4.The artificial intelligence information auditing method based on a distributed micro-service architecture according to claim 1, wherein, The S3 comprises attaching a decision flow fingerprint to all atomic audit decisions of the same information stream, and collecting a decision sequence in a sliding time window based on the decision flow fingerprint, the length of the time window being dynamically adjusted according to the emergency degree. 5.The artificial intelligence information auditing method based on the distributed micro-service architecture according to claim 4, characterized in that, The generation of the final globally consistent decision through time window conflict resolution processing comprises: If there is a decision type conflict in the decision sequence in the time window, activate a hidden state backtracker to perform: i. inputting the atomic decision sequence as an observation value into a hidden Markov model; ii. Decode the optimal hidden state sequence by the Viterbi algorithm, and output the globally consistent decision.

6. An artificial intelligence information auditing system based on a distributed microservice architecture, used to implement the artificial intelligence information auditing method based on the distributed microservice architecture according to any one of claims 1-5, characterized in that, Comprise: Gateway layer module: for receiving information to be audited and performing semantic parsing, extracting lightweight decision context including information type, semantic key frame and urgency label; Microservice instance cluster: for binding target microservice instances according to the decision context, and preloading the decision context to the memory work area of the target microservice instances; Audit engine module: deployed in each target microservice instance, configured to perform multi-modal analysis based on the decision context, and generate atomic audit decisions; Decision aggregation and conflict resolution module: for time window aggregation and decision type conflict determination of multiple atomic audit decisions of the same information stream, and output the final globally consistent decision.

Images