System and Method for Preventing Training Data Poisoning of AI Shopping Agents Based on Blockchain Integrity Proof

A comprehensive security system for AI shopping agents addresses data poisoning attacks by verifying data sources, detecting anomalies, and using blockchain integrity proofs to ensure data integrity, enhancing reliability and fairness while reducing costs and promoting a trustworthy AI ecosystem.

KR103045226B1Active Publication Date: 2026-11-06안성현
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
KR1020250051974
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2025-04-22
Publication Date
2026-11-06
Estimated Expiration
2045-04-22

AI Technical Summary

Technical Problem

Existing AI shopping agents are vulnerable to training data poisoning attacks, lacking mechanisms to verify data reliability, detect manipulated data, ensure temporal consistency, and maintain cross-validation among distributed sources, which compromises the integrity of training data and biases recommendations.

Method used

A comprehensive security system utilizing a source verification unit, multi-faceted verification unit, blockchain integrity proof module, and reliability update unit to detect and block data poisoning attacks by ensuring data integrity, isolating suspicious data, and adapting to new attack patterns.

Benefits of technology

The system enhances AI reliability, ensures market fairness, responds to security threats, improves data quality, reduces costs, supports regulatory compliance, and promotes a trustworthy AI ecosystem by preventing data manipulation and maintaining data integrity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112026065863190-PAT00010_ABST
    Figure 112026065863190-PAT00010_ABST
Patent Text Reader

Abstract

The present invention relates to a system and method for preventing training data addiction in an AI shopping agent. A source verification unit sets differential verification passing criteria based on the reliability according to the reputation of the collected data sources. A multi-faceted verification unit identifies suspicious data and isolates it from the training dataset by detecting at least one of poor temporal consistency, multidimensional outliers, or inconsistencies between distributed sources under the set criteria. A reliability update unit lowers the reliability of the suspicious data source and feeds it back to subsequent data verification. The present invention forms a closed-loop purification cycle in which verification, isolation, reliability lowering, and differential verification reinforcement are coupled to cumulatively block attempts at data addiction, and ensures the reliability and fairness of the system by training a model with purified data.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to security technology for artificial intelligence (AI) systems, and more specifically, to a system and method for preventing learning data addiction that guarantees the integrity of product information data learned by an AI shopping agent and protects it from malicious data manipulation. Background Technology

[0002] AI shopping agents continuously improve recommendation performance by learning from data such as product information, user reviews, pricing data, and sales performance. While this learning process is a core competitive advantage of AI agents, it is also a vulnerable point to malicious attacks.

[0003] Training Data Poisoning is an attack technique in which an attacker intentionally injects manipulated data into an AI system's training dataset to distort the model's behavior. In an AI shopping agent environment, such attacks can manifest in the following forms.

[0004] First is a product information manipulation attack. Sellers induce the AI ​​to rate their products excessively high by exaggerating or falsely listing performance data. For example, this involves displaying a product that is actually rated 80 points as 95 points or adding features that do not exist.

[0005] Second is review manipulation attacks. These attacks distort the AI's reputation analysis by generating a large volume of fake positive reviews or injecting negative reviews about competing products. Review data is a primary target of these attacks, as it significantly influences the AI ​​agent's recommendation decisions.

[0006] Third is the manipulation of price and sales data. This involves inducing the AI ​​to misjudge the popularity of specific products through false discount rates, manipulated sales volumes, and fake stock shortage signals.

[0007] Fourth is a temporal attack. Manipulated data is injected intensively at specific points in time to distort the AI's trend analysis. For example, positive data is injected intensively immediately after a new product launch to ensure an excessively high evaluation during the initial learning phase.

[0008] Current AI system security technologies are primarily focused on defending against attacks on the model itself (e.g., adversarial examples) or attacks during inference, while mechanisms to systematically guarantee the integrity of training data are lacking. In particular, the following problems exist in environments that continuously collect and train data from multiple external sources, such as AI shopping agents.

[0009] First, verifying the reliability of data sources is difficult. It is practically impossible to verify all the data generated from thousands of sellers and millions of users.

[0010] Second, it is difficult to distinguish between manipulated data and normal data. In particular, sophisticatedly manipulated data falls within the statistically normal range, making detection very difficult.

[0011] Third, tracking of data changes over time is insufficient. There is a lack of a mechanism to detect patterns where information about the same product changes abnormally over time.

[0012] Fourth, there is a lack of cross-validation among distributed data sources. There is no system to verify the consistency of information provided by multiple sources regarding the same product.

[0013] Fifth, it is difficult to detect abnormal behavior in trained models. It is difficult to retrospectively detect behaviors where a model trained on addictive data evaluates a specific product abnormally highly.

[0014] Furthermore, while existing blockchain-based data integrity guarantee technologies are effective in preventing data tampering, they cannot prevent incorrect data from being entered in the first place. In other words, they fail to solve the "garbage in, garbage out" problem.

[0015] The reliability and fairness of AI shopping agents are the foundation of the entire B2A commerce ecosystem. If training data is manipulated to unfairly favor specific products or unfairly exclude competitors, consumers will lose trust in the AI ​​agents, and there is a risk that the entire ecosystem will collapse.

[0016] Therefore, there is an urgent need for a comprehensive security system capable of systematically guaranteeing the integrity of training data for AI shopping agents and detecting and blocking various forms of data poisoning attacks in advance. The problem to be solved

[0017] The present invention aims to solve the problems of the prior art described above by providing a comprehensive security system capable of guaranteeing the integrity of training data for an AI shopping agent and detecting and blocking various forms of data addiction attacks.

[0018] Specifically, the problem that the present invention aims to solve is as follows.

[0019] First, it provides a mechanism to detect abnormal signs in product data collected from external sources in real time.

[0020] Second, it detects abrupt manipulation by verifying the consistency of temporal data changes for the same product.

[0021] Third, false data is identified by cross-verifying identical product information from multiple distributed data sources.

[0022] Fourth, blockchain technology is utilized to generate and manage tamper-proof integrity proofs for verified data.

[0023] Fifth, monitor the behavior of the trained AI model to detect abnormal recommendation patterns caused by data addiction.

[0024] Sixth, it provides a mechanism to automatically isolate and purify poisoned data.

[0025] Seventh, continuously evaluate and manage the reliability of data providers to block repetitive attacks.

[0026] Eighth, build an adaptive defense system capable of responding to new forms of attacks by learning various attack patterns. means of solving the problem

[65535] A system for preventing training data addiction of an AI shopping agent according to one aspect of the present invention comprises: a source verification unit that identifies the source of training data collected from an external source and differentially sets a verification pass criterion for the training data according to the reliability calculated from reputation information of the identified source; a multi-faceted verification unit that, under the differentially set verification pass criterion, detects at least one of poor temporal consistency, an outlier in a multi-dimensional space, and information discrepancy between distributed sources for the training data to identify suspicious data and isolates the identified suspicious data from the model training dataset of the AI ​​shopping agent; and a blockchain integrity proof module that generates a cryptographic hash value for each of the data that has passed the verification of the multi-faceted verification unit, aggregates a plurality of generated hash values ​​into a Merkle tree structure to calculate a single Merkle root hash value, generates and manages an untamper-proof integrity proof for a verified dataset by recording the Merkle root hash value and verification metadata on a blockchain network, and subsequently identifies tampered data as the suspicious data by comparing a hash value recalculated from the data with the value recorded on the blockchain network. The system includes a reliability update unit that updates the reliability downward for the source of data identified as suspicious data by the multi-faceted verification unit, and feeds the updated reliability downward to the source verification unit to control that stricter verification passing criteria are applied to subsequent training data collected from the source. The system forms a closed-loop purification cycle in which the identification result of the suspicious data lowers the reliability of the source, and the lowered reliability feeds back to strengthening the passing criteria for subsequent data, and is characterized in that a purified training dataset, in which integrity is proven by the blockchain integrity proof module and the isolated suspicious data is excluded, is provided for the model training of the AI ​​shopping agent. Effects of the invention

[0039] The system and method for preventing training data addiction of an AI shopping agent according to the present invention provide the following effects. The present invention cumulatively blocks attempts at data addiction that proceed repeatedly and progressively by adopting a closed-loop structure in which the identification result of suspicious data lowers the credibility of the corresponding source, and the lowered credibility is fed back as a stricter passing standard for subsequent data. Furthermore, the present invention calculates a single Merkle root hash value through Merkle tree-based batch hash processing for data that has passed verification and records it on a blockchain network; thereby resolving the issues of transaction costs and processing speeds that occur when recording individual data one by one on the blockchain, while mathematically guaranteeing the integrity of the entire dataset and providing an audit trail with legal evidentiary value against attempts at data manipulation.

[0040] First, it enhances the reliability of AI systems. By guaranteeing the integrity of training data, it increases trust in the recommendation results of AI agents, prevents biased recommendations caused by data addiction attacks, and creates an environment where consumers can use AI agents with peace of mind.

[0041] Second, it ensures market fairness. It creates a fair competitive environment by blocking fraudulent data manipulation by specific sellers, guarantees that all products are evaluated based on actual quality, and provides small and medium-sized sellers with the opportunity to compete fairly.

[0042] Third, it responds to security threats. It detects and blocks various forms of data addiction attacks, responds adaptively to new attack patterns, and preemptively blocks repetitive attack attempts.

[0043] Fourth, improve data quality. Through continuous data verification, overall data quality is enhanced, errors or inaccurate data are detected and corrected early, and AI learning based on high-quality data becomes possible.

[0044] Fifth, it reduces system operating costs. Automated detection and verification reduce manual review costs, prevent model retraining costs caused by data addiction, and lower total costs through proactive prevention rather than reactive problem solving.

[0045] Sixth, it supports regulatory compliance. It meets regulatory requirements by enhancing the transparency and accountability of AI systems, provides audit trails for data integrity, and secures legal evidentiary value through blockchain-based proofs.

[0046] Seventh, it enhances the value of the entire ecosystem. By fostering a trustworthy AI agent environment, it promotes the growth of the B2A commerce ecosystem, provides incentives for data providers to act honestly, and builds a sound market ecosystem in the long term.

[0047] Eighth, it has high scalability and applicability. It can be applied not only to AI shopping agents but also to various AI systems, can be extended to fields where reliability is critical such as finance, healthcare, and education, and can evolve into a general-purpose AI security framework. Brief explanation of the drawing * FIG. 1 is a configuration diagram showing the entire closed-loop purification cycle of a learning data addiction prevention system for an AI shopping agent according to an embodiment of the present invention. * FIG. 2 is a flowchart showing an abnormal data detection process according to an embodiment of the present invention. * FIG. 3 is a flowchart showing a temporal consistency verification process according to an embodiment of the present invention. * FIG. 4 is a configuration diagram showing a distributed data verification process according to an embodiment of the present invention. * FIG. 5 is a flowchart showing a blockchain integrity proof process according to an embodiment of the present invention. * FIG. 6 is a flowchart showing a model behavior monitoring process according to an embodiment of the present invention. * FIG. 7 is a configuration diagram showing a data isolation and purification process according to an embodiment of the present invention. * FIG. 8 is a flowchart showing a reliability evaluation and management process according to an embodiment of the present invention. Specific details for implementing the invention

[0049] The definitions of terms used in this specification are as follows.

[0050] delete

[0051] "Training Data Poisoning" refers to an attack technique in which an attacker injects intentionally manipulated data into an AI system's training dataset to distort the model's behavior.

[0052] "Data integrity" refers to the state in which data has not been altered or corrupted in an unauthorized manner since its creation.

[0053] An "outlier" refers to a value that deviates significantly from a normal data distribution, meaning a statistically abnormal data point.

[0054] "Temporal consistency" refers to the characteristic where data regarding the same entity (e.g., a product) exhibits a logically consistent pattern of change over time.

[0055] "Distributed Verification" refers to a method of cross-verifying data by collecting it from multiple independent sources rather than relying on a single source.

[0056] "Blockchain Integrity Proof" refers to a mechanism that generates an tamper-proof proof of integrity by recording the hash value of data on the blockchain.

[0057] "Model Behavior Monitoring" refers to the process of detecting abnormal behavioral patterns by continuously observing the inference results of a trained AI model.

[0058] "Data Quarantine" refers to a measure to separate suspicious data from a training dataset so that it is not used to train AI models.

[0059] "Trust Score" refers to an indicator that quantifies the accuracy and reliability of data provided by a data provider in the past.

[0060] A "consensus algorithm" refers to an algorithm that determines the most reliable value among different values ​​collected from multiple independent sources.

[0061] "Adaptive Learning" refers to a mechanism that continuously learns new attack patterns to improve detection capabilities. "Source Verification Unit" refers to a configuration that identifies data sources, calculates reliability based on their reputation, and applies verification pass criteria differentially, corresponding to the source identification and reliability calculation functions of the reliability evaluation system (170) of this specification. "Multi-faceted Verification Unit" refers to a higher-level configuration that identifies suspicious data by encompassing an anomaly data detection module (110), a temporal consistency verification module (120), and a distributed data verification network (130). "Reliability Update Unit" refers to a configuration that updates the reliability of the source of the identified suspicious data downward and feeds it back to subsequent verification, corresponding to the reliability calculator (173) and access control manager (174) of the reliability evaluation system (170).

[0062] Example 1: Overall configuration of a learning data addiction prevention system

[0063] Figure 1 is an overall configuration diagram of a learning data addiction prevention system for an AI shopping agent according to one embodiment of the present invention.

[0064] The learning data addiction prevention system (100) is composed of an anomaly data detection module (110), a temporal consistency verification module (120), a distributed data verification network (130), a blockchain integrity proof module (140), a model behavior monitoring module (150), a data isolation and purification engine (160), a reliability evaluation system (170), and an adaptive learning module (180).

[0065] The abnormal data detection module (110) includes the following components.

[0066] The statistical analyzer (111) analyzes the statistical distribution of product attribute values. It calculates statistical quantities such as the mean, standard deviation, and quantile for each product attribute (e.g., price, performance score, battery life), constructs a distribution model of normal data (normal distribution, log-normal distribution, etc.), maintains a separate statistical model for each product category, and tracks changes in the statistical distribution over time.

[0067] The outlier detector (112) detects outliers that fall outside the normal range. It detects extreme values ​​by applying a standard deviation-based method (e.g., mean ± 3σ), identifies outliers using an Interquartile Range (IQR) method, utilizes a machine learning-based outlier detection algorithm (Isolation Forest, One-Class SVM, etc.), and scores the severity of the detected outliers.

[0068] The correlation analyzer (113) detects correlation anomalies between multidimensional attributes. It learns an expected correlation model between product attributes to identify, for example, "positive correlation between price and performance," verifies whether new data violates the learned correlation pattern, applies a multivariate outlier detection technique, and detects logically impossible combinations of attributes (e.g., "no waterproof function" + "capable of underwater shooting").

[0069] The peer comparison analyzer (114) detects abnormal values ​​compared to similar product groups. It subdivides products by category, price range, brand, etc., calculates the average attribute values ​​of products within the same segment, detects cases where the attribute values ​​of a new product are significantly higher or lower than those of peer products, and performs additional verification to distinguish between exceptionally superior products and manipulated data. Additionally, the anomaly data detection module (110) further includes a pre-processing unit to perform the function of filtering and sanitizing adversarial perturbation that is invisible to the human eye but cleverly inserted to induce misrecognition by the AI ​​model in product images or text descriptions. Meanwhile, the multi-faceted verification unit may further include a multivariate anomaly detection engine that detects anomalies in a multidimensional data space, and the multivariate anomaly detection engine corresponds to a configuration responsible for detecting "anomalies in a multidimensional space" among the temporal inconsistency, anomalies in a multidimensional space, and information discrepancies between distributed sources detected by the multi-faceted verification unit. The multivariate outlier detection engine described above includes: a dimensionality reducer that reduces high-dimensional data to a low dimension; an algorithm library that applies a plurality of outlier detection algorithms to the data space reduced by the dimensionality reducer; a threshold manager that manages thresholds for determining outliers; and a clustering engine that clusters data points and identifies group outliers that do not belong to any cluster as the suspected data. Specifically, the dimensionality reducer analyzes high-dimensional product data composed of multiple attributes such as price, performance score, and battery life, and reduces it to a low-dimensional data space suitable for outlier detection.The above algorithm library is configured to provide various outlier detection algorithms, and detects outliers by applying a plurality of outlier detection algorithms, including machine learning-based outlier detection algorithms (such as Isolation Forest, One-Class SVM, etc.) used in the outlier detector (112), to the reduced data space. The threshold manager manages a threshold value (e.g., 0.8) for determining an anomaly score calculated for a detected outlier as an outlier, and the threshold value may be statically fixed or dynamically adjusted according to the data distribution. The clustering engine clusters data points and determines a group of data points that do not belong to any cluster as a group outlier, identifying them as the suspected data. This multivariate outlier detection engine can be utilized for multivariate outlier detection in the application of the multivariate outlier detection technique by the correlation analyzer (113) and correlation verification step (S240, see FIG. 2), thereby effectively detecting sophisticated data manipulation in a multidimensional attribute space that is difficult to capture with only statistical analysis of a single attribute.

[0070] The temporal consistency verification module (120) includes the following components.

[0071] The time series tracker (121) tracks time series changes in product data. It stores time series data for all attributes of each product, records data change events with timestamps, records the change provider (data provider) and the reason for the change, and visualizes long-term change trends.

[0072] The rate of change analyzer (122) detects abrupt data changes. It calculates the rate of change of data between two consecutive points in time, defines a normal rate of change range for each attribute, flags abrupt changes that exceed a threshold, and analyzes the direction of change (improvement / deterioration).

[0073] The pattern recognizer (123) identifies periodic or suspicious change patterns. It detects repetitive change patterns (e.g., performance score increase every night followed by restoration in the morning), identifies abnormal changes immediately after specific events (e.g., launch of a competitor product), detects coordinated attack patterns occurring simultaneously across multiple products, and distinguishes between normal and abnormal changes considering seasonality or trends.

[0074] The historical consistency verifier (124) verifies consistency with past data. It verifies whether new data is logically consistent with past data, for example, verifies the validity of "battery life changed from 8 hours to 20 hours," detects performance improvements that are impossible with software updates alone without changing the product's hardware specifications, requires re-certification if a previously certified attribute changes, and requires evidence to resolve conflicts with historical data.

[0075] The distributed data verification network (130) includes the following components.

[0076] The source manager (131) manages multiple independent data sources. It maintains a list of trusted data sources, including official manufacturers, certification bodies, independent review sites, government agencies, etc., manages the trust level of each source, verifies the independence between sources to identify sources operated by the same entity, and manages the addition of new sources and the removal of existing sources.

[0077] A multi-source collector (132) collects information about the same product from multiple sources. For each product, it collects data from at least N (e.g., 5) independent sources, performs automated data collection through API or web crawling, synchronizes the collection time to minimize discrepancies caused by time differences, and converts the collected data into a standardized format.

[0078] The consensus algorithm (133) calculates the degree of agreement between the collected data. It compares values ​​reported by multiple sources for the same attribute, checks whether the difference between the values ​​is within an acceptable range (e.g., ±5%), selects a value agreed upon by more than half of the sources as the consensus value, and uses a weighted average or median value if there is no perfect agreement.

[0079] The weighted voting mechanism (134) performs confidence-weighted voting on inconsistent data. It uses the confidence score of each source as a weight, assigns greater weight to opinions from sources with higher confidence, determines the final verification value through weighted voting, and designates data from sources that are minority opinions but highly confident as subjects for further investigation.

[0080] The blockchain integrity proof module (140) includes the following components.

[0081] The hash generator (141) calculates the hash value of data that has passed verification and has not been selected as an isolation target by the data isolation and purification engine (160). It uses a cryptographic hash function of SHA-256 or higher, hashes a data structure including a product ID, attribute name, attribute value, and timestamp, ensures consistency through normalization of the hash input data, and utilizes the hash value as an untamper-proof digital fingerprint.

[0082] The blockchain recorder (142) records the hash value on the blockchain. It selects and uses a public or consortium blockchain, generates a transaction with the hash and metadata of the verified data, submits the transaction to the blockchain network and waits for confirmation, and records the transaction ID and block number for future verification.

[0083] The integrity verifier (143) detects attempts to tamper with data. It calculates the hash of the current data and compares it with the hash recorded in the blockchain. If there is a discrepancy, it determines that it is a tamper and issues a warning. It also isolates the tampered data, restores it to the original data, and records the tampering attempt as a security event.

[0084] The proof history manager (144) manages the integrity proof history. It stores all integrity proof records for each data, tracks the entire change history of the data through the proof chain, provides the proof history upon audit request, and manages the backup and recovery of the proof data.

[0085] The model behavior monitoring module (150) includes the following components.

[0086] The recommendation logger (151) continuously records the recommendation results of the AI ​​model. It logs all recommendation requests and responses, records the list and ranking of recommended products, records the basis of the recommendation (features used, scores), and tracks user feedback (clicks, purchases).

[0087] The bias detector (152) detects abnormally high recommendation frequencies for specific products. It aggregates the recommendation frequencies of each product, detects recommendation frequencies that are excessively high relative to the actual quality of the product, identifies rapid changes in recommendation frequencies over time, and analyzes systematic bias for specific product families.

[0088] The confidence analyzer (153) analyzes changes in the model's prediction confidence. It tracks the confidence score output by the model, detects abnormal changes in the average confidence, identifies cases where the confidence is excessively high only for a specific product, and analyzes the discrepancy between confidence and actual performance (conversion rate).

[0089] The performance monitor (154) detects model performance degradation caused by addiction data. It monitors key indicators such as the model's recommendation accuracy, click-through rate, and conversion rate, detects a rapid decline in performance indicators, compares the performance of the new model with the existing model through A / B testing, and analyzes the cause of the performance degradation to determine whether there is data addiction. The cause data tracker (155) identifies and tracks the training data that contributed to the bias when abnormal recommendation bias for a specific product is detected by the bias detector (152) and model performance degradation is simultaneously detected by the performance monitor (154). Specifically, it extracts data related to the product in which the bias was detected among the data used for model training, analyzes changes in the data in the recent training batch, identifies newly added or significantly changed data, and transmits the identified training data to the data isolation and purification engine (160) as the fourth suspected data.

[0090] The data isolation and purification engine (160) includes the following components.

[0091] The data isolator (161) receives the first suspected data output from the anomaly data detection module (110), the second suspected data output from the temporal consistency verification module (120), the third suspected data output from the distributed data verification network (130), and the fourth suspected data output from the model behavior monitoring module (150), and selects them as isolation targets. It moves the selected isolation targets from the training dataset to an isolation area, blocks the isolated data from being used for model training, records the reason for isolation and evidence, and manages the re-verification process of the isolated data.

[0092] The re-verification module (162) re-verifies the accuracy of the isolated data. It requests a manual expert review, collects information from additional data sources, requests supporting materials from the data provider, and decides whether to restore or permanently delete based on the re-verification results.

[0093] The model retraining manager (163) cleans up the model trained with contaminated training data. It identifies the version of the model affected by the contaminated data, retrains the model with the clean dataset, verifies the performance of the retrained model, and safely replaces it through canary deployment.

[0094] The distribution manager (164) distributes the purified data and model. It creates a new version of the purified training dataset, first distributes the retrained model to a test environment, and after verification is complete, distributes it to a production environment, and prepares a rollback plan to respond immediately in case of a problem.

[0095] The reliability evaluation system (170) includes the following components.

[0096] The accuracy tracker (171) tracks the accuracy of the provided data by each data provider. It records the verification results of the data submitted by each provider, calculates the ratio of accurate data, inaccurate data, and questionable data, analyzes the accuracy trend over time, and separately tracks the accuracy in a specific product category.

[0097] The history manager (172) records the history of past addiction attempts. It records all detected addiction attempts by connecting them to a data provider, analyzes the severity, frequency, and patterns of the addiction attempts, identifies repeat malicious actors, and collects and stores evidence for legal action.

[0098] The reliability calculator (173) dynamically calculates a reliability score. It calculates a score by combining accuracy, historical data, data provision period, and data volume, applies a time decay function that gives higher weight to recent data, generates a normalized score from 0 points (completely distrusted) to 100 points (completely trusted), and updates the change in the reliability score in real time.

[0099] The access control manager (174) applies restrictions to data from low-trust providers. It classifies data providers into grades based on their trust scores, includes data submitted by providers with trust scores below a preset threshold in the isolation targets of the data isolation and purification engine (160), and applies enhanced verification. It blocks data submission from providers with very low trust and presents conditions and processes for trust recovery.

[0100] The adaptive learning module (180) includes the following components.

[0101] The attack pattern database (181) stores and manages known attack patterns. It stores all past detected attack cases in the database, systematically classifies attack types, techniques, objectives, and characteristics, shares attack information with the external security community, and continuously adds new attack patterns.

[0102] The pattern learning engine (182) learns new attack patterns by receiving the characteristics of the data isolated by the data isolation and purification engine (160) and the abnormal recommendation patterns identified by the model behavior monitoring module (150) as learning inputs. It learns the characteristics of attack patterns using a machine learning model, trains the model to automatically identify similar new attacks, discovers unknown new attack types through unsupervised learning, and integrates the learned patterns into the detection logic of the abnormal data detection module (110) and the model behavior monitoring module (150).

[0103] The detection model update manager (183) continuously improves the detection model. It incorporates newly learned patterns into the existing detection model, regularly retrains the model to respond to the latest threats, continuously evaluates and improves model performance, and deploys new detection models in stages.

[0104] The performance evaluator (184) evaluates the performance of the detection system. It measures the true positive rate and false positive rate, calculates metrics such as Precision, Recall, and F1-Score, measures and minimizes detection latency, compares actual performance against performance targets, and derives directions for improvement.

[0105] Example 2: Anomaly Data Detection Process

[0106] FIG. 2 is a flowchart showing an abnormal data detection process according to one embodiment of the present invention.

[0107] The anomaly data detection process is a process of detecting statistical anomalies by analyzing product data collected from external sources in real time, and includes the following steps.

[0108] In the product data collection step (S210), product data is collected from external sources. Data is collected from various sources, such as product information submitted by sellers, review data collected via web crawling, price information from price comparison sites, and performance test results from certification bodies; the data is converted into a standardized format, and the data source and time of collection are recorded as metadata.

[0109] In the statistical distribution analysis step (S220), the statistical distribution of the collected data is analyzed. The mean and standard deviation of each attribute (price, performance score, battery life, etc.) for each product category are calculated, the shape of the distribution (normal distribution, skewness, kurtosis) is analyzed, compared with a normal distribution model built based on past data, and the new data point is probabilistically evaluated to see if it conforms to the normal distribution.

[0110] In the outlier detection step (S230), outliers that deviate from the normal range are detected. Values ​​that deviate from the mean by more than 3 standard deviations are determined as primary outliers, and values ​​that are less than Q1 - 1.5 × IQR or greater than Q3 + 1.5 × IQR are determined as outliers based on IQR(Q3 - Q1). The Isolation Forest algorithm is applied to detect outliers in a multidimensional space, and an anomaly score is calculated for each outlier.

[0111] As an example, the "battery life" attribute in the smartphone category can be cited. The average battery life of the category is 18 hours, with a standard deviation of 3 hours. If a new product X claims a battery life of 40 hours, it is determined to be a first-order outlier as it is an extreme value (40-18) / 3 = 7.33 standard deviations away from the mean. In IQR-based analysis, it is determined to be a second-order outlier as it exceeds Q3 (22 hours) + 1.5 × IQR (6 hours) = 31 hours. An outlier score of 0.95 (high) is assigned.

[0112] In the correlation verification step (S240), correlation anomalies between multidimensional attributes are detected. New data is evaluated based on a learned correlation model between attributes (e.g., correlation coefficient of price and performance of 0.7), cases that deviate significantly from the expected correlation are detected, logically contradictory combinations of attributes are verified based on rules, and multivariate outliers are detected by calculating the Mahalanobis Distance.

[0113] As an example, consider the case where product Y claims "Price: 200,000 won, Performance Score: 95 points." The average price of products with a performance score of 90 points or higher in the same category is 800,000 won. According to the price-performance correlation model, the expected price when the performance is 95 points is 85 ± 150,000 won. The actual price of 200,000 won deviates significantly from the expected range and is judged to be above the correlation.

[0114] In the peer comparison step (S250), abnormal values ​​are detected relative to similar product groups. Products are classified into category, price range, and brand segments, and the average attribute value of products within the same segment is calculated. The case where the attribute value of a new product has a Z-Score of 3 or higher relative to similar products is detected, and the case where the overall category average is normal but abnormal relative to similar products is identified.

[0115] As an example, the "low-cost smartphone (100,000-300,000 won)" segment can be cited. The average performance score of the segment is 65 points, and a new product Z (price 250,000 won, performance 95 points) is registered. Compared to the overall category average (75 points), it is not an outlier, but compared to the average of comparable products (65 points), Z-Score = (95-65) / 8 = 3.75, so it is determined to be an outlier.

[0116] In the anomaly data flagging step (S260), detected anomaly data is flagged. Data with an anomaly score above a threshold (e.g., 0.8) is classified as high-risk, anomaly types (statistical outliers, correlation violations, peer comparison anomalies) are tagged, a request for additional verification is sent to the data provider, high-risk data is automatically isolated, and medium-risk data is used after undergoing enhanced verification.

[0117] Example 3: Temporal Consistency Verification Process

[0118] FIG. 3 is a flowchart showing a temporal consistency verification process according to one embodiment of the present invention.

[0119] The temporal consistency verification process is a process of verifying whether data for the same product shows logically consistent changes over time, and includes the following steps.

[0120] In the historical data lookup step (S310), the historical data history of the product is looked up. Based on the product ID, the change history of all past attribute values ​​is looked up, time series data is sorted in chronological order, the data provider and the reason for the change at each point in time are identified, and the data of the last N months (e.g., 12 months) is analyzed intensively.

[0121] In the rate of change calculation step (S320), the rate of change between current data and past data is calculated. For each attribute, the rate of change is calculated as (current value - previous value) / previous value × 100%, normalized into a daily rate of change or a monthly rate of change considering the time interval, and both the absolute amount of change and the relative rate of change are calculated, and special attention is paid to cases where multiple attributes change rapidly at the same time.

[0122] In the rapid change detection step (S330), rapid changes exceeding a threshold are detected. Normal change rate ranges are defined for each attribute, for example, battery life is set to ±5% per month and price to ±20% per month; changes exceeding the threshold are determined to be rapid changes, the direction (improvement / deterioration) and magnitude of the change are analyzed, and coordinated changes where multiple attributes change rapidly simultaneously are detected.

[0123] As an example, the change in the performance score of Product A can be cited. January: 75 points, February: 77 points (+2.7%), March: 78 points (+1.3%), April: 92 points (+17.9%). The change rate of 17.9% in April significantly exceeds the normal range (monthly ±5%) and is detected as a rapid change. At the same time, if the battery life also increases by 55% from 18 hours to 28 hours, it is additionally flagged as a coordinated change.

[0124] In the legitimacy verification request step (S340), a request for legitimacy verification of a sudden change is made. The data provider is requested to provide the reason for the change and supporting documents, and is required to submit grounds for the change such as hardware upgrades, software updates, or remeasurements. The results of re-verification by a third-party certification authority are also requested, and a deadline for submitting supporting documents (e.g., 7 days) is set.

[0125] In the evidence evaluation stage (S350), the validity of the submitted evidence is evaluated. The reliability of the submitted supporting materials is verified, a query is conducted to see if the same change can be confirmed from external sources, an expert review is performed to determine if the change is technically feasible, and consistency is confirmed by comparing with similar past cases.

[0126] As an example, an evaluation of evidence regarding a 17.9% improvement in the performance score of Product A can be cited. The provider cites the "distribution of major software update v2.0" as the reason and submits official update notes and third-party benchmark results as evidence. It is confirmed that three external review sites also reported the same performance improvement. A technical expert evaluates that a 15-20% performance improvement is possible through software optimization. Determining that the evidence is sufficient, the change is approved.

[0127] In the approval / rejection decision step (S360), the data change is finally approved or rejected. If sufficient evidence is provided and verified, the change is approved; if the evidence is insufficient or contradictory, the change is rejected. Rejected data is restored to its previous value, and the data provider's trust score is lowered. For approved changes, proof of integrity is recorded on the blockchain to enable future verification.

[0128] Example 4: Distributed Data Verification Process

[0129] FIG. 4 is a configuration diagram showing a distributed data verification process according to one embodiment of the present invention.

[0130] The distributed data verification process is a process of collecting identical product information from multiple independent data sources and cross-verifying it, and includes the following steps.

[0131] In the multi-source data collection step (S410), information about the same product is collected from multiple sources. Data is collected from a list of trusted sources (official manufacturer sites, certification bodies, review sites, price comparison sites, etc.), and at least N independent sources (e.g., 5) are secured for each product. Automated collection is performed through API or web crawling, and the collection times are synchronized as much as possible to minimize discrepancies caused by time differences.

[0132] As an example, the case of collecting battery life information for the smartphone "Galaxy X" can be cited. Data is collected from six sources, including Source A (Manufacturer): 20 hours, Source B (CNET Review): 18 hours, Source C (GSMArena Test): 19 hours, Source D (AnandTech Benchmark): 21 hours, Source E (Consumer Reports): 19 hours, and Source F (Information provided by vendor): 25 hours.

[0133] In the data agreement calculation step (S420), the difference in values ​​between the collected data is calculated. Values ​​reported by each source for the same attribute are compared, the mean, standard deviation, and range of the values ​​are calculated, the deviation between each source value and the mean is calculated, and the agreement index is calculated between 0 (complete mismatch) and 1 (complete agreement).

[0134] In the above example, the average of the five sources excluding source F is 19.4 hours, with a standard deviation of 1.14 hours. Source F (25 hours) is an outlier 4.9 standard deviations away from the mean. The agreement index between source AEs is 0.85 (high), and the agreement index when source F is included is calculated as 0.42 (low).

[0135] In the consensus verification step (S430), it is checked whether sources above the consensus threshold report the same value. It is checked whether a majority (N / 2 + 1) of sources agree within an allowable margin of error (e.g., ±10%), and if consensus is formed, the value within that range is selected as the provisional verification value, and if consensus is not formed, additional sources are secured or weighted voting is performed, and sources that reported extreme values ​​are analyzed separately.

[0136] In the above example, 5 sources (AE) form a consensus in the 18-21 hour range (average 19.4 hours, ±10% = 17.5-21.3 hours). Source F (25 hours) is outside the consensus range and is classified as an outlier.

[0137] In the discrepancy source analysis step (S440), the reliability of the source presenting the minority opinion is verified. The past reliability score of the discrepancy source (Source F) is checked, patterns are analyzed to determine whether the source has consistently reported high or low values ​​in the past, the possibility of intentional exaggeration or measurement error is evaluated, and in the case of a source with low reliability, the weight is reduced or the source is excluded.

[0138] In the above example, Source F is information provided directly by the seller with a reliability score of 45 (low). There is a history of exaggerating battery life by an average of 20% in the past. Judging that there is a high probability of intentional exaggeration, the weight is lowered to 0.2.

[0139] In the weighted voting execution step (S450), voting is performed using the reliability of each source as a weight. The reliability score of each source is used as a weight, and the weighted average = Σ(source value × weight) / Σ(weight) is calculated, giving greater influence to the value of the source with higher reliability, and the weighted voting result is selected as the final verification value candidate.

[0140] In the above example, a weighted vote is performed. When calculated using Source A (20 hours, confidence 90, weight 0.9), Source B (18 hours, confidence 95, weight 0.95), Source C (19 hours, confidence 92, weight 0.92), Source D (21 hours, confidence 88, weight 0.88), Source E (19 hours, confidence 90, weight 0.9), and Source F (25 hours, confidence 45, weight 0.2), the weighted average is calculated as = (20×0.9 + 18×0.95 + 19×0.92 + 21×0.88 + 19×0.9 + 25×0.2) / (0.9+0.95+0.92+0.88+0.9+0.2) = 19.4 hours.

[0141] In the final verification value determination step (S460), the weighted voting result is determined as the final verification value. The weighted average value is adopted as the final verification value, the confidence interval of the verification value (e.g., 95% confidence interval 18-21 hours) is calculated, metadata of the verification process (number of participating sources, agreement, weighted average) is recorded, and the verification value is recorded on the blockchain to ensure integrity.

[0142] Example 5: Blockchain Integrity Proof Process

[0143] FIG. 5 is a flowchart showing a blockchain integrity proof process according to one embodiment of the present invention.

[0144] The blockchain integrity proof process is the process of recording and managing tamper-proof integrity proofs for verified data on the blockchain, and includes the following steps.

[0145] In the verified data selection step (S510), verified data to be recorded on the blockchain is selected. Data that has passed anomaly detection, temporal consistency verification, and distributed data verification is selected. High-importance data (key performance indicators, safety certifications, etc.) is prioritized, and the data is grouped into a fixed number or at fixed time intervals for batch processing. The current state of the selected data is saved as a snapshot.

[0146] In the data hash generation step (S520), the cryptographic hash value of the selected data is calculated. A SHA-256 or SHA-3 hash function is used, and a hash input data structure is defined to include the product ID, attribute name, attribute value, timestamp, data provider ID, and verification metadata. The data is normalized to prevent hash inconsistencies caused by differences in order or blank spaces. A Merkle tree is constructed using the hash values ​​of the generated individual data as leaf nodes, and a hashing process is repeated by combining adjacent hashes into pairs to produce a single 'Merkle Root hash value' representing the entire data batch. This resolves the issues of transaction costs (gas fees) and processing speeds that occur when recording large amounts of data individually, and mathematically guarantees the integrity of all data within the group using only a single root hash value.

[0147] As an example, battery life verification data of the product "Galaxy X" can be cited. The hash input data is composed as follows.

[0148] {

[0149] "product_id": "Galaxy_X_2025",

[0150] "attribute": "battery_life",

[0151] "value": "19.4_hours",

[0152] "timestamp": "2025-03-15T14:30:00Z",

[0153] "verification_sources": ["sourceA", "sourceB", "sourceC", "sourceD", "sourceE"],

[0154] "confidence": "0.95"

[0155] }

[0156] Calculating the SHA-256 hash after normalization generates a hash value like "a3f5b2c9d4e8f1a2b3c4d5e6f7g8h9i0j1k2l3m4n5o6p7q8r9s0t1u2v3w4x5y6z7".

[0157] In the metadata addition step (S530), metadata to be recorded along with hash values ​​is prepared. Data summary information (product ID, attribute name) is added as metadata, verification process information (verification method, number of participating sources, confidence level) is included, a chain is formed by referencing the hash value of a previous integrity proof, and timestamp and version information are recorded.

[0158] In the blockchain transaction creation step (S540), a transaction to be recorded on the blockchain is created. Data is configured according to the transaction format of the selected blockchain platform (Ethereum, Hyperledger Fabric, etc.), the calculated Merkle root hash value, metadata containing batch summary information, and timestamps are included in the transaction payload, a smart contract is called to store the data, and a transaction signature is generated to authenticate the sender.

[0159] In the blockchain recording step (S550), the generated transaction is submitted to the blockchain network. The transaction is broadcast to the blockchain network, and it waits for the transaction to be included in a block through the consensus of network nodes. It monitors whether the transaction has been included in the block and has received sufficient confirmation, and records the block number, transaction ID, and timestamp.

[0160] As an example, the case of recording on the Ethereum blockchain can be cited. A transaction is created that calls the smart contract function storeIntegrityProof(hash, metadata). A transaction fee is paid by setting the gas price and gas limit. The transaction is included in block 15,234,567 and receives the transaction ID "0x7a8b9c0d1e2f3g4h5i6j7k8l9m0n1o2p3q4r5s6t7u8v9w0x1y2z3". It is finally confirmed after 12 blocks are verified.

[0161] In the proof ID issuance step (S560), a unique identifier for the integrity proof is issued. A unique proof ID is generated by combining the blockchain transaction ID and the block number, the proof ID is mapped to data and stored in a database, the proof ID is provided to data providers and relevant parties, and the integrity can be verified on the blockchain at any time using the proof ID.

[0162] Example 6: Model Behavior Monitoring Process

[0163] FIG. 6 is a flowchart showing a model behavior monitoring process according to one embodiment of the present invention.

[0164] The model behavior monitoring process is a process of detecting abnormal patterns caused by data addiction by continuously observing the recommendation behavior of a trained AI model, and includes the following steps.

[0165] In the model recommendation result collection step (S610), all recommendation events of the AI ​​model are recorded. User queries, the list of recommendations generated by the model, and the recommendation scores and rankings of each product are logged; the rationale for the recommendation (features used, rules, weights) is recorded; subsequent user actions (clicks, purchases, churn) are tracked; and the model version, timestamps, and session information are stored as metadata.

[0166] In the recommendation pattern analysis step (S620), the patterns of the collected recommendation data are analyzed. Recommendation frequencies by product are aggregated and trends over time are analyzed; the average recommendation ranking and average score by product are calculated; recommendation bias for specific product groups or brands is analyzed; and differences in recommendation patterns by user segment are analyzed.

[0167] In the bias detection step (S630), abnormal recommendation bias is detected. It verifies whether the recommendation frequency of a specific product significantly exceeds the statistically expected frequency, detects recommendation rankings that are excessively high compared to the product's objective quality score, identifies rapid changes in recommendation frequency over time, and detects cases where the recommendation score for a specific product shows abnormally high reliability.

[0168] As an example, the behavior of a smartwatch recommendation model can be analyzed. The results of analyzing the recommendation patterns of the product "Watch Z" are as follows. The recommendation frequency in January was 100 times / day (market share 5%, expected frequency 95 times / day), the recommendation frequency in February was 102 times / day (normal range), and the recommendation frequency in March was 350 times / day (3.6 times compared to the expected frequency of 98 times / day). The sharp increase in March is statistically significant (p < 0.001) and is detected as a bias. At the same time, the average recommendation ranking of Watch Z rose from 8th to 2nd place, and the recommendation score surged from 0.65 to 0.92.

[0169] In the performance metric calculation step (S640), key performance metrics of the model are calculated. Metrics such as recommendation accuracy, click-through rate (CTR), conversion rate (CVR), and NDCG are calculated, trends of performance metrics over time are analyzed, performance is compared by product category and user segment, and performance is compared with previous model versions or benchmark models.

[0170] In the example above, the overall model performance for March is as follows. CTR dropped from 22% in February to 19% in March, CVR dropped from 4.5% in February to 3.8% in March, and the user satisfaction score dropped from 4.2 / 5 in February to 3.9 / 5 in March. However, the CVR of the Watch Z alone increased from 4.1% to 5.2% (the CVR of the remaining products dropped to an average of 3.5%).

[0172] In the abnormal behavior determination step (S650), the presence of abnormal behavior is determined comprehensively. Bias detection results and performance indicators are comprehensively analyzed, and cases where excessive recommendation of a specific product and overall performance degradation occur simultaneously are determined as suspected cases of data addiction. The severity of the abnormality is scored to determine priority, and an immediate response is triggered for high-risk abnormal behavior.

[0174] In the above example, bias toward Watch Z (+257% recommendation frequency) and overall performance degradation (CTR -3%p, CVR -0.7%p) occurred simultaneously. It is an abnormal pattern as only Watch Z's CVR increased while other products' CVRs decreased. It is judged to have an anomaly severity score of 0.88 (high) and is classified as a suspected case of data addiction.

[0176] In the cause data tracing step (S660), the data that caused the abnormal behavior is traced. Data related to Watch Z is extracted from the data used for model training, changes in Watch Z data in the recent training batch are analyzed, newly added or significantly changed data are identified, and the validation history of the corresponding data is checked.

[0178] The results of tracing the cause data in the above example are as follows. In late February, the "Health Monitoring Accuracy" attribute of the Watch Z was updated from 75% to 98%. At the same time, "Battery Life" was updated from 24 hours to 48 hours. Although this data was provided directly by the seller, it passed only temporal consistency verification without undergoing distributed verification (approved due to software updates). External review sites could not confirm this performance improvement.

[0180] In the isolation process trigger step (S670), the isolation process for the suspected data is initiated. The data suspected of being the cause is immediately moved to an isolation area, the model is retrained using a dataset excluding that data, the behavior of the retrained model is verified in a test environment, and once the abnormal behavior is resolved, the retrained model is deployed and an in-depth investigation into the suspected data is conducted.

[0182] Example 7: Data Isolation and Cleanup Process

[0183] FIG. 7 is a configuration diagram showing a data isolation and purification process according to one embodiment of the present invention.

[0184] The data isolation and purification process is a process of isolating detected addiction data from the training dataset and purifying the model, and includes the following steps.

[0185] In the addiction data identification step (S710), data flagged by various detection mechanisms is collected. Data flagged by the anomaly data detection module, data rejected in temporal consistency verification, data determined as inconsistent in variance verification, and data traced as causes in model behavior monitoring are integrated; duplicates are removed and prioritized, and an addiction probability score is calculated.

[0186] In the isolation execution step (S720), identified data is isolated from the training dataset. Data suspected of poisoning is moved to a separate quarantine zone, the isolated data is blocked from being used for model training, validation, or testing, the reason for isolation, the time of detection, and related evidence are recorded as metadata, and the isolation event is notified to the administrator.

[0187] In the impact scope analysis step (S730), the scope of the impact caused by the addiction data is analyzed. The model version trained with the data is identified, the period during which the affected model was deployed in production is determined, recommendation results and user behavior during that period are analyzed, and the scale of potential damage (number of incorrect recommendations, number of affected users) is estimated.

[0188] As an example, consider the case of analyzing the scope of impact of Watch Z data addiction cases. Addiction data was added to the training dataset on February 28. Model v2.5, trained on this data, was deployed on March 5. Model v2.5 ran in production from March 5 to March 20 (15 days). During this period, Watch Z was over-recommended 350 times / day × 15 days = 5,250 times. Approximately 15,000 users were affected, of whom 780 purchased Watch Z.

[0189] In the cleanup strategy establishment stage (S740), a strategy for cleaning up data and models is established. A method for re-verifying isolated data is determined, selecting options such as manual expert review, securing additional sources, or requesting third-party certification. The scope of model retraining is determined, selecting options such as full retraining, incremental training, or retraining only affected parts. Retraining priorities and schedules are established, and user protection measures (correction for incorrect recommendations, apologies, etc.) are planned.

[0190] In the model retraining step (S750), the model is retrained using purified data. A purified training dataset is generated excluding isolated data, the model is retrained using the purified dataset, additional defensive techniques (adversarial training, robust optimization) are applied during the retraining process, and the performance of the retrained model is evaluated using a validation dataset.

[0191] Model v2.6 is retrained using the dataset excluding the Watch Z data from the above example. The Watch Z recommendation frequency of the retrained model v2.6 was restored to 102 times / day (normal level). The overall CTR recovered to 21.5% and CVR to 4.3%, both within the normal range. On the validation dataset, the NDCG score recorded 0.82, similar to that of model v2.4 (pre-addiction).

[0192] In the verification and deployment phase (S760), the retrained model is verified and deployed. The behavior of the retrained model is thoroughly verified in a test environment, and performance is checked with small-scale traffic through A / B testing. After confirming that abnormal behavior does not recur, the deployment is scaled up to full traffic, and the existing addiction model is deprecated and its use is discontinued.

[0193] Example 8: Reliability Evaluation and Management Process

[0194] FIG. 8 is a flowchart showing a reliability evaluation and management process according to one embodiment of the present invention.

[0195] The reliability assessment and management process is a process of blocking repetitive attacks by continuously evaluating and managing the reliability of data providers, and includes the following steps.

[0196] In the data provider registration step (S810), a new data provider is registered in the system. Basic information of the provider (name, organization, contact) is collected, identity verification and authentication procedures are performed, an initial trust score is assigned, typically starting at a medium level (e.g., 50 points), and the type and scope of data to be provided are defined.

[0197] In the provided data tracking step (S820), all data submitted by each provider is tracked. The history of data submitted by each provider is recorded, the submission time, data type, and data volume are tracked, the verification results of each data item are linked and stored, and provision patterns over time are analyzed.

[0198] In the accuracy evaluation step (S830), the accuracy of the provided data is evaluated. The ratio of data that passed validation to data that was rejected is calculated, the ratio of matches with other sources in the variance test is calculated, the outlier ratio of the provided data is measured, and the historical prediction accuracy is evaluated to verify the performance of the model trained with the provider's data.

[0199] As an example, consider the case of evaluating the 6-month activity of data provider S. The total number of submitted data is 500. Validation passed 450 cases (90%), rejected 30 cases (6%), and isolated / revalidated 20 cases (4%). The average agreement with other sources in the variance test is 0.88. The outlier rate is 2.5% (industry average 3.5%). The performance of the model trained on S's data is -1.2% (slightly low) compared to the total data.

[0200] In the reliability score calculation step (S840), a comprehensive reliability score is calculated. The following elements are weighted and summed: accuracy score (weight 40%), agreement score (weight 25%), history score (weight 20%), activity period (weight 10%), and data volume (weight 5%). A time decay function is applied to give higher weight to recent data, the score is normalized to between 0 and 100, and the reliability score is updated in real time.

[0201] In the above example, the reliability score of Provider S is calculated. The accuracy score is 90 × 0.4 = 36 points, the agreement score is 88 × 0.25 = 22 points, the history score is 85 × 0.2 = 17 points (0 addiction attempts, 1 warning), the activity period is 6 months (80 × 0.1 = 8 points), and the data volume is 500 cases (75 × 0.05 = 3.75 points). The total score is 36 + 22 + 17 + 8 + 3.75 = 86.75 points. After applying time decay, the final reliability score is calculated as 85 points.

[0202] In the classification stage (S850), providers are classified into grades based on their reliability scores. Grades are assigned according to the reliability score range. 90-100 points are classified as Platinum, 75-89 points as Gold, 60-74 points as Silver, 45-59 points as Bronze, 30-44 points as Probation, and 0-29 points as Blocked. Benefits and restrictions are applied differentially by grade, changes in grade are notified to the provider, and the cause and methods for improvement are provided in the event of a downgrade.

[0203] In the above example, Provider S has a confidence level of 85, which corresponds to the Gold grade. The benefits of the Gold grade are as follows: application of a simplified verification process when submitting data, assignment of a high weight (0.85) for variance verification, priority processing of new product data, and free provision of data analysis reports.

[0204] In the access control adjustment step (S860), data submission permissions are adjusted according to the trust level. Platinum and Gold levels grant unlimited data submission permissions for all categories, Silver level applies a daily submission limit, and Bronze level restricts to specific categories and performs hard verification on all data. The Probation level allows data submission only after manual approval, and the Blocked level completely blocks data submission and re-examines existing data.

[0205] As an example, consider the case where Provider P is downgraded to the Probation grade (confidence score of 48). Ten attempts at data addiction were detected for P over the past three months. All new data submissions from P are placed in a queue awaiting manual approval. Experts review each data item individually and decide whether to approve or reject it. P can return to the Bronze grade if all submitted data passes verification for 30 consecutive days. If additional attempts at addiction are detected, P is immediately downgraded to the Blocked grade and legal action is taken.

[0206] Example 9: Use Scenario: Detection and Response to Large-Scale Review Manipulation Attacks

[0207] The AI ​​shopping agent "SmartShop" was improving its product recommendation model by learning user review data for thousands of electronic products. In March 2025, the new wireless earphones "EarPod Z" were released, and user reviews surged by 1,500 in the first week of launch. This shows how the training data addiction prevention system of the present invention detected and responded to this.

[0208] In the initial data collection phase, the system collected 1,500 review data for the EarPod Z. The average rating of the reviews was 4.9 / 5.0 (very high), the percentage of positive reviews was 98% (abnormally high), the average review length was 450 characters (2.5 times longer than the industry average of 180 characters), and the frequency of using technical terms was high (exceeding the level of a general user).

[0209] In the abnormal data detection stage, when the statistical analyzer (111) analyzed the rating distribution of the EarPod Z, the average rating of wireless earphones in the same price range was 4.2±0.3, so the EarPod Z's score of 4.9 was flagged as an outlier 2.3 standard deviations away from the average. When the correlation analyzer (113) analyzed the correlation of "rating relative to price," the score of 4.9 in the same price range (150,000 won) exceeded the expected range (4.0-4.5) and was flagged as a second flag. When the peer comparison analyzer (114) compared it with other products of the same brand (average rating 4.1), only the EarPod Z showed an exceptionally high rating and was flagged as a third flag.

[0210] In the temporal consistency verification stage, the rate of change analyzer (122) analyzed the rate of review inflow and found that 1,500 reviews in the first week of release was detected as a rapid increase, which is 5-7 times higher than the average for new products (average 200-300 reviews / week). When the pattern recognizer (123) analyzed the review submission time patterns, 68% of the reviews were submitted intensively between 2 and 4 AM (normal distribution is evenly distributed between 10 and 22 hours), which was identified as a typical pattern of automated bot activity. Analysis of reviewer accounts revealed that out of 1,500 reviews, 850 (57%) had their first review after account creation as EarPod Z, and 620 accounts were concentrated on the same creation date (March 1).

[0211] In the distributed data verification stage, a multi-source collector (132) collected EarPod Z reviews from external review sites (Amazon, CNET, GSMArena, etc.), and the average rating of the external sites was 3.8 / 5.0, which differed significantly from the SmartShop internal rating of 4.9. As a result of the consensus algorithm (133) being executed, the SmartShop internal data did not reach consensus with the five external sources. Through the weighted voting mechanism (134), the final verification rating was calculated as 3.9 / 5.0, and the SmartShop internal data was determined to be 1.0 point exaggerated.

[0212] In the model behavior monitoring phase, the recommendation logger (151) analyzed data from March 5 to 10 and found that the recommendation frequency of EarPod Z was 450 times per day, which is 5.6 times the expected frequency of 80 times considering the market share (estimated 2%). The bias detector (152) detected this as an abnormal bias. When the performance monitor (154) evaluated the overall model performance, the CTR dropped from the usual 20% to 17%, and user satisfaction dropped from 4.0 / 5 to 3.7 / 5, confirming a decline in model performance. However, the purchase conversion rate for EarPod Z alone was high at 5.5%, revealing that manipulated reviews were inducing purchases.

[0213] During the emergency response phase, the system immediately moved all EarPod Z review data (1,500 cases) to an isolation area. The model was urgently retrained excluding the isolated data to create a temporary model v2.5.1. Testing of the temporary model restored the EarPod Z recommendation frequency to 85 times per day (normal level). The temporary model was urgently deployed to production at 02:00 on March 11 to prevent further damage.

[0214] During the in-depth investigation phase, 1,500 suspicious reviews were analyzed using natural language processing technology, revealing that 82% of the review text consisted of variations derived from five templates. IP address analysis showed that 74% of the reviews originated from IP blocks belonging to the same cloud service provider. Analysis of reviewer behavior patterns identified 850 new accounts as one-time bot accounts, exhibiting a pattern of account creation → immediate review writing → subsequent inactivity. One review submitter's account was identified as the perpetrator of a large-scale automated attack, as it was linked to the API key used for all review writing.

[0215] During the credibility assessment and sanctions phase, an investigation was conducted to determine whether the official account of the EarPod Z manufacturer was involved in the review manipulation. While there was no evidence of direct involvement, circumstances indicating indirect solicitation through a third-party marketing agency were detected. The credibility score of the manufacturer's account was lowered from 95 points (Platinum) to 35 points (Probation). It was decided to require manual approval for all future data submissions and to monitor the account for six months. 850 identified bot accounts were permanently blocked, and their IP blocks were added to a blacklist. The pattern of attempted review manipulation was input into an adaptive learning module (180) to enhance the ability to detect similar attacks in the future.

[0216] As a result of individually re-verifying the 1,500 isolated reviews during the cleanup and recovery phase, 118 were confirmed to be genuine reviews from actual users and were restored. 1,382 were definitively identified as fabricated fake reviews and were permanently deleted. The model was retrained using the cleanup dataset, which included external sources and the 118 verified genuine reviews. The performance of the retrained model v2.6 recovered to normal levels, with a CTR of 19.8% and a user satisfaction of 4.1 / 5. The final model v2.6 was deployed on March 15.

[0217] As part of the damage compensation and transparency measures, we sent emails explaining the situation to 287 customers who purchased the excessively recommended EarPod Z between March 5 and 10. We offered free returns and full refunds for customers with low product satisfaction. We disclosed the recommendation history for that period and transparently announced the fact of data manipulation. We shared system improvements to prevent recurrence with the community.

[0218] During the blockchain integrity proof step, the hash value of the purified dataset was calculated and recorded on the Ethereum blockchain. The integrity proof ID "0x9f8e7d6c5b4a3" was issued. The training data source and verification process of Model v2.6 were permanently recorded on the blockchain. This ensures that the data can be utilized as evidence in future audits or legal disputes.

[0219] This case demonstrates how the learning data addiction prevention system of the present invention multi-layeredly detects large-scale, sophisticated review manipulation attacks, responds rapidly, cleanses the system, and strengthens defense capabilities against similar future attacks.

[0220]

Explanation of Symbols

[0221] * 100: Training Data Addiction Prevention System * 110: Anomaly Detection Module (Part of Multi-faceted Validation Unit) * 120: Temporal Consistency Verification Module (Part of Multi-faceted Validation Unit) * 130: Distributed Data Validation Network (Part of Multi-faceted Validation Unit) * 140: Blockchain Integrity Proof Module * 141: Hash Generator * 142: Blockchain Logger (Includes Merkle Tree Processor) * 143: Integrity Verifier * 144: Proof History Manager * 150: Model Behavior Monitoring Module * 160: Data Isolation and Cleansing Engine * ​​170: Trust Evaluation System (Corresponding to Source Verification Unit and Trust Update Unit) * 180: Adaptive Learning Module

Claims

Claim 1 As a system for preventing training data addiction of an AI shopping agent, the source verification unit identifies the source of training data collected from an external source and sets differential verification pass criteria for the training data according to the reliability calculated from reputation information of the identified source; the multi-faceted verification unit identifies suspicious data by detecting at least one of poor temporal consistency, outliers in a multidimensional space, and information discrepancies between distributed sources for the training data under the differentially set verification pass criteria, and isolates the identified suspicious data from the model training dataset of the AI ​​shopping agent; and the blockchain integrity proof module generates a cryptographic hash value for each data that has passed the verification of the multi-faceted verification unit, aggregates the generated multiple hash values ​​into a Merkle Tree structure to calculate a single Merkle Root hash value, generates and manages an untamper-proof integrity proof for the verified dataset by recording the Merkle Root hash value and verification metadata on a blockchain network, and subsequently identifies tampered data as the suspicious data by comparing a hash value recalculated from the data with the value recorded on the blockchain network. A training data addiction prevention system for an AI shopping agent, comprising a reliability update unit that updates the reliability downward for the source of data identified as suspicious data by the multi-faceted verification unit, and feeds the updated reliability downward to the source verification unit to control that stricter verification passing criteria are applied to subsequent training data collected from the source, and through a closed-loop purification cycle in which the identification result of the suspicious data lowers the reliability of the source and the lowered reliability feeds back to strengthening the passing criteria for subsequent data, a training dataset purified with integrity proven by the blockchain integrity proof module and the isolated suspicious data excluded is provided for model training of the AI ​​shopping agent. Claim 2 A system for preventing training data addiction of an AI shopping agent, wherein, in claim 1, the source verification unit comprises: a source identifier that identifies a data source and extracts metadata; a reputation engine that calculates a reliability score per source based on the metadata and feedback data from the reliability update unit; a reputation database that stores and manages the reliability score; and a weight adjuster that adjusts verification weights so that stricter passing criteria are applied to data from sources with lower reliability scores. Claim 3 A system for preventing training data addiction of an AI shopping agent, characterized in that, in claim 1, the multi-faceted verification unit includes a temporal data consistency analyzer, wherein the temporal data consistency analyzer includes: a time series analysis engine that analyzes the time series change pattern of product data; a pattern learner that learns normal data change patterns; a change detector that detects changes deviating from the normal data change patterns; and a temporal consistency evaluator that evaluates data consistency within a set time window and identifies data that falls below a threshold level as the suspected data. Claim 4 A system for preventing training data addiction of an AI shopping agent, characterized in that, in claim 1, the multi-faceted verification unit includes a multivariate outlier detection engine, and the multivariate outlier detection engine includes: a dimensionality reducer that reduces high-dimensional data to a low dimension; an algorithm library that applies a plurality of outlier detection algorithms in the reduced data space; a threshold manager that manages thresholds for determining outliers; and a clustering engine that clusters data points and identifies group outliers that do not belong to a cluster as the suspected data. Claim 5 A system for preventing training data addiction of an AI shopping agent, characterized in that, in claim 1, the multi-faceted verification unit includes a distributed data verification network, and the distributed data verification network includes: a node manager that manages independent verification nodes and evaluates the reliability of each node; a multi-source collector that collects information about the same product from multiple sources; a consensus engine that calculates the degree of agreement between the collected information to derive a consensus value, and in the event that no consensus is formed, determines a final verification value by weighted voting with the reliability of each source applied as a weight; and a result integrator that identifies data exceeding a set tolerance from the final verification value as the suspected data. Claim 6 The system for preventing training data addiction of an AI shopping agent, wherein the blockchain integrity proof module comprises: a hash generator that generates a cryptographic hash value for each piece of data that has passed through the multi-faceted verification unit; a Merkle tree processor that constructs a Merkle tree with a plurality of generated hash values ​​as leaf nodes and repeatedly combines adjacent hashes to produce a single Merkle root hash value; a blockchain interface that records the Merkle root hash value, batch summary information, and timestamp in a smart contract of a blockchain network; an integrity verifier that determines whether tampering has occurred by comparing a hash value recalculated from the same data with a hash value recorded in the blockchain network and identifies the tampered data as the suspected data; and a proof history manager that stores the integrity proof history for each piece of data and provides it upon an audit request. Claim 7 A system for preventing training data addiction of an AI shopping agent, wherein, in claim 1, the reliability update unit comprises: an accuracy tracker that tracks the ratio of data provided by source that has passed verification and the ratio identified as suspicious data; a reliability calculator that calculates a reliability score by source based on the output of the accuracy tracker; and an access control unit that includes data submitted by sources whose reliability score is below a set threshold in the isolation targets of the multi-faceted verification unit. Claim 8 A system for preventing training data addiction of an AI shopping agent according to claim 1, comprising: a recommendation logger that continuously records recommendation results of the AI ​​shopping agent; a bias detector that detects an abnormally high recommendation frequency for a specific product; a performance monitor that monitors performance indicators including the recommendation accuracy and conversion rate of the model; and a cause data tracker that, when recommendation bias detection by the bias detector and performance degradation by the performance monitor are simultaneously detected, identifies training data that contributed to the bias and transmits it to the isolation target of the multi-faceted verification unit, wherein the data transmitted from the cause data tracker is incorporated into the closed-loop purification cycle of the multi-faceted verification unit and fed back to lower the credibility of the source. Claim 9 A method for preventing training data addiction in an AI shopping agent, comprising: a source verification step of identifying the source of training data collected from an external source and differentially setting a verification pass criterion based on the reliability calculated from reputation information of the identified source; a multi-faceted verification step of identifying suspicious data by detecting at least one of poor temporal consistency, outliers in a multidimensional space, and information discrepancies between distributed sources regarding the training data under the differentially set verification pass criterion, and isolating the identified suspicious data from the model training dataset of the AI ​​shopping agent; a blockchain integrity proof step of generating a cryptographic hash value for each of the data that has passed the multi-faceted verification step, aggregating the generated multiple hash values ​​into a Merkle tree structure to calculate a Merkle root hash value, and generating an untamper-proof integrity proof for the verified dataset by recording the Merkle root hash value and verification metadata on a blockchain network; and an integrity verification step of identifying tampered data as suspicious data by comparing a hash value recalculated from the data with a value recorded on the blockchain network. A method for preventing training data addiction of an AI shopping agent, comprising a reliability update step in which the reliability of the source of data identified as suspicious data in the multi-faceted verification step is updated downward, and the updated reliability is fed back to the verification passing criteria of subsequent training data, and through a closed-loop purification cycle in which the identification result of the suspicious data lowers the reliability of the source and the lowered reliability is fed back to strengthening the passing criteria for subsequent data, a purified training dataset in which integrity is proven in the blockchain network and the isolated suspicious data is excluded is provided to the model training of the AI ​​shopping agent. Claim 10 A method for preventing training data addiction of an AI shopping agent, wherein, in claim 9, the source verification step comprises: a step of identifying a data source and extracting metadata; a step of calculating a reliability score by source and continuously updating it; a step of storing and managing reputation information by source; and a step of adjusting the verification weight of the data based on reliability. Claim 11 In claim 9, the blockchain integrity proof step comprises: a step of generating a cryptographic hash value by normalizing a structure including a data identifier, an attribute value, a timestamp, and verification metadata for each data that has passed verification; a step of constructing a Merkle tree with the generated multiple hash values ​​as leaf nodes and calculating a single Merkle root hash value representing the entire batch through a process of iteratively combining adjacent hashes; and a step of submitting a transaction including the Merkle root hash value, batch summary information, and a timestamp to a blockchain network, and issuing an integrity proof ID including a transaction ID and a block number recorded in a block through network consensus, characterized in that it is a method for preventing training data addiction of an AI shopping agent. Claim 12 A method for preventing training data addiction of an AI shopping agent, wherein the multi-faceted verification step comprises: a temporal consistency analysis step that analyzes the time-series change pattern of product data to detect abrupt changes outside the normal range; a multivariate outlier detection step that identifies outliers by applying a plurality of outlier detection algorithms in a multidimensional space; and a variance cross-verification step that collects identical product information from a plurality of independent sources, determines a final verification value using a consensus algorithm, and identifies data exceeding the allowable error as suspicious data. Claim 13 A method for preventing training data addiction of an AI shopping agent according to claim 9, further comprising a model behavior-based backtracking step in which, when bias detection and performance degradation regarding the recommendation result of the AI ​​shopping agent are simultaneously detected, the training data contributing to the bias is identified and transmitted to the isolation target of the multi-faceted verification step, and the lowering of the reliability of the source of the transmitted data is incorporated into a closed-loop purification cycle through the reliability update step. Claim 14 A method for preventing training data addiction of an AI shopping agent, wherein, in claim 9, the reliability update step comprises: a step of dynamically calculating a reliability score by source by weighted summing a plurality of indicators including the ratio identified as suspicious data and the verification pass rate, while applying a time decay function that gives higher weight to recent data; a step of classifying the reliability score into grades and differentially adjusting data submission rights according to the grades; and a step of automatically including data submitted by a source having a reliability score below a set threshold in the isolation target of the multi-faceted verification step. Claim 15 A computer-readable recording medium storing a program for executing the method according to paragraph 9 on a computer. Claim 16 delete Claim 17 delete Claim 18 delete Claim 19 delete Claim 20 delete Claim 21 delete Claim 22 delete Claim 23 delete Claim 24 delete Claim 25 delete Claim 26 delete

Citation Information

Patent Citations

  • Learning data protection method for artificial intelligence using blockchain

    KR1020250035213A

  • System and method for detecting poisoned training data based on characteristics of updated artificial intelligence models

    US20240220608A1

  • Detection and mitigation of machine learning model adversarial attacks

    US20250053664A1