Live broadcast e-commerce compliance auditing method and system based on penetration type supervision

Abstract

The invention discloses a live broadcast e-commerce compliance auditing method and system based on penetration type supervision, and relates to the technical field of e-commerce compliance supervision, and the method comprises the steps: carrying out the Hash abstract calculation of a judgment result data set and a multi-source data set, carrying out the partitioning storage of the data set in a block chain, verifying the storage process through a timestamp and a multi-node consensus protocol, and carrying out the verification of the verification result; forming an evidence storage data chain; automatically matching the evidence storage data chain with a compliance rule, identifying potential illegal behaviors and risk levels, and forming a compliance result set; performing fusion calculation on the compliance result set, historical violation records and user feedback data to generate a comprehensive risk score, and automatically triggering real-time compliance intervention measures according to the comprehensive risk score to form an intervention result set; and re-storing the intervention result set, the comprehensive risk score and the violation behavior in the block chain to form a closed-loop evidence storage data chain. According to the invention, supervision transparency is enhanced, data integrity is ensured, and reliable support is provided for real-time intervention and closed-loop auditing.

Description

Technical Field

[0001] This invention relates to the field of e-commerce compliance supervision technology, and in particular to a method and system for compliance auditing of live-streaming e-commerce based on penetrating supervision. Background Technology

[0002] With the rapid development of internet technology, e-commerce models are gradually shifting from text and image displays to real-time interaction, with emerging formats such as live-streaming e-commerce experiencing explosive growth globally. Through real-time interaction between hosts and users, consumers can complete the entire process of understanding products, asking questions, and making immediate purchases in a short time, significantly shortening the decision-making chain in traditional e-commerce models. In recent years, artificial intelligence technologies such as computer vision, natural language processing, and big data analytics have been introduced into e-commerce compliance management to identify non-compliant behaviors such as false advertising, illegal advertising, and the sale of counterfeit goods. Meanwhile, blockchain technology, due to its distributed storage, immutability, and traceability characteristics, is also increasingly being applied to the trusted storage and transaction supervision aspects of e-commerce.

[0003] However, in the rapid development of live-streaming e-commerce, existing technologies still face certain challenges in compliance verification. On the one hand, current compliance verification methods mostly focus on single-dimensional data detection, such as filtering based on text keywords or verifying the authenticity of goods based on image recognition, lacking deep integration between multi-source data, making it difficult to form a holistic compliance judgment in complex live-streaming scenarios. On the other hand, existing evidence storage technologies usually remain at the level of storing single events or transaction records on the blockchain, and have not yet formed a dynamic, closed-loop real-time evidence storage chain, resulting in limitations in compliance intervention, risk feedback, and traceability due to insufficient coordination. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a live-streaming e-commerce compliance audit method based on penetrating supervision to solve the problems of real-time compliance determination and traceable evidence preservation throughout the entire live-streaming process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] Firstly, this invention provides a method for compliance review of live-streaming e-commerce based on penetrating supervision, which includes:

[0008] Initial data is collected and preprocessed to form a multi-source dataset;

[0009] A deep convolutional neural network is used to automatically compare the appearance features, brand logo features, and packaging features of product images to form a judgment result dataset.

[0010] The judgment result dataset and the multi-source dataset are hash digested, stored in blocks on the blockchain, and the storage process is verified through timestamps and multi-node consensus protocols to form a chain of evidence storage data.

[0011] The evidence storage data chain is automatically matched with compliance rules to identify potential violations and risk levels, forming a compliance result set;

[0012] The system integrates and calculates compliance result sets, historical violation records, and user feedback data to generate a comprehensive risk score. Based on the comprehensive risk score, it automatically triggers real-time compliance intervention measures, forming an intervention result set.

[0013] The intervention results set, comprehensive risk score, and violations are re-stored on the blockchain to form a closed-loop evidence storage data chain.

[0014] As a preferred embodiment of the live-streaming e-commerce compliance review method based on penetrating supervision described in this invention, the specific steps for forming the multi-source dataset are as follows:

[0015] Collect the video stream, audio stream, text comments, product images, and user behavior of the host during the live broadcast, perform timestamp alignment and event-driven sampling, and generate initial data;

[0016] The system extracts keyframes and generates perceptual fingerprints from video streams, separates sound sources and transcribs semantics from audio streams, desensitizes and semantically annotates text bullet comments, performs region detection and OCR parsing on product images, and models click sequences and paths of user behavior, fusing them together to form feature-based data.

[0017] The feature-rich data undergoes cross-modal temporal fusion, quality enhancement, and privacy tokenization to generate a multi-source dataset.

[0018] As a preferred embodiment of the live-streaming e-commerce compliance review method based on penetrating supervision described in this invention, the specific steps for forming the judgment result dataset are as follows:

[0019] The product image is input into a deep convolutional neural network to extract appearance features, brand logo features and packaging features, and then compared one by one with the product feature vectors extracted and stored in a standard product image library to obtain a comparison result set.

[0020] The results set is compared to determine its authenticity and confidence level, forming a judgment result dataset.

[0021] As a preferred embodiment of the live-streaming e-commerce compliance review method based on penetrating supervision described in this invention, the specific steps for forming the evidence storage data chain are as follows:

[0022] The judgment result dataset and the multi-source dataset are formatted and deduplicated to generate a data package to be stored.

[0023] The data packets to be stored are hash digested to generate a unique digest, and a data block index list is established to generate hash index data.

[0024] The data packet to be stored is divided into several data blocks based on the hash index data, and a timestamp and digest verification are added to each block to generate a set of data blocks to be stored on the chain.

[0025] Submit the set of data blocks to be stored on the blockchain to initiate a write transaction. After multi-node consensus verification, the block height and transaction identifier are returned, and an on-chain confirmation record is generated.

[0026] Based on the on-chain confirmation records, a mapping is established between the evidence storage data block summary and the on-chain identifier, and they are linked together in chronological order to form an evidence storage data chain.

[0027] As a preferred embodiment of the live-streaming e-commerce compliance review method based on penetrating supervision described in this invention, the compliance rules are obtained by sorting out normative documents, industry standards, platform management specifications and historical violation cases, classifying, coding and itemizing the prohibited clauses, restrictive requirements and review points involved, and then reviewing and revising them.

[0028] As a preferred embodiment of the live-streaming e-commerce compliance review method based on penetrating supervision described in this invention, the specific steps for forming the compliance result set are as follows:

[0029] The evidence storage data chain is broken down and organized according to time sequence and business objects to obtain a matching dataset;

[0030] The matching dataset is compared with each compliance rule, and the existing violations and their corresponding risk levels are marked to obtain the marked result set;

[0031] The tagged result set is summarized and structured to generate a compliance result set.

[0032] As a preferred embodiment of the live-streaming e-commerce compliance review method based on penetrating supervision described in this invention, the historical violation records and user feedback data are obtained by auditing and archiving the platform's completed live-streaming records and transaction records, summarizing and structuring user complaints, evaluations, reports and customer service handling results during the live-streaming process, and then reviewing and classifying them.

[0033] As a preferred embodiment of the live-streaming e-commerce compliance review method based on penetrating supervision described in this invention, the specific steps for forming the intervention result set are as follows:

[0034] The compliance result set, historical violation records, and user feedback data are summarized and organized to obtain a unified multi-source data set;

[0035] The unified multi-source data set is associated and classified according to live streaming behavior and time sequence to generate structured data tables;

[0036] Multidimensional data fusion and compliance rule comparison are performed on structured data tables to calculate the comprehensive risk factor value for each live streaming behavior and generate a comprehensive risk score;

[0037] Based on the comprehensive risk score, corresponding real-time compliance intervention measures are triggered, and the operational behaviors, modification status, warnings and restrictions during the execution process are summarized and organized to form an intervention result set.

[0038] As a preferred embodiment of the live-streaming e-commerce compliance review method based on penetrating supervision described in this invention, the specific steps for forming a closed-loop evidence storage data chain are as follows:

[0039] The intervention result set, comprehensive risk score, and marked violations are summarized and organized to generate a sorted data set;

[0040] The organized data set is formatted and structured to generate blockchain storage and evidence-keeping data blocks;

[0041] Hash the blockchain-stored evidence data blocks and add timestamps to form verifiable block data, then write it into the blockchain to form a closed-loop evidence data chain.

[0042] Secondly, this invention provides a compliance review system for live-streaming e-commerce based on penetrating supervision, including:

[0043] The data acquisition module is used to collect initial data and preprocess it to form a multi-source dataset;

[0044] The feature comparison module is used to automatically compare the appearance features, brand logo features, and packaging features of product images using a deep convolutional neural network to form a judgment result dataset.

[0045] The evidence storage and chain building module is used to store the judgment result dataset and multi-source datasets in a unified manner on the blockchain to form an evidence storage data chain;

[0046] The rule matching module is used to automatically match the evidence storage data chain with compliance rules, identify potential violations and risk levels, and form a compliance result set.

[0047] The risk assessment module is used to integrate and calculate compliance result sets, historical violation records, and user feedback data to generate a comprehensive risk score. Based on the comprehensive risk score, it automatically triggers real-time compliance intervention measures to form an intervention result set.

[0048] The closed-loop evidence storage module is used to re-store the intervention result set, comprehensive risk score and violation behavior on the blockchain to form a closed-loop evidence storage data chain.

[0049] The beneficial effects of this invention are as follows: By calculating the hash digest of the judgment result and multi-source data, a unique identifier is generated and segmented into block-based data. A timestamp is added when the data is uploaded to the blockchain, and verification is performed through a multi-node consensus protocol. This achieves verifiability and tamper-proofness of the storage process, thus forming an immutable and traceable evidence storage data chain. This provides a secure and reliable evidence storage foundation for the entire live streaming process. Simultaneously, binding each data action to its time sequence provides a clear traceability basis for subsequent compliance rule matching and risk scoring, enhancing regulatory transparency and ensuring data integrity, and providing reliable support for real-time intervention and closed-loop auditing. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of a compliance review method for live-streaming e-commerce based on penetrating supervision.

[0052] Figure 2 This is a schematic diagram of a compliance review system for live-streaming e-commerce based on penetrating supervision.

[0053] Figure 3 A flowchart for the formation of a data chain for evidence storage.

[0054] Figure 4 This is a flowchart for the formation of the intervention outcome set. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for compliance review of live-streaming e-commerce based on penetrating supervision, including the following steps:

[0059] S1. Collect initial data and preprocess it to form a multi-source dataset.

[0060] S1.1 Collect the video stream, audio stream, text comments, product images, and user behavior of the host during the live broadcast, perform timestamp alignment and event-driven sampling, and generate initial data.

[0061] Specifically, to generate initial data, the video stream, audio stream, text comments, product images, and user behavior of the anchor during the live broadcast are collected in parallel and processed in sequence: the video stream is subjected to frame rate unification, key frame extraction and perception fingerprint calculation, and the key frame sequence is output.

[0062] The keyframe sequence is time-aligned with the speech stream. After speech activity detection, sound source separation and noise reduction, the speech stream is semantically transcribed and output as a time-aligned speech transcript.

[0063] Perform deduplication, anonymization, and semantic annotation on the text bullet comments, and output a structured bullet comment record;

[0064] Perform region detection, barcode / QR code recognition, and OCR parsing on product images, and output product image metadata;

[0065] Collect click, dwell time, add to cart, order placement and refund events and construct a behavioral time series, then output the behavioral event sequence;

[0066] The keyframe sequence, time-aligned speech transcription, structured bullet screen recording, product image metadata, and behavioral event sequence are aligned by timestamp. Based on event-driven sampling, higher resolution data is extracted at key moments such as product display, order placement, or abnormal interaction (e.g., sampling 5 seconds before and after product display or order placement). The data is then aggregated, formatted, and used to generate the initial data.

[0067] S1.2. Extract keyframes and generate perceptual fingerprints from the video stream, separate the sound source and transcribe the speech stream, desensitize and semantically annotate the text bullet comments, perform region detection and OCR parsing on the product images, and model the click sequence and path of user behavior, and fuse them to form feature data.

[0068] Specifically, the video stream is subjected to frame rate normalization, and key frames are extracted using compressed domain I-frame recognition or key frame extraction based on inter-frame difference (including I-frames, significant scene switching frames, and frames with obvious inter-frame differences in the video stream) to obtain a key frame sequence. Then, a perceptual fingerprint generation algorithm (such as pHash) is applied to the key frame sequence to extract similarity features and obtain a perceptual fingerprint.

[0069] The speech stream is processed by speech activity detection, blind source separation and noise reduction to distinguish the broadcaster's audio source from background noise, and speech recognition is performed on the separated broadcaster audio to obtain semantic transcription text.

[0070] Simultaneously, the text bullet comments are deduplicated and desensitized, unified character encoding is performed, word segmentation and named entity recognition are performed to complete semantic annotation and output structured bullet comment records;

[0071] First, perform region detection on the product image to locate the display area and extract high-resolution sub-images. Then, perform OCR parsing and barcode / QR code recognition on the high-resolution sub-images to obtain the region information and recognition text of the product image.

[0072] Collect user behavior data such as clicks, dwell time, add-to-cart, order placement, and refund events, and construct click sequences and path models according to time series to obtain a behavior path table;

[0073] The keyframe sequence, perceptual fingerprint set, semantic transcription text, structured bullet screen records, image region information of the product, and recognition text, as well as the behavior path table, are aligned by timestamps and extracted at the trigger time point based on event-driven sampling, and then fused to generate feature data.

[0074] S1.3 Perform cross-modal temporal fusion, quality enhancement, and privacy tokenization on the feature-enhanced data to generate a multi-source dataset.

[0075] Specifically, the feature data is synchronized and resampled according to timestamps, and linear interpolation is used to align multimodal features with time deviations and output a temporally aligned feature stream as input for cross-modal temporal fusion.

[0076] Cross-modal temporal fusion performs feature standardization, feature concatenation, and weighted aggregation on cross-modal temporally aligned feature streams, and uses a sliding window or event window aggregation strategy to generate a fused feature sequence in the temporal dimension as input for quality enhancement;

[0077] The quality enhancement performs integrity checks and noise suppression on the fused feature sequence, including signal-to-noise ratio evaluation, linear interpolation / temporal reconstruction for missing frames or words, outlier filtering, and segmented resampling. After generating quality labels for each segment, the quality enhancement data is output as input for privacy tokenization processing.

[0078] Privacy tokenization processes irreversibly hash (e.g., SHA-256) or pseudo-anonymize user identifiers, fingerprints, and sensitive fields in quality-enhancing data to generate revocable tokens. Simultaneously, it minimizes field retention and indexes the data, outputting a multi-source dataset with privacy-maintaining annotations.

[0079] S2. Utilize a deep convolutional neural network to automatically compare the appearance features, brand logo features, and packaging features of product images to form a judgment result dataset.

[0080] S2.1 Input the product image into a deep convolutional neural network to extract appearance features, brand logo features and packaging features, and compare them one by one with the product feature vectors extracted and stored in the standard product image library to obtain a comparison result set.

[0081] Specifically, the product images used for training are preprocessed, including scaling by resolution, normalizing color channels, and standardizing pixel values, to obtain the training input images;

[0082] The deep convolutional neural network is trained by inputting the input image into the deep convolutional neural network, extracting features through forward propagation and combining them with product category labels for backpropagation training, iteratively updating the parameters of the deep convolutional neural network until convergence, thus completing the training of the deep convolutional neural network.

[0083] In the application phase, the same preprocessing operation is performed on the product images to be compared to obtain preprocessed images;

[0084] The preprocessed image is input into the trained deep convolutional neural network to extract appearance feature vectors, brand logo feature vectors, and packaging feature vectors, and output product feature vectors.

[0085] Read the pre-extracted and stored product feature vectors from the standard product image library, and calculate the similarity value between each product feature vector and the standard product feature vector. The cosine similarity calculation is used, and the expression is:

[0086]

[0087] Where S(A,B) represents the similarity value between product feature vector A and standard product feature vector B, where A represents the product feature vector and B represents the standard product feature vector. i Let B represent the component value of the product feature vector A in the i-th dimension. iLet n represent the component value of the standard commodity feature vector B in the i-th dimension, where n represents the number of dimensions and i represents the dimension index.

[0088] The calculated similarity values ​​are then output as similarity records.

[0089] Based on similarity records, the results are categorized and sorted according to comparison rules to generate a comparison result set.

[0090] It should also be noted that the sources of the standard product image library include authoritative product images provided by brand owners, manufacturers, and e-commerce platforms, as well as images obtained by screening and verifying compliant product images from certified historical transaction data. During the establishment process, the collected original product images undergo uniform resolution conversion, background interference removal, and color correction to ensure consistent image quality. Each product image is labeled with its product category, brand logo, and packaging features. After verification, these are stored as standard product feature data, thus forming the basic data source and basis for the establishment of the standard product image library.

[0091] S2.2. Compare the results set to determine its authenticity and assess its confidence level, forming a judgment result dataset.

[0092] Specifically, the system receives the comparison result set and performs normalization processing on the similarity records, OCR parsed text, and barcode / QR code matching results in the comparison record, and outputs standardized comparison records.

[0093] Perform consistency judgment on each standardized comparison record. If the barcode / QR code matching result is consistent, output a preliminary judgment as true and save the corresponding evidence. If the barcode / QR code matching result is inconsistent, output a preliminary judgment and save the corresponding evidence based on the majority consistency of the appearance feature similarity, brand logo feature comparison result and packaging feature comparison result, and output a preliminary judgment record.

[0094] The preliminary judgment records are combined with the historical matching frequency in the standard commodity image library and the similarity information at the time of comparison. The similarity, OCR consistency, barcode / QR code matching and historical matching frequency are weighted and summarized and normalized to generate a confidence score. The output is a judgment result dataset containing the authenticity judgment, confidence score and judgment evidence.

[0095] Preferably, deep convolutional neural networks are applied to product image recognition and automatic comparison in live e-commerce scenarios. By simultaneously extracting the appearance features, brand logo features and packaging features of the product, and performing a fine comparison with the feature vectors in the standard product image library, the authenticity of the product and confidence score can be achieved.

[0096] Compared with existing technologies that rely on manual review, template matching, or histogram comparison, traditional image matching algorithms can improve the accuracy and efficiency of recognition, while supporting large-scale, real-time processing and reducing human error and delays.

[0097] It achieves automated and high-precision product authenticity identification, providing a reliable and quantifiable data foundation for subsequent compliance audits. This ensures that product information during live-stream e-commerce transactions is verifiable and trustworthy, thereby enhancing the platform's regulatory capabilities and user trust, and creating a penetrating regulatory effect that existing technologies cannot directly achieve.

[0098] S3. Calculate the hash digest of the judgment result dataset and the multi-source dataset, store them in blocks on the blockchain, and verify the storage process through timestamps and multi-node consensus protocols to form a chain of evidence-based data.

[0099] S3.1. Perform format unification and deduplication verification on the judgment result dataset and the multi-source dataset to generate a data package to be stored.

[0100] Specifically, after receiving the judgment result dataset and the multi-source dataset, the field names, data types and timestamps are normalized according to the unified data format specifications, and normalized records are output.

[0101] Standardize the units and codes of the records and output verifiable records;

[0102] For verifiable records, generate a deduplication key based on the hash value concatenated from the fields, remove duplicates and merge conflicting entries, and output a unique record set;

[0103] The unique record set is checked for integrity and constraints. After removing unqualified records, it is encapsulated with metadata to generate a data package to be stored.

[0104] S3.2 Calculate the hash digest of the data packet to be stored, generate a unique digest, and establish a data block index list to generate hash index data.

[0105] Specifically, the data packets to be stored are first processed by normalization and serialization, and the fields are arranged in order, such as by product identifier, session identifier, timestamp, source type, authenticity judgment, confidence score and event type, and then converted into a continuous byte stream using UTF-8 encoding.

[0106] Perform hash digest calculation on the byte stream, using the SHA-256 algorithm to calculate a unique digest value for each record or the entire byte stream;

[0107] After obtaining a unique digest value, the byte stream is divided into blocks according to the digest value and a preset block division strategy. For example, the blocks are divided into blocks with each data block size of one megabyte or with each thousand records as a block. Multiple data blocks are generated, and a block-level hash value is calculated for each data block. The data blocks and the corresponding hash value set are then output.

[0108] A data block index list is established based on the set of evidence storage data blocks and the corresponding set of hash values. Each record in the index list includes the block sequence number, the starting byte offset position, the ending byte offset position, the block-level hash value, the range of records included, and the generated timestamp, forming an initial index table.

[0109] The initial index table is digitally signed, for example, by using an elliptic curve digital signature algorithm to generate a digital signature, and then metadata such as version number and generation time are appended to the index table to form hash index data.

[0110] It should also be noted that the preset block partitioning strategy is based on the needs of data storage efficiency and verification operability. On the one hand, it avoids query delays caused by excessively large single blocks, taking into account the differences in data scale such as video streams, audio streams, product images and user behavior in the live e-commerce scenario. On the other hand, it refers to common blockchain storage principles, such as partitioning by fixed size or fixed number of records, to balance write performance and verification efficiency, and meet the real-time requirements of compliance review for live e-commerce.

[0111] S3.3. Divide the data packet to be stored into several data blocks based on the hash index data, and add a timestamp and digest verification to each block to generate a set of data blocks to be uploaded to the blockchain.

[0112] Specifically, based on the block sequence number and start and end byte offset recorded in the hash index data, the data packet to be stored is segmented byte by byte, and the original set of data blocks for storage is output.

[0113] Calculate the SHA-256 hash for each evidence data block in the original evidence data block set, and attach the calculated hash value as the digest verification value to the corresponding evidence data block, outputting a block record set with digest verification;

[0114] Add timestamp information (e.g., using UTC time zone and represented in ISO-8601 format) to each block in the block record set with digest verification, and write the timestamp and the corresponding digest verification value into the block element information, and output the set of data blocks to be uploaded to the chain for evidence storage with timestamps;

[0115] Consistency verification is performed on the set of data blocks to be uploaded to the blockchain with timestamps. The verification is completed by comparing the digest verification value in the block element information with the hash value in the hash index data. Data blocks that pass the verification are marked as eligible to be uploaded to the blockchain and the set of data blocks to be uploaded to the blockchain is output.

[0116] S3.4 Submit the set of data blocks to be uploaded to the blockchain network to initiate a write transaction. After multi-node consensus verification, the block height and transaction identifier are returned, and an on-chain confirmation record is generated.

[0117] Specifically, the set of data blocks to be uploaded to the blockchain for evidence storage is packaged according to the transaction format specified by the blockchain network to generate a write transaction request, and then broadcast to the blockchain node cluster through network nodes;

[0118] After receiving a write transaction request, each blockchain node performs transaction verification according to the blockchain consensus protocol, including verifying the consistency of the hash, timestamp, block sequence number, and hash of the previous block of the evidence storage data block, and outputs the verification result.

[0119] After the legality of the write transaction is confirmed by a majority of nodes' consensus vote, the transaction is packaged into a new block, and the block height and transaction identifier of the new block are generated.

[0120] The block height and transaction identifier are mapped to the set of data blocks to be uploaded to the blockchain for evidence storage, forming an on-chain confirmation record.

[0121] S3.5. Based on the on-chain confirmation record, establish a mapping between the evidence storage data block summary and the on-chain identifier, and connect them in chronological order to form an evidence storage data chain.

[0122] Specifically, based on the on-chain confirmation records, the hash digest of each evidence storage data block and the corresponding blockchain block height and transaction identifier are extracted, and a mapping table is established between the hash digest and the on-chain identifier.

[0123] Sort the data blocks in the mapping table in ascending order according to their timestamp information, and arrange them in chronological order.

[0124] The mapping tables arranged in chronological order are connected in a chain structure, where the hash digest of each evidence storage data block points to the hash digest of the previous evidence storage data block, and the corresponding block height and transaction identifier are appended.

[0125] The integrity of the chain connection is verified to ensure that the hash digest of each evidence data block matches the on-chain identifier without omission, thus generating the evidence data chain.

[0126] Preferably, after performing hash digest calculation on the judgment result dataset and the multi-source dataset, the data is stored in blocks on the blockchain, and the storage process is verified by combining timestamps and multi-node consensus protocols, thereby forming a traceable and tamper-proof evidence data chain.

[0127] Compared to existing technologies that rely solely on centralized databases or simple encrypted storage, this approach utilizes blockchain's block hashing and multi-node consensus mechanism to achieve data integrity, time traceability, and multi-party verification capabilities, avoiding the risk of single-point tampering. It ensures the credibility of live-streaming e-commerce transaction and audit data during storage and transfer, providing reliable data protection for subsequent compliance rule comparisons and intervention measures. Simultaneously, it enhances regulatory transparency and security, demonstrating a closed-loop, verifiable, and real-time traceable penetrating regulatory capability that existing technologies cannot provide.

[0128] S4. Automatically match the evidence storage data chain with compliance rules to identify potential violations and risk levels, and form a compliance result set.

[0129] S4.1 Compliance rules are derived by organizing normative documents, industry standards, platform management specifications, and historical violation cases, classifying, coding, and itemizing the relevant prohibitive clauses, restrictive requirements, and audit points, and then reviewing and revising them.

[0130] Specifically, prohibited clauses, restrictive requirements, and key review points are extracted from normative documents, industry standard documents, platform management regulations, and historical violation case records, and then classified according to clause type, applicable objects, and risk level. For example, prohibited clauses involving false advertising are classified as content compliance clauses, restrictive requirements involving restrictions on live streaming permissions are classified as entity qualification clauses, and key review points involving the sale of high-risk product categories are classified as transaction security clauses.

[0131] The categorized clauses are coded and assigned unique identifiers, and the clause content, scope of application, and key review points are organized into a structured item table.

[0132] The structured item list is professionally reviewed to verify the completeness, accuracy, and logical relationships between items.

[0133] Based on the proofreading results, revisions are made, missing clauses are supplemented, unclear or ambiguous clauses are corrected, and compliance rules are generated after being sorted, coded and reviewed.

[0134] S4.2. The evidence storage data chain is split and organized according to time sequence and business objects to obtain a matching dataset.

[0135] Specifically, based on the timestamp information in the evidence storage data chain, the evidence storage data chain is arranged in ascending order according to time sequence;

[0136] Based on the business object identifier, the business objects to which each evidence storage data block belongs in the evidence storage data chain are grouped; within each group, the evidence storage data blocks are sequentially numbered according to time order to generate serialized data records;

[0137] Perform integrity verification on each set of serialized data records, remove duplicate or missing evidence data blocks, and ensure that each record corresponds to a unique time step and business object;

[0138] The processed serialized data records are summarized to form a matching dataset that can be directly matched according to time sequence and business objects.

[0139] It should also be noted that the business objects are derived from the entity information that has been registered or recorded on the live-streaming e-commerce transaction platform, including anchor accounts, merchant accounts, unique product identifiers and transaction order numbers, etc., and can be confirmed and labeled in combination with internal business management records and user behavior data to ensure that each piece of evidence data clearly corresponds to a specific business object.

[0140] S4.3. Compare the matching dataset with each compliance rule, mark the existing violations and their corresponding risk levels, and obtain the marked result set.

[0141] Specifically, for each record in the matching dataset, the field information related to the business object is extracted one by one, and matched and compared with the prohibitive clauses, restrictive requirements and audit points in the compliance rules. The violation of the clause is determined by logical judgment and setting severity thresholds, frequency thresholds and impact scope thresholds. Each record with a violation is marked with the violation type and corresponding risk level. For example, high risk is marked as level three and low risk is marked as level one. A marked result set containing violation behavior markings and risk level information is generated.

[0142] It should also be explained that determining whether a violation of the terms has occurred through logical judgment and setting severity thresholds, frequency thresholds, and impact scope thresholds specifically includes: collecting and summarizing historical live streaming behavior data, historical violation records, and user feedback data; extracting a set of original indicators for measuring severity, frequency, and impact scope according to violation type and impact indicators; cleaning and normalizing each original indicator to obtain standardized severity values, standardized frequency values, and standardized impact scope values; calculating distribution statistics for standardized indicators on historical samples, and initially selecting candidate severity thresholds, frequency thresholds, and impact scope thresholds based on business risk tolerance and regulatory requirements; backtesting the candidate severity thresholds, frequency thresholds, and impact scope thresholds on independent verification samples, calculating false positive and false negative rates, and adjusting the candidate severity thresholds, frequency thresholds, and impact scope thresholds accordingly; organizing a review after backtesting and adjustment, and confirming and revising the thresholds in conjunction with normative documents and platform operation rules; writing the confirmed severity thresholds, frequency thresholds, and impact scope thresholds into the rule base and recording the version number and effective time for subsequent comparison and periodic recalibration;

[0143] Through logical judgment, the actual severity, frequency, and scope of impact of the behavior are compared with the severity threshold, frequency threshold, and scope of impact threshold, respectively. When any indicator reaches or exceeds the severity threshold, frequency threshold, or scope of impact threshold, it is judged as a violation of the corresponding clause, thereby forming a violation mark and risk level.

[0144] S4.4 Summarize and structure the marked result set to generate a compliance result set.

[0145] Specifically, all violations in the marking result set are grouped according to business records, and the number of violations, the distribution of violation categories, and the distribution of risk levels corresponding to each business record are counted to form a preliminary statistical table;

[0146] Based on the severity threshold, frequency threshold, and scope threshold of the violation, the violation of each business record is weighted and calculated to generate a comprehensive violation score;

[0147] The comprehensive violation score, violation category distribution, and risk level distribution of each business record are stored in a table to generate a compliance result set.

[0148] S5. The compliance result set, historical violation records and user feedback data are integrated and calculated to generate a comprehensive risk score. Based on the comprehensive risk score, real-time compliance intervention measures are automatically triggered to form an intervention result set.

[0149] S5.1 Historical violation records and user feedback data are obtained by auditing and archiving the platform's completed live broadcast records and transaction records, combining user complaints, evaluations, reports and customer service handling results during the live broadcast process, summarizing and structuring them, and then reviewing and classifying them.

[0150] Specifically, audit data is extracted from the platform's completed live streaming and transaction records in chronological order, missing values ​​are filled and outliers are checked to form a preliminary historical record table;

[0151] Collect user complaints, comments, reports, and customer service handling results during the live stream, and link them to the corresponding live stream records and transaction records according to the business object;

[0152] The associated data is then standardized in terms of fields, categorized and coded, and deduplicated to form a preliminary structured table.

[0153] After reviewing the accuracy of the violation category, processing result, and user feedback for each record, the records are categorized and labeled according to preset classification standards to generate a dataset of historical violation records and user feedback.

[0154] It should also be noted that the sources of the preset classification standards include the provisions on online transactions and live streaming behavior in relevant normative documents, industry standards and operating guidelines issued by industry associations or internal organizations, their own management rules, and the analysis results of historical violation cases. These are then organized, compared, and abstracted to form a classification standard applicable to the structured organization of historical violation records and user feedback data.

[0155] S5.2 Summarize and organize the compliance result set, historical violation records and user feedback data to obtain a unified multi-source data set.

[0156] Specifically, the compliance result set, historical violation records, and user feedback data are first standardized according to timestamps, business object identifiers, and data field names. The data fields are then unified into a single field format. Duplicate records are deduplicated, missing values ​​are filled with interpolation or default values, and outliers are corrected or marked according to anomaly judgment rules. The compliance result set, historical violation records, and user feedback data are then categorized and summarized according to preset classification standards. Finally, they are sorted and structured according to time order, business object, and violation type to generate a unified multi-source data set containing unified fields, classification labels, and time series information. For example, transaction violation records, user complaint information, and historical audit results are integrated into a single structured record.

[0157] S5.3. Associate and classify the unified multi-source data set according to live streaming behavior and time sequence to generate structured data tables.

[0158] Specifically, for a unified multi-source dataset, data with the same live streaming behavior are sorted in chronological order based on the live streaming behavior identifier field and the timestamp field.

[0159] Extract key fields such as business object identifier, violation type, user feedback, and transaction information from each record, and standardize the field names to a field format;

[0160] Records are categorized and archived according to the types of live streaming behavior, for example, product display behavior, interactive behavior, and transaction behavior are classified into different categories.

[0161] Multiple records under the same live streaming behavior are linked to generate a structured data table containing time series information, business object information, and behavior classification tags. Each record contains a complete set of fields and a classification identifier, ensuring that the data in the table is arranged in chronological order for easy subsequent analysis.

[0162] S5.4 Perform multi-dimensional data fusion and compliance rule comparison on the structured data table, calculate the comprehensive risk factor value for each live broadcast behavior, and generate a comprehensive risk score.

[0163] Specifically, for structured data tables, fields such as violation type identifier, user feedback score, historical violation count, and transaction anomaly indicators are extracted from each record, and the fields are standardized into a risk assessment field format;

[0164] The comprehensive risk factor value is calculated by weighting and summing fields such as violation type identifier, user feedback rating, historical violation count, and transaction anomaly indicators for the same live streaming behavior. The expression is as follows:

[0165] R = w v ×V+w u ×U+w h ×H+w t ×T;

[0166] Where R represents the comprehensive risk factor value, w v V represents the weighting coefficient for the violation type identifier, and w represents the violation type identifier. u This represents the weighting coefficient of user feedback ratings, where U represents the user feedback rating, and w h The weighting coefficient represents the number of historical violations, where H represents the number of historical violations, and w t The weighting coefficient of the abnormal trading indicator is represented by T, where T represents the abnormal trading indicator.

[0167] It should be noted that the weight coefficients of the violation type identifier, user feedback rating, historical violation count, and transaction anomaly indicator in the formula are obtained by statistical analysis of historical violation case records, and the importance is calculated using the analytic hierarchy process. Finally, the results are normalized so that the sum of each weight coefficient is 1.

[0168] The comprehensive risk factor value of each live broadcast behavior is compared with the severity threshold, frequency threshold, and scope of impact threshold one by one. When the comprehensive risk factor value reaches or exceeds the severity threshold, frequency threshold, or scope of impact threshold, the corresponding threshold level judgment result is obtained. The risk level of the live broadcast behavior is determined comprehensively based on the comparison results of the severity threshold, frequency threshold, and scope of impact threshold.

[0169] The calculation results are populated into a structured data table to generate comprehensive risk score data that includes behavior identifiers, timestamps, business objects, and comprehensive risk scores.

[0170] S5.5. Based on the comprehensive risk score, trigger corresponding real-time compliance intervention measures, and summarize and organize the operational behaviors, modification status, warnings and restrictions during the execution process to form an intervention result set.

[0171] Specifically, based on the comprehensive risk score data, each live broadcast behavior is compared with the preset risk grading threshold to determine the category of real-time compliance intervention measures that should be triggered.

[0172] Implement intervention measures, including recording operational behaviors, modifying the status of live broadcast content, generating warning messages, and restricting operational results. Each operation is recorded with a timestamp, behavior identifier, and execution result.

[0173] The intervention process and corresponding operation status of each live broadcast behavior are summarized and organized to form an intervention result set that includes behavior identifier, risk level, intervention category, operation behavior, modification status, warning and restriction results.

[0174] It should also be explained that the specific steps for setting the risk classification threshold are as follows: collect historical live streaming behavior data, historical violation records, and user feedback data; statistically analyze the distribution of risk indicators for each type of violation and abnormal operation, such as the percentage of violations, severity scores, and frequency; determine the classification boundaries of risk indicators based on the statistical results, setting high-risk behavior as an example value greater than or equal to 0.8, medium-risk behavior as an example value between 0.5 and 0.8, and low-risk behavior as an example value less than 0.5; review and adjust the classification boundaries to form the preset risk classification threshold for comprehensive risk scoring comparison.

[0175] S6. Restore the intervention result set, comprehensive risk score and violation behavior to the blockchain to form a closed-loop evidence storage data chain.

[0176] S6.1 Summarize and organize the intervention result set, comprehensive risk score and marked violations to generate an organized data set.

[0177] Specifically, it reads the operational behavior, modification status, warning and restriction results data in the intervention result set, reads the risk score of each live broadcast behavior in the comprehensive risk score, and reads the violation behavior and risk level information in the marked result set;

[0178] By associating live streaming behaviors with unique identifiers, intervention result sets, comprehensive risk scores, and marked violations under the same live streaming behavior are merged to form a unified record;

[0179] The merged records are sorted in chronological order, and duplicate or redundant information is removed.

[0180] Each record is labeled with corresponding fields for intervention behavior, risk score, and violation behavior, such as operation behavior field, risk score field, and violation behavior field, to generate a sorted data set.

[0181] S6.2. Format and structure the sorted data set to generate blockchain storage and evidence data blocks.

[0182] Specifically, the system reads the operation behavior, risk score, and violation fields of each record in the processed dataset.

[0183] In accordance with the field order and data type requirements of blockchain storage, the operation behavior field, risk score field, violation behavior field and timestamp field are uniformly encoded and formatted. For example, text fields are converted to standard character encoding and numeric fields retain a specified number of decimal places.

[0184] The formatted records are grouped into fixed-size evidence storage blocks in chronological order, for example, each block contains one thousand records;

[0185] A unique identifier and digest information are generated for each evidence storage data block to form a blockchain storage evidence storage data block. The start time, end time and data volume fields are recorded for each evidence storage data block to obtain the blockchain storage evidence storage data block.

[0186] S6.3. Perform hash calculations on the blockchain-stored evidence data blocks and add timestamps to form verifiable block data, and write it into the blockchain to form a closed-loop evidence data chain.

[0187] Specifically, it reads the record content and field information of each evidence storage data block in the blockchain storage evidence storage data block set;

[0188] A predetermined hash algorithm is used to perform hash operations on all fields of each evidence storage data block to generate unique digest information. For example, the SHA-256 algorithm is used to obtain a hash value with a length of 256 bits.

[0189] The generated hash digest is bound to the current timestamp field to form verifiable block data;

[0190] Verifiable block data is written to the blockchain network block by block in chronological order, and the successful writing is verified through multi-node consensus.

[0191] For each written evidence storage data block, the block height, transaction identifier, and previous block hash are recorded and sequentially linked to form a closed-loop evidence storage data chain, ensuring the integrity and traceability of each block.

[0192] This embodiment also provides a compliance review system for live-streaming e-commerce based on penetrating supervision, including: a data acquisition module for collecting initial data and preprocessing it to form a multi-source dataset; a feature comparison module for automatically comparing the appearance features, brand logo features, and packaging features of product images using a deep convolutional neural network to form a judgment result dataset; an evidence storage and chaining module for storing the judgment result dataset and the multi-source dataset in a unified blockchain to form an evidence storage data chain; a rule matching module for automatically matching the evidence storage data chain with compliance rules to identify potential violations and risk levels, forming a compliance result set; a risk assessment module for integrating and calculating the compliance result set, historical violation records, and user feedback data to generate a comprehensive risk score, and automatically triggering real-time compliance intervention measures based on the comprehensive risk score to form an intervention result set; and a closed-loop evidence storage module for re-storing the intervention result set, comprehensive risk score, and violations in the blockchain to form a closed-loop evidence storage data chain.

[0193] This embodiment also provides a computer device applicable to the compliance review method for live-streaming e-commerce based on penetrating supervision, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the compliance review method for live-streaming e-commerce based on penetrating supervision as proposed in the above embodiment.

[0194] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0195] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the compliance audit method for live-streaming e-commerce based on penetration-based supervision as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0196] In summary, this invention achieves verifiability and tamper-proof storage by: calculating hash digests of the judgment results and multi-source data to generate unique identifiers; dividing the data into blocks; attaching timestamps when uploading to the blockchain; and verifying the data through a multi-node consensus protocol. This forms an immutable and traceable evidence storage data chain. It provides a secure and reliable evidence storage foundation for the entire live streaming process, while binding each data action to its time sequence, ensuring clear traceability for subsequent compliance rule matching and risk scoring. This enhances regulatory transparency and guarantees data integrity, providing reliable support for real-time intervention and closed-loop auditing.

[0197] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A compliance review method for live-streaming e-commerce based on penetrating supervision, characterized in that: include, Initial data is collected and preprocessed to form a multi-source dataset; A deep convolutional neural network is used to automatically compare the appearance features, brand logo features, and packaging features of product images to form a judgment result dataset. The judgment result dataset and the multi-source dataset are hash digested, stored in blocks on the blockchain, and the storage process is verified through timestamps and multi-node consensus protocols to form a chain of evidence storage data. The evidence storage data chain is automatically matched with compliance rules to identify potential violations and risk levels, forming a compliance result set; The system integrates and calculates compliance result sets, historical violation records, and user feedback data to generate a comprehensive risk score. Based on the comprehensive risk score, it automatically triggers real-time compliance intervention measures, forming an intervention result set. The intervention results set, comprehensive risk score, and violations are re-stored on the blockchain to form a closed-loop evidence storage data chain.

2. The live-streaming e-commerce compliance review method based on penetrating supervision as described in claim 1, characterized in that: The specific steps for forming a multi-source dataset are as follows. Collect the video stream, audio stream, text comments, product images, and user behavior of the host during the live broadcast, perform timestamp alignment and event-driven sampling, and generate initial data; The system extracts keyframes and generates perceptual fingerprints from video streams, separates sound sources and transcribs semantics from audio streams, desensitizes and semantically annotates text bullet comments, performs region detection and OCR parsing on product images, and models click sequences and paths of user behavior, fusing them together to form feature-based data. The feature-rich data undergoes cross-modal temporal fusion, quality enhancement, and privacy tokenization to generate a multi-source dataset.

3. The live-streaming e-commerce compliance review method based on penetrating supervision as described in claim 1, characterized in that: The specific steps for forming the judgment result dataset are as follows: The product image is input into a deep convolutional neural network to extract appearance features, brand logo features, and packaging features. These features are then compared one by one with the product feature vectors that have been pre-extracted and stored in a standard product image library to obtain a comparison result set. The results set is compared to determine its authenticity and confidence level, forming a judgment result dataset.

4. The live-streaming e-commerce compliance review method based on penetrating supervision as described in claim 1, characterized in that: The specific steps for forming the evidence storage data chain are as follows: The judgment result dataset and the multi-source dataset are formatted and deduplicated to generate a data package to be stored. The data packets to be stored are hash digested to generate a unique digest, and a data block index list is established to generate hash index data. The data packet to be stored is divided into several data blocks based on the hash index data, and a timestamp and digest verification are added to each block to generate a set of data blocks to be stored on the chain. Submit the set of data blocks to be stored on the blockchain to initiate a write transaction. After multi-node consensus verification, the block height and transaction identifier are returned, and an on-chain confirmation record is generated. Based on the on-chain confirmation records, a mapping is established between the evidence storage data block summary and the on-chain identifier, and they are linked together in chronological order to form an evidence storage data chain.

5. The live-streaming e-commerce compliance review method based on penetrating supervision as described in claim 1, characterized in that: The compliance rules are derived by organizing normative documents, industry standards, platform management specifications, and historical violation cases, classifying, coding, and itemizing the prohibited clauses, restrictive requirements, and key review points involved, and then reviewing and revising them.

6. The live-streaming e-commerce compliance review method based on penetrating supervision as described in claim 1, characterized in that: The specific steps for forming the compliance result set are as follows: The evidence storage data chain is broken down and organized according to time sequence and business objects to obtain a matching dataset; The matching dataset is compared with each compliance rule, and the existing violations and their corresponding risk levels are marked to obtain the marked result set; The tagged result set is summarized and structured to generate a compliance result set.

7. The live-streaming e-commerce compliance review method based on penetrating supervision as described in claim 1, characterized in that: The historical violation records and user feedback data are obtained by auditing and archiving the platform's completed live broadcast records and transaction records, summarizing and structuring user complaints, evaluations, reports and customer service handling results during the live broadcast, and then reviewing and classifying them.

8. The live-streaming e-commerce compliance review method based on penetrating supervision as described in claim 1, characterized in that: The specific steps for forming the intervention result set are as follows: The compliance result set, historical violation records, and user feedback data are summarized and organized to obtain a unified multi-source data set; The unified multi-source data set is associated and classified according to live streaming behavior and time sequence to generate structured data tables; Multidimensional data fusion and compliance rule comparison are performed on structured data tables to calculate the comprehensive risk factor value for each live streaming behavior and generate a comprehensive risk score; Based on the comprehensive risk score, corresponding real-time compliance intervention measures are triggered, and the operational behaviors, modification status, warnings and restrictions during the execution process are summarized and organized to form an intervention result set.

9. The live-streaming e-commerce compliance review method based on penetrating supervision as described in claim 1, characterized in that: The specific steps for forming a closed-loop evidence storage data chain are as follows. The intervention result set, comprehensive risk score, and marked violations are summarized and organized to generate a sorted data set; The organized data set is formatted and structured to generate blockchain storage and evidence-keeping data blocks; Hash the blockchain-stored evidence data blocks and add timestamps to form verifiable block data, then write it into the blockchain to form a closed-loop evidence data chain.

10. A compliance audit system for live-streaming e-commerce based on penetrating supervision, based on the compliance audit method for live-streaming e-commerce based on penetrating supervision as described in any one of claims 1 to 9, characterized in that: include, The data acquisition module is used to collect initial data and preprocess it to form a multi-source dataset; The feature comparison module is used to automatically compare the appearance features, brand logo features, and packaging features of product images using a deep convolutional neural network to form a judgment result dataset. The evidence storage and chain building module is used to store the judgment result dataset and multi-source datasets in a unified manner on the blockchain to form an evidence storage data chain; The rule matching module is used to automatically match the evidence storage data chain with compliance rules, identify potential violations and risk levels, and form a compliance result set. The risk assessment module is used to integrate and calculate compliance result sets, historical violation records, and user feedback data to generate a comprehensive risk score, and automatically trigger real-time compliance intervention measures based on the comprehensive risk score to form an intervention result set; The closed-loop evidence storage module is used to re-store the intervention result set, comprehensive risk score and violation behavior on the blockchain to form a closed-loop evidence storage data chain.

Images