Block chain-based electronic commerce platform counterfeit commodity big data identification method

Abstract

The invention discloses an e-commerce platform counterfeit commodity big data identification method based on a block chain, and relates to the technical field of computer data processing, and the method comprises the steps: 1, constructing commodity multi-modal feature fusion representation, and fusing text, image and logistics track features; step 2, establishing block chain credible evidence storage and on-chain and off-chain bidirectional index mapping; 3, realizing rapid approximate retrieval of massive commodities by adopting locality sensitive hashing; step 4, executing multi-level counterfeit suspicion judgment and early warning based on comparison of a similarity threshold value and an evidence; and 5, constructing a closed-loop feedback mechanism to realize adaptive updating of the model and continuous enriching of block chain data. According to the method, the accuracy, the real-time performance and the system credibility of counterfeit commodity recognition are remarkably improved, and the application requirements in a high-efficiency power generation merchant scene are met.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and in particular to a blockchain-based method for identifying counterfeit goods on e-commerce platforms using big data. Background Technology

[0002] With the large-scale development of e-commerce platforms, commodity transaction data is growing exponentially, and the concealment and spread speed of counterfeit goods have also increased significantly. E-commerce platforms rely on traditional keyword matching and manual review mechanisms for product compliance screening, which are insufficient to cope with the complex semantic spoofing and cross-platform collaborative counterfeiting behaviors in the massive amount of product information. Especially in high-concurrency transaction scenarios, product description text, image features, and supply chain traceability information need to be cross-validated within milliseconds, which poses a severe challenge to the system's real-time analysis capabilities, data credibility assurance, and the accuracy of multimodal feature fusion.

[0003] Among them, blockchain-based e-commerce data storage technology provides an immutable underlying support for product traceability. Its core lies in anchoring key operation records throughout the product's lifecycle to a distributed ledger in the form of hash values. However, existing identification models typically separate blockchain-stored data from product characteristic features, failing to establish a dynamic mapping relationship between trusted on-chain data and heterogeneous off-chain information from multiple sources.

[0004] Existing technologies have several shortcomings in the process of identifying counterfeit goods: First, the feature extraction of product images and text descriptions uses independent encoders, which leads to the accumulation of cross-modal semantic alignment errors and cannot effectively identify counterfeit behaviors such as discrepancies between images and text; second, blockchain evidence storage is only used for post-event traceability and lacks a feedback coupling mechanism with real-time identification models, which prevents abnormal transaction behaviors from triggering immediate alerts on the chain; third, fine-grained comparison of similar products in a big data environment relies on global hash indexes, which has high computational complexity and is difficult to meet the throughput requirements of e-commerce platforms that add tens of thousands of new products per second. Summary of the Invention

[0005] The purpose of this invention is to provide a blockchain-based method for identifying counterfeit goods on e-commerce platforms using big data, in order to solve the problems of low accuracy in identifying counterfeit goods, difficulty in cross-modal information fusion, and insufficient data credibility in existing technologies.

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

[0007] A blockchain-based method for identifying counterfeit goods on e-commerce platforms using big data includes the following specific steps:

[0008] Step 1: Construct a multimodal feature fusion representation of the product. The product description text, product main image and detail image sequence, and supply chain logistics trajectory data are acquired in real time through the data acquisition interface deployed on the e-commerce platform server. For the text data, a bidirectional long short-term memory network based on the attention mechanism is used to extract semantic feature vectors. For the image data, a deep residual network is used to extract visual feature vectors. For the logistics data, a temporal convolutional network is used to extract trajectory feature vectors. The above three types of feature vectors are input into the cross-modal feature alignment module. The semantic space is aligned and fused through a multi-head attention mechanism to generate a unified multimodal feature representation of the product.

[0009] Step 2: Establish a blockchain-based trusted evidence storage and dynamic mapping mechanism. The product identification code at the time of product listing, the supplier's digital identity certificate, and the hash value of the multimodal feature representation generated in Step 1 are used to form an evidence storage data unit. This evidence storage data unit is written into the distributed ledger through a consensus algorithm. At the same time, a two-way index mapping table is constructed between the on-chain evidence storage hash value and the off-chain product multimodal feature representation. This mapping table supports two-way queries using either the product identification code or the evidence storage hash value.

[0010] Step 3: Implement fast retrieval of approximate nearest neighbors based on locality-sensitive hashing. Construct a locality-sensitive hash index structure for the full product database of the e-commerce platform. Map the multimodal feature representation of the products generated in Step 1 to a low-dimensional Hamming space using a random projection function to generate a fixed-length hash signature. Then, construct a multi-table hashing structure. Using hash buckets, similar product feature vectors are grouped into the same or adjacent hash buckets, reducing the retrieval time complexity from... Reduce to ;

[0011] Step 4: Perform multi-level anti-counterfeiting judgment and warning. For the product to be detected, firstly, the Top-K most similar products are quickly retrieved through the local sensitive hash index in Step 3. Then, the cosine similarity between the product to be detected and each similar product in the multimodal feature space is calculated. The first-level similarity threshold is set to 0.85. When there are products with similarity greater than 0.85, the blockchain evidence records of the two are further compared. If the supplier digital identity certificate in the evidence record is inconsistent, an anti-counterfeiting warning is triggered.

[0012] Step 5: Construct a closed-loop feedback and model adaptive update system. Add the counterfeit product cases and their multimodal feature representations confirmed in Step 4 as negative samples to the training dataset. Periodically retrain the feature extraction network and cross-modal alignment module in Step 1. At the same time, add the confirmed counterfeit behavior records and their processing results as new evidence storage data units to the blockchain, forming a positive feedback loop of improved recognition model performance and rich blockchain evidence storage data.

[0013] Preferably, in step 1, the cross-modal feature alignment module adopts a 12-head attention mechanism, with each attention head having a dimension of 64. It generates query vectors, key vectors, and value vectors by performing linear transformations on text feature vectors, image feature vectors, and logistics feature vectors. Then, it calculates cross-modal attention weights by scaling dot product attention, and finally outputs a unified product multimodal feature representation with a dimension of 768.

[0014] Preferably, in step 1, text feature extraction uses a bidirectional long short-term memory network with a vocabulary size of 50,000, a hidden layer dimension of 512, and a dropout rate of 0.1; image feature extraction uses a ResNet50 deep residual network, the input image size is adjusted to 224 pixels by 224 pixels, and the global average pooling feature of the last convolutional layer is extracted with a dimension of 2048; logistics feature extraction uses a 5-layer temporal convolutional network with a kernel size of 3 and the number of channels being 64, 128, 256, 512, and 512 respectively.

[0015] Preferably, in step 2, the evidence storage data unit needs to undergo SHA256 hash operation to generate a 32-byte hash value before being written to the blockchain. The consensus algorithm adopts an improved practical Byzantine fault-tolerant algorithm, which requires more than two-thirds of the nodes to reach a consensus before the data can be uploaded to the blockchain. The data upload time delay is controlled within 2 seconds.

[0016] Preferably, the bidirectional index mapping table in step 2 is implemented using a distributed key-value database, which supports more than 10,000 concurrent read and write operations per second. The index data retains a hot data copy in memory, with a response time of less than 10 milliseconds. At the same time, the index snapshot is persisted to the solid-state drive every hour.

[0017] Preferably, in step 3, the number of random projection functions for locality-sensitive hashing is set to 256, the Hamming space dimension is 64, and each hash table contains... Each hash bucket is used to build a multi-table index through eight independent hash tables. During retrieval, only a linear scan within one hash bucket is needed to find the approximate nearest neighbor, ensuring a retrieval accuracy of over 95%.

[0018] Preferably, step 4, the multi-level counterfeit suspicion determination, also includes a second-level determination: when the cosine similarity is between 0.7 and 0.85, fine-grained feature comparison is initiated, and the Mahalanobis distance between the two in the key attribute dimension is calculated, using the formula:

[0019]

[0020] Where x and y are the feature vectors of the two products, and S is the feature covariance matrix. When the Mahalanobis distance is greater than 3, a counterfeit suspicion warning is also triggered.

[0021] Preferably, after the warning signal is generated in step 4, the system automatically generates a visual report containing the product identification code, a list of similar products, similarity values, and evidence comparison results, and pushes the warning information to the platform review terminal in real time through a message queue. Reviewers need to complete manual review and confirmation within 30 minutes.

[0022] Preferably, in step 5, the model update adopts an incremental learning strategy, and a model fine-tuning is initiated every 1000 new training samples. During training, the basic network parameters are kept fixed, and only the parameters of the last 3 fully connected layers are updated. The learning rate is set to 0.001, and the stochastic gradient descent algorithm with momentum is used for optimization. The training cycle is 50 rounds.

[0023] Preferably, the closed-loop feedback system in step 5 also includes a performance monitoring module that continuously tracks the accuracy, recall and F1 score of the identification model. When the F1 score drops by more than 5% for three consecutive statistical periods, the model is automatically retrained using all data from the last 30 days.

[0024] Preferably, the method is deployed in a cloud computing environment and adopts a containerized architecture for horizontal scaling. A single container instance can concurrently process 100 product identification requests with an average response time of less than 500 milliseconds. The system supports elastic scaling and can automatically adjust computing resources according to traffic fluctuations during e-commerce platform promotional activities.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] This invention establishes a dual guarantee mechanism of multimodal feature fusion representation of commodities and blockchain trusted evidence storage, realizing dynamic mapping and collaborative analysis of on-chain trusted data and off-chain multi-source heterogeneous information, effectively solving the problem of cross-modal semantic alignment error.

[0027] This invention employs Locality Sensitive Hash Index (LSI) technology to significantly reduce the computational complexity of comparing massive amounts of goods, enabling the system to meet the real-time recognition requirements of e-commerce platforms under high-concurrency scenarios.

[0028] The multi-level anti-counterfeiting judgment system and closed-loop feedback mechanism established by this invention not only improves the recognition accuracy, but also realizes the continuous self-optimization of the recognition model and the continuous enrichment of blockchain evidence data, ultimately forming a virtuous cycle in which recognition ability and data credibility promote each other. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall technical solution architecture of a blockchain-based big data identification method for counterfeit goods on e-commerce platforms proposed in this invention.

[0030] Figure 2This is a schematic diagram illustrating the core principle framework of the product multimodal feature fusion and blockchain trusted evidence dynamic mapping mechanism proposed in this invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0032] Currently, with the large-scale development of e-commerce platforms, commodity transaction data is growing exponentially, significantly increasing the concealment and spread speed of counterfeit goods. E-commerce platforms rely on traditional keyword matching and manual review mechanisms for product compliance screening, which struggles to cope with the complex semantic spoofing and cross-platform collaborative counterfeiting behaviors within massive amounts of product information. Especially in high-concurrency transaction scenarios, product description text, image features, and supply chain traceability information need to be cross-validated within milliseconds, posing a severe challenge to the system's real-time analysis capabilities, data credibility assurance, and the accuracy of multimodal feature fusion. Blockchain-based e-commerce data storage technology provides tamper-proof underlying support for product traceability, its core being the anchoring of key operation records throughout the product's lifecycle to a distributed ledger in the form of hash values. However, existing identification models typically separate blockchain-stored data from product representation features, failing to construct a dynamic mapping relationship between on-chain trusted data and off-chain multi-source heterogeneous information. To address these technical problems, this invention proposes a counterfeit goods intelligent identification method that deeply integrates blockchain trusted storage and multimodal big data analysis, and applies it to a blockchain-based e-commerce platform counterfeit goods big data identification method.

[0033] refer to Figure 1 The overall technical architecture of the blockchain-based e-commerce platform counterfeit goods big data identification method proposed in this invention includes five core steps:

[0034] Step 1: Construct a multimodal feature fusion representation of the product;

[0035] Step 2: Establish a blockchain-based trusted evidence storage and dynamic mapping mechanism;

[0036] Step 3 implements fast approximate nearest neighbor retrieval based on locality-sensitive hashing;

[0037] Step 4: Perform multi-level anti-counterfeiting assessment and early warning;

[0038] Step 5: Construct a closed-loop feedback and model adaptive update system.

[0039] The above steps work together to form a counterfeit product identification system with high real-time performance, high accuracy, and continuous evolution capabilities.

[0040] Step 1: Construct a multimodal feature fusion representation for the product. This involves acquiring product description text, product main image and detail image sequences, and supply chain logistics trajectory data in real time through a data acquisition interface deployed on the e-commerce platform server. Specifically, the data acquisition interface is integrated into the e-commerce platform's backend service cluster as an API gateway, supporting the processing of no less than 5000 product listing or editing events per second. For the product description text, a bidirectional long short-term memory network based on an attention mechanism is used to extract semantic feature vectors; for the product main image and detail image sequences, a deep residual network is used to extract visual feature vectors; and for the supply chain logistics trajectory data, a temporal convolutional network is used to extract trajectory feature vectors. Subsequently, these three types of feature vectors are input into a cross-modal feature alignment module, where a multi-head attention mechanism is used to align and fuse the semantic space, generating a unified multimodal feature representation for the product.

[0041] Specifically, the text feature extraction process in step 1 is as follows: First, the original product description text undergoes standardized preprocessing, including HTML tag removal, unified encoding format, word segmentation, and stop word filtering, ultimately mapping it to a predefined vocabulary. This vocabulary has a size of 50,000 words, covering high-frequency product terms and brand names from mainstream e-commerce platforms. The preprocessed word sequence is input into a bidirectional long short-term memory network, which contains two LSTM layers (forward and backward), with a hidden layer dimension of 512 and a dropout rate of 0.1 to prevent overfitting. The network output is the hidden state at each time step, which, after global max pooling, yields a fixed-dimensional semantic feature vector with a dimension of 512.

[0042] Specifically, the image feature extraction process in step 1 is as follows: The product main image and details image sequences are first uniformly adjusted to 224-pixel multiplied 224-pixel RGB format images, which are then used as input to the ResNet50 deep residual network. This network is pre-trained on the ImageNet dataset, retaining all convolutional layer weights and only fine-tuning the final classification layer. After the input image is processed through 50 residual blocks, the output feature map of the last convolutional layer is extracted, and a global average pooling operation is performed on this feature map to generate a visual feature vector with a dimension of 2048. This vector effectively captures high-level semantic information such as the product's global appearance, texture, color distribution, and key component structure.

[0043] Specifically, the logistics trajectory feature extraction process in step 1 is as follows: Supply chain logistics trajectory data is provided in the form of a timestamp sequence, including fields such as origin of shipment, transit station, destination of shipment, mode of transport, estimated arrival time, and actual arrival time. This sequence is first encoded into a numerical vector; for example, geographical location is represented by latitude and longitude coordinates, and mode of transport is mapped using one-hot encoding. The encoded sequence is input into a 5-layer temporal convolutional network, with each layer having a kernel size of 3 and channels of 64, 128, 256, 512, and 512 respectively. The activation function is ReLU, and batch normalization is used between layers. The network output is the feature vector of the last time step, with a dimension of 512, representing the spatiotemporal regularity and degree of abnormal deviation of the logistics path.

[0044] After extracting the three types of single-modal features, the cross-modal feature alignment module in step 1 receives text feature vectors, image feature vectors, and logistics feature vectors as input. To unify the input dimensions, the three are mapped to a 768-dimensional space through independent linear projection layers. Subsequently, the cross-modal feature alignment module employs a 12-head attention mechanism, with each attention head having a dimension of 64. Specifically, it performs linear transformations on the feature vectors of the three modalities to generate query vectors. Key vector Sum value vector For the i-th attention head, calculate the scaled dot product attention weights:

[0045]

[0046] in, The outputs of 12 attention heads are concatenated and then mapped back to 768 dimensions through a linear layer, resulting in a unified multimodal feature representation of the product with a dimension of 768. This representation integrates complementary information from textual semantics, visual appearance, and logistics behavior in a unified semantic space, effectively eliminating cross-modal semantic alignment errors.

[0047] refer to Figure 2 Step 2 involves establishing a blockchain-based trusted evidence storage and dynamic mapping mechanism. This mechanism combines the product identifier code at the time of product listing, the supplier's digital identity certificate, and the hash value of the multimodal feature representation generated in Step 1 to form the evidence storage data unit. Specifically, the product identifier code is a unique product ID assigned by the e-commerce platform. The supplier's digital identity certificate is issued by an authoritative CA (Certificate Authority) and includes the supplier's public key, company information, and validity period. The multimodal feature representation first undergoes a SHA256 hash operation to generate a 32-byte hash value, ensuring the integrity and irreversibility of the feature data. These three data items are combined into a structured JSON object, serving as the evidence storage data unit.

[0048] The evidence-based data unit is written to the distributed ledger via a consensus algorithm. The consensus algorithm employs an improved Practical Byzantine Fault Tolerance (PBT) algorithm, requiring more than two-thirds of the nodes to reach a consensus before data can be uploaded to the blockchain, with the data upload time latency controlled to within 2 seconds. The blockchain network is jointly maintained by the e-commerce platform, third-party regulatory agencies, and core suppliers, with a minimum of seven nodes to ensure the system's decentralization and attack resistance.

[0049] Simultaneously, a bidirectional index mapping table is constructed between on-chain evidence storage hash values ​​and off-chain product multimodal feature representations. This mapping table is implemented using a distributed key-value database, such as Redis Cluster or Apache Cassandra, supporting over 10,000 concurrent read / write operations per second. Hot copies of the index data are maintained in memory, with a response time of less than 10 milliseconds, and index snapshots are persisted to solid-state drives every hour. The mapping table supports two query modes: one is to query the corresponding evidence storage hash value and multimodal feature representation using the product identifier code; the other is to query the product identifier code and related metadata using the evidence storage hash value. This bidirectional indexing mechanism achieves dynamic mapping between trusted on-chain data and multi-source heterogeneous off-chain information, providing fundamental support for subsequent real-time comparison and early warning.

[0050] Step 3: Implement fast near-neighbor retrieval based on Locality Sensitive Hashing (LSH). Construct an LSH index structure for the entire product database of the e-commerce platform. Specifically, the multimodal feature representation of the products generated in Step 1 is mapped to a low-dimensional Hamming space using random projection functions to generate a fixed-length hash signature. The number of random projection functions for LSH is set to 256, and the Hamming space dimension is 64. Each random projection function is a 768-dimensional random Gaussian vector. After performing a dot product with the input feature vector and taking the sign, a 1-bit binary output is obtained. The 256 projection results are concatenated to form a 64-byte hash signature.

[0051] Eight independent hash tables are constructed using multi-table hashing technology, each containing... There are 8 hash buckets. During index building, the hash signature of each product is inserted into the corresponding position in each of the 8 hash tables. During retrieval, a hash signature is generated for the product to be detected, and then its corresponding hash bucket and adjacent buckets (buckets with a Hamming distance of 1) are queried in each of the 8 hash tables. After merging all candidate products and removing duplicates, the Top-K most similar products are obtained. This mechanism clusters similar product feature vectors into the same or adjacent hash buckets, reducing the retrieval time complexity from... Reduce to Search accuracy is guaranteed at That's all. In actual deployment, a single search only requires a linear scan within one hash bucket to find the approximate nearest neighbor, greatly improving throughput in high-concurrency scenarios.

[0052] Step 4: Perform multi-level counterfeit suspicion judgment and early warning. For the product to be detected, firstly, the Top-K most similar products are quickly retrieved using the Local Sensitive Hash Index (LSH) from Step 3, with K set to 50 by default. Then, the cosine similarity between the product to be detected and each similar product in the multimodal feature space is calculated. The cosine similarity calculation formula is the product of the two vectors' dot product divided by their moduli, with the result ranging from -1 to 1. The closer the value is to 1, the more similar the product is.

[0053] The first-level similarity threshold is set at 0.85. When products with a similarity greater than 0.85 are found, their blockchain-based evidence records are further compared. Specifically, the comparison focuses on the supplier's digital identity certificate within the evidence records. If the supplier's digital identity certificates are inconsistent, a counterfeit suspicion alert is triggered. This judgment logic is based on the business assumption that "highly similar products should come from the same legitimate supplier," effectively identifying counterfeit activities with highly similar images and text but different origins.

[0054] Furthermore, step 4, the multi-level counterfeit suspicion determination, also includes a second-level determination. When the cosine similarity is between 0.7 and 0.85, fine-grained feature comparison is initiated. At this point, the system no longer relies on overall feature similarity but focuses on key attribute dimensions. Key attribute dimensions are determined through statistical analysis of historical counterfeit cases, such as the brand logo area, product model label, and main color scheme of the packaging. The Mahalanobis distance between the two in the key attribute dimension is calculated using the following formula:

[0055]

[0056] Here, x and y are the feature sub-vectors of two products in the key attribute dimension, and S is the feature covariance matrix in that dimension, calculated from a full sample of genuine products. Mahalanobis distance considers the correlation and scale difference between features and is more discriminative than Euclidean distance. A Mahalanobis distance greater than 3 also triggers a counterfeit suspicion warning.

[0057] Once an alert is generated, the system automatically produces a visual report containing the product identification code, a list of similar products, similarity scores, and evidence comparison results. This report is presented in a web interface, allowing auditors to intuitively view the comparison of images and text, differences in logistics tracking, and details of blockchain evidence storage. Simultaneously, the alert information is pushed to the platform's auditing terminal in real time via a message queue, requiring auditors to complete manual review and confirmation within 30 minutes. Alerts not processed in a timely manner will automatically escalate and notify senior auditors.

[0058] Step 5: Construct a closed-loop feedback and model adaptive update system. Add the counterfeit product cases and their multimodal feature representations, verified manually in Step 4, as negative samples to the training dataset. Simultaneously, genuine product samples are continuously accumulated as positive samples. Periodically retrain the feature extraction network and cross-modal alignment module from Step 1.

[0059] The model updates employ an incremental learning strategy, initiating a fine-tuning process every 1000 new training samples. During training, the basic network parameters are kept fixed; the weights of the ResNet50 convolutional layers, LSTM recurrent layers, and TCN convolutional layers are frozen. Only the parameters of the last three fully connected layers are updated, including the projection layer and attention layer of the cross-modal alignment module. The learning rate is set to... The model was optimized using stochastic gradient descent with momentum, a momentum coefficient of 0.9, and a training cycle of 50 epochs. This strategy ensures model stability while efficiently absorbing new knowledge.

[0060] In addition, the closed-loop feedback system in step 5 also includes a performance monitoring module. This module continuously tracks the accuracy, recall, and F1 score of the identification model, with a statistical period of once per hour. If the F1 score drops by more than [a certain percentage] for three consecutive statistical periods, [the system will detect the F1 score]. When this happens, the entire model retraining process is automatically triggered, using all data from the last 30 days for training to restore model performance.

[0061] The entire methodology is deployed in a cloud computing environment and scales horizontally using a containerized architecture. Each functional module is encapsulated as an independent Docker container and orchestrated and managed using Kubernetes. The functional modules include data acquisition, feature extraction, LSH indexing, a decision engine, and a feedback system. A single container instance can concurrently process 100 product recognition requests, with an average response time of less than 500 milliseconds. The system supports elastic scaling and can automatically adjust computing resources based on traffic fluctuations during e-commerce platform promotional activities. For example, during the Double 11 shopping festival, it can automatically scale to 1000 container instances, achieving a peak throughput of 100,000 recognition requests per second.

[0062] To verify the effectiveness of this invention, a simulated e-commerce platform environment was constructed, containing data on 10 million products, of which counterfeit products accounted for a certain percentage. Experimental results show that the method of this invention achieves a counterfeit product identification accuracy of 98.7% and a recall rate of [missing information]. F1 score is The traditional keyword matching method has an F1 score of 1. The traditional keyword matching method has an F1 score of 1. This significantly outperforms traditional keyword matching methods and single-modal deep learning methods. Furthermore, under a stress test with tens of thousands of QPS, the system's average response time remains stable at 420 milliseconds, meeting the requirements for high-concurrency real-time recognition.

[0063] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A blockchain-based big data identification method for counterfeit goods on e-commerce platforms, characterized in that, The specific steps include the following: Step 1: Construct a multimodal feature fusion representation of goods. Obtain goods data in real time through the data acquisition interface deployed on the e-commerce platform server. Extract feature vectors from the goods text data, image data and logistics data. Input the above three types of feature vectors into the cross-modal feature alignment module to generate a unified multimodal feature representation of goods. Step 2: Establish a blockchain trusted evidence storage and dynamic mapping mechanism. The product identification code at the time of product listing, the supplier's digital identity certificate, and the hash value of the multimodal feature representation generated in Step 1 are used to form an evidence storage data unit. This evidence storage data unit is written into the distributed ledger, and a two-way index mapping table is constructed at the same time. Step 3: Implement fast retrieval of approximate nearest neighbors based on locality-sensitive hashing, construct a locality-sensitive hash index structure, map the multimodal feature representation of the goods generated in Step 1 to a low-dimensional Hamming space through a random projection function, generate hash signatures, and construct hash buckets to aggregate similar product feature vectors into the same or adjacent hash buckets. Step 4: Perform multi-level anti-counterfeiting judgment and warning. For the product to be detected, firstly, the Top-K most similar products are quickly retrieved through the local sensitive hash index in Step 3. Then, the cosine similarity between the product to be detected and each similar product in the multimodal feature space is calculated. The first-level similarity threshold is set to 0.

85. When there are products with similarity greater than 0.85, the blockchain evidence records of the two are further compared. If the supplier digital identity certificate in the evidence record is inconsistent, an anti-counterfeiting warning is triggered. Step 5: Construct a closed-loop feedback and model adaptive update system. Add the counterfeit product cases and their multimodal feature representations confirmed in Step 4 as negative samples to the training dataset. Periodically retrain the feature extraction network and cross-modal alignment module in Step 1. At the same time, add the confirmed counterfeit behavior records and their processing results as new evidence storage data units to the blockchain.

2. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: In step 1, the cross-modal feature alignment module adopts a 12-head attention mechanism, with each attention head having a dimension of 64. It generates query vectors, key vectors, and value vectors by performing linear transformations on text feature vectors, image feature vectors, and logistics feature vectors. Then, it calculates cross-modal attention weights by scaling dot product attention, and finally outputs a unified product multimodal feature representation with a dimension of 768.

3. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: In step 1, text feature extraction uses a bidirectional long short-term memory network with a vocabulary size of 50,000, a hidden layer dimension of 512, and a dropout rate of 0.

1. Image feature extraction uses a ResNet50 deep residual network, with the input image size adjusted to 224 pixels by 224 pixels, and the global average pooling feature of the last convolutional layer is extracted with a dimension of 2048. Logistics feature extraction uses a 5-layer temporal convolutional network with a kernel size of 3 and the number of channels being 64, 128, 256, 512, and 512 respectively.

4. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: In step 2, the evidence storage data unit needs to undergo SHA256 hash operation to generate a 32-byte hash value before being written to the blockchain. The consensus algorithm adopts an improved practical Byzantine fault-tolerant algorithm, which requires more than two-thirds of the nodes to reach a consensus before the data can be uploaded to the blockchain. The data upload time delay is controlled within 2 seconds.

5. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: In step 2, the bidirectional index mapping table is implemented using a distributed key-value database, which supports more than 10,000 concurrent read and write operations per second. The index data retains a hot data copy in memory, with a response time of less than 10 milliseconds. At the same time, the index snapshot is persisted to the solid-state drive every hour.

6. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: In step 3, the number of random projection functions for locality-sensitive hashing is set to 256, the Hamming space dimension is 64, and each hash table contains... Each hash bucket is used to build a multi-table index using eight independent hash tables.

7. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: The multi-level anti-counterfeiting judgment in step 4 also includes a second level of judgment. When the cosine similarity is between 0.7 and 0.85, fine-grained feature comparison is initiated to calculate the Mahalanobis distance between the two in the key attribute dimension. When the Mahalanobis distance is greater than 3, an anti-counterfeiting warning is also triggered.

8. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: After the warning signal is generated in step 4, the system automatically generates a visual report containing the product identification code, a list of similar products, similarity values, and evidence comparison results. The warning information is then pushed to the platform's review terminal in real time via a message queue. Reviewers must complete manual review and confirmation within 30 minutes.

9. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: In step 5, the model update adopts an incremental learning strategy. The model fine-tuning is initiated once every 1000 new training samples are accumulated. During training, the basic network parameters are kept fixed, and only the parameters of the last 3 fully connected layers are updated. The learning rate is set to 0.001, and the stochastic gradient descent algorithm with momentum is used for optimization. The training cycle is 50 rounds.

10. The method for identifying counterfeit goods on e-commerce platforms based on blockchain according to claim 1, characterized in that: The closed-loop feedback system in step 5 also includes a performance monitoring module that continuously tracks the accuracy, recall and F1 score of the identification model. When the F1 score drops by more than 5% for three consecutive statistical periods, the model is automatically retrained using all data from the last 30 days.

Images