Determining the Authenticity of Information Uploaded by Blockchain Based on Neural Networks—For Seed Traceability

Abstract

Traditional seed supply chains face several hidden risks. Certain regulatory departments tend to focus primarily on entity circulation while neglecting the origin and accuracy of data in seed quality supervision, resulting in limited precision and low credibility of traceability information related to quality and safety. Blockchain technology offers a systematic solution to key issues such as data source distortion and insufficient regulatory penetration in the seed supply chain by enabling data rights confirmation, tamper-proof traceability, smart contract execution, and multi-node consensus mechanisms. In this study, we developed a system that integrates blockchain and neural networks to provide seed traceability services. When uploading seed traceability information, the neural network models are employed to verify the authenticity of information provided by humans and save the tags on the blockchain. Various neural network architectures, such as Multilayer Perceptron, Recurrent Neural Network, Fully Convolutional Neural Network, and Long Short-term Memory model architectures, have been tested to determine the authenticity of seed traceability information. Among these, the Long Short-term Memory model architecture demonstrated the highest accuracy, with an accuracy rate of 90.65%. The results demonstrated that neural networks have significant research value and potential to assess the authenticity of information in a blockchain. In the application scenario of seed quality traceability, using blockchain and neural networks to determine the authenticity of seed traceability information provides a new solution for seed traceability. This system empowers farmers by providing trustworthy seed quality information, enabling better purchasing decisions and reducing risks from counterfeit or substandard seeds. Furthermore, this mechanism fosters market circulation of certified high-quality seeds, elevates crop yields, and contributes to the sustainable growth of agricultural systems.

1. Introduction

High-quality seeds serve as the cornerstone of agricultural development, directly influencing the productivity and scalability of the planting industry. In the context of growing global food security challenges, high-quality seeds represent a vital factor in enhancing agricultural productivity. The utilization of superior-grade seeds constitutes a significant component for improving agricultural output across all cultivation systems. This factor has gained unprecedented importance in ensuring adequate food security for the world’s expanding population. Additionally, it serves as an effective marketing instrument for boosting potential crop sales, particularly in today’s highly competitive marketplace. Therefore, seed certification is essential for the agricultural industry [1,2]. The seed supply chain involves various manufacturers and regulatory departments. Inefficient management across this chain is largely attributed to inadequate traceability within the seed certification process [3]. This inefficiency can increase the operational costs of the seed certification process. The presence of counterfeit seeds poses significant threats to crop yields and quality in agricultural production, thereby undermining the integrity of the seed market. Extensive research has been conducted on seed traceability systems, including DNA molecular marking technologies [4,5,6,7,8] and Internet of Things (IoT)-based wireless radio frequency technology [9,10,11]. Among these, DNA recognition technologies offer high identification accuracy and biological-level traceability. However, the high costs associated with equipment and instruments present significant barriers to large-scale public adoption. Internet of Things (IoT) and Radio Frequency Identification (RFID) technologies enhance automation level and operational efficiency within traceability systems. Nevertheless, problems such as high initial investment, ongoing maintenance costs, signal interference, and limited coverage remain unresolved. With the advancement of blockchain technology, ensuring data transparency and immutability through decentralization and encryption methods has become a widely discussed topic. Moreover, blockchain deployment costs are low, and the system is easy to use and maintain. Therefore, this is considered an effective method for enhancing the transparency and credibility of traceability systems. In recent years, the integration of blockchain technology into agricultural supply chains has attracted increasing scholarly attention [12,13,14,15,16,17]. Specifically, integrating blockchain technology into traditional seed quality traceability systems has improved the serious centralization of traditional traceability methods, low data credibility, and insecure information storage. Blockchain can significantly reduce both time and financial expenditure in seed supply chain management, thereby reducing costs and improving operational efficiency [18]. Hyperledger Fabric, a licensed blockchain architecture, offers enhanced confidentiality, flexibility, and scalability compared with other alternatives. It supports pluggable implementations of different components and can accommodate complex enterprise applications present in the ecosystem. It is suitable for a range of industrial scenarios such as banking, finance, insurance, healthcare, human resources, and supply chains. For example, El Hajji employed Hyperledger Fabric—based on the builder design pattern—to manage an agri-food supply chain with a focus on maintaining food relationships, authorization, and accurate traceability throughout the supply chain. Performance evaluation using the Caliper tool demonstrated the feasibility and efficiency of the proposed system by measuring network throughput and transaction latency [19]. Similarly, Rehan et al. applied Industrial IoT technologies in conjunction with Hyperledger Fabric to enhance traceability, security, and transparency in supply chain operations [20]. Some traceability systems leverage the security and transparency characteristics of blockchain networks in their architecture. However, manual data entry is often required, which introduces the possibility of human error in the information stored on the blockchain, ultimately affecting the security and authenticity of traceability systems. With the advancement of artificial intelligence, numerous scholars have employed neural networks to process information processing tasks [21,22,23,24,25,26]. Some researchers have specifically applied neural network models to detect anomalous data. For instance, Maram et al. introduced an EfficientNet feedforward neural network (EFFNN) model for detecting ransomware in blockchain systems. They extracted sequence-based statistical features from blockchain data and applied linear normalization. Feature selection was then performed using the Kumar–Hassebrook and Czekanowski similarity metrics to enhance relevant feature recognition. Finally, a hybrid EFNNN model integrating EfficientNet with a deep feedforward neural network was used for detection. This model achieved high performance, with a true negative rate of 0.9600, an accuracy rate of 0.9550, and a true positive rate of 0.8210, demonstrating its effectiveness in identifying ransomware within blockchain systems [27]. Despite these promising results, the model exhibits a core limitation in input modality flexibility. It only processes numerical features and cannot handle non-numerical data types such as text inputs. This limitation significantly undermines the practical effectiveness of the method in real-world anomaly detection scenarios, where comprehensive analysis often requires multimodal data integration. In a separate study, Wang et al. proposed a method for anomaly detection in smart grids. Their approach incorporated a blockchain encoder within an adversarial multi-agent gradient neural network to identify anomalies in a smart grid network. By leveraging edge computing, the system reduced traffic and latency communication by shifting processing, data handling, and services from a centralized cloud to edge servers. Various smart-grid anomaly analysis datasets have been experimentally studied in terms of their prediction accuracy, service quality, scalability, and abnormal detection rate. This model achieved scalability of 89.00%, prediction accuracy of 95.00%, service quality of 92.00%, and an abnormal detection rate of 85.00% [28]. While neural networks show strong potential for anomaly detection, their application in the context of seed traceability remains limited. Furthermore, the neural network architecture proposed by Wang et al. encounters scalability challenges when applied to complex smart grid environments. In real-world deployments, where numerous smart meters generate massive volumes of heterogeneous data streams, the reliance of the model on multiple agents to collaboratively process and analyze data for accurate anomaly detection fundamentally compromises its operational efficiency. The distributed decision-making mechanism introduces substantial computational overhead, resulting in delayed response times, rendering the approach impractical for real-time anomaly detection in operational power grid systems.

To address the limitations identified in previous studies, in this study, an innovative seed traceability system that integrated blockchain and neural network technologies was developed. The blockchain component leveraged its core strengths—decentralization and tamper-resistant data storage—to securely maintain traceability records. Meanwhile, the neural network utilized advanced data processing and intelligent decision-making techniques to verify the validity of traceability information. This clear functional division marks a significant advancement over earlier approaches. Specifically, the proposed system overcomes the single data-type input constraint observed in Maram et al.’s implementation by supporting multiple data types. Furthermore, we enhanced system efficiency by adopting a single-agent framework for blockchain interaction, as opposed to the multi-agent architecture used in Wang’ et al.’s study, which introduced performance bottlenecks. We validated these improvements through extensive system testing, confirming faster response times and greater operational stability.

2. Materials and Methods

2.1. Blockchain

Blockchain is a distributed ledger technology that ensures data immutability, decentralization, and transparency through the combined use of encryption algorithms, consensus mechanisms, and distributed storage technologies. Originally proposed as the foundational technology for Bitcoin, blockchain is now widely applied in fields such as finance, supply chain management, healthcare, and the IoT [29]. Blockchain systems use encryption algorithms to create a decentralized, fraud-proof network that enhances the overall system security and transparency. Transactions in a blockchain network can be stored across distributed devices and form chains in a blockchain-specific distributed ledger, thereby reducing vulnerability to cyberattacks. This innovation process creates an unalterable list of records (known as blocks) connected via cryptographic hashes to ensure the integrity and security of data storage [30,31,32,33].

This study adopts Hyperledger Fabric as the blockchain framework for system architecture. Hyperledger Fabric is a modular, permissioned distributed ledger platform that provides high levels of confidentiality and flexibility. It supports pluggable implementations of different components and can accommodate highly complex applications within a diverse ecosystem [34]. Unlike other blockchains, the Hyperledger Fabric features a uniquely scalable and configurable architecture. Therefore, the front-end and back-end of the traceability system can be built more conveniently using a Hyperledger Fabric. The structure of the Hyperledger Fabric network used in this study is shown in Figure 1 and the pseudocode is shown in

Table 1. Hyperledger Fabric v2.5 Pseudocode.

#Network initialization org1 = Organization(“Org1MSP”, peers = [“peer0.org1.example.com”]) org2 = Organization(“Org2MSP”, peers = [“peer0.org2.example.com”]) channel = Channel(“mychannel”, orderer = “orderer.example.com”, orgs = [org1, org2]) channel.create_genesis_block() #Chaincode deployment cc_package = package_chaincode(“mycc”, “github.com/chaincode”, “go”, “1.0”) peer0.org1.install_chaincode(cc_package) peer0.org2.install_chaincode(cc_package) peer0.org1.approve_chaincode(“mychannel”, “mycc”, “1.0”, “AND(‘Org1MSP.member’,‘Org2MSP.member’)”) peer0.org2.approve_chaincode(“mychannel”, “mycc”, “1.0”, “AND(‘Org1MSP.member’,‘Org2MSP.member’)”) peer0.org1.commit_chaincode(“mychannel”, “mycc”)

#transaction proposal = client.org1.new_proposal( channel = “mychannel”, chaincode = “mycc”, fcn = “TransferAsset”, args = [“asset1”, “org1”, “org2”], peers = [peer0.org1, peer0.org2] ) endorsement1 = peer0.org1.endorse(proposal) endorsement2 = peer0.org2.endorse(proposal) transaction = client.org1.send_to_orderer([endorsement1, endorsement2], “orderer.example.com”) block = orderer.deliver_block(“mychannel”, transaction) peer0.org1.commit_block(block) peer0.org2.commit_block(block) #query response = peer0.org1.query( channel = “mychannel”, chaincode = “mycc”, fcn = “GetAsset”, args = [“asset1”] ) print(response)

.

A Hyperledger Fabric network consists of multiple components, each with a specific functionality. A node (peer) is the core component of the network and is responsible for maintaining the ledger and executing the chaincode. There are two types of peer nodes: a committing node (committing peer), which verifies the transactions and writes them to the ledger, and an endorsing node (endorsing peer), which executes the chaincode and endorses the transaction. Chaincode, the smart contract mechanism of Hyperledger Fabric, is used to implement the business logic. The chaincode can be written in multiple programming languages (such as Go and JavaScript) and is executed on peer nodes. The ledger in Fabric comprises two parts: the World State, which stores the current state using key-value pair databases such as LevelDB and CouchDB, and the Blockchain, which stores immutable, historical records of all transactions. Fabric also supports channels, which are private communication subnets that allow different organizations to create independent communication and data storage spaces within the same network. Data within a channel is only accessible to the channel members. Membership Service providers (MSPs) manage the identities and permissions across a network. In the context of network security, MSPs can assist in managing digital identities and permissions by leveraging Microsoft’s identity solutions, such as Azure Active Directory (Azure AD) and Public Key Infrastructure (PKI). To ensure secure access, MSP utilizes certificate-based authentication to verify the identity of nodes and users within a network. The Certificate Authority (CA), a core component of PKI, issues and manages these certificates, ensuring that only authorized entities can participate in the network.

The Hyperledger Fabric network used in this study consists of two organizations, each with a peer node and a local CA. These organizations share a sorting service cluster (orderer cluster). The peer node of Organization 1 is responsible for endorsement, submission, gossip communication, and related functions. Its local certificate organization (CA_1) operates as part of the identity management organization of the network. The peer node of Organization 2 performs symmetrical functions to the production node, and its local certificate organization (CA_2) is responsible for independently issuing identity credentials. The orderer cluster adopts the Raft consensus protocol [35] to ensure that the transaction sequence is globally consistent. Each organization manages its members through a local CA and uses X.509 certificates and elliptic curve cryptography to ensure security.

2.2. Neural Network Architecture

This experiment employed four neural network architectures to evaluate the authenticity of seed traceability information: Multilayer Perceptron (MLP), Fully Convolutional Neural Network (FCNN), Recurrent Neural Network (RNN), and Long Short-term Memory (LSTM). Considering that seed traceability information includes both numerical and textual information, we used a multi-feature fusion method, as shown in Figure 2, to construct neural network models. This method merges text feature columns, such as seed origin and seed maturity period, into a string, which is then vectorized using TfidfVectorizer. TfidfVectorizer converts the text into a Term Frequency–Inverse Document Frequency ($TF-IDF)$ Weight matrix, which in turn converts text information into a numerical representation. The $TF-IDF$ matrix of numerical and text features is spliced together in the column direction to form a new feature matrix that contains all the features used for training. This method is preferred over simpler techniques such as CountVectorizer because it balances word frequency and importance across the dataset, resulting in more informative feature representations [36]. The following paragraph explains how to combine text and numerical features.

We have two text feature columns, namely seed origin and seed category. First, we need to merge these two text features into a string. Suppose we have n samples, and the text characteristics of each sample are $t_i\left(i=\mathrm{1,2},...,n\right)$, then these text features can be merged into a string list $T=[t_1,t_2,...,t_n]$. Next, use TfidfVectorizer to convert text features to a $TF-IDF$ weight matrix. $TF-IDF$ is a commonly used text vectorisation method that considers Term Frequency $\left(TF\right)$ and Inverse Document Frequency $\left(IDF\right)$.

$TF$refers to the frequency with which a word appears in a document. Assume that the frequency of word ω appears in the article $d$ is $f_{\omega ,d}$, then the word frequency $TF(\omega ,d)$ can be expressed as follows:

$$TF(\omega ,d)={\displaystyle \frac{f_{\omega ,d}}{{\sum }_{{\omega }^{\prime }\in d}{f}_{\omega ,d}}}$$

(1)

The $IDF$ measures the importance of words in the corpus. Assuming that N documents are present in the corpus and the word ω appears in $n_{\omega }$ documents, the $IDF\left(\omega \right)$ can be expressed as follows:

$$IDF(\omega )=\log {\displaystyle \frac{N}{n_{\omega }}}$$

(2)

$TF-IDF$ is the product of word frequency and inverse document frequency:

$$TF-IDF(\omega ,d)=TF(\omega ,d)\times IDF\left(\omega \right)$$

(3)

After using TfidfVectorizer, the text feature $T$ is converted into a $TF-IDF$ weight matrix $X_{text}$, where each row corresponds to a sample and each column corresponds to a word’s $TF-IDF$ value.

Suppose we have m numerical features, and the numerical features of each sample can be represented as a vector $x_i=\left[x_{i1},x_{i2},...,x_{im}\right]$. The numerical characteristics of all samples can be represented as a matrix $X_{num}$, where each row corresponds to a sample and each column corresponds to a numerical characteristics. To combine text and numerical features, we need to splice the $TF-IDF$ matrix $X_{text}$ and the numerical feature matrix $X_{num}$ in the column direction to form a new feature matrix $X$. Assuming that $X_{text}$ is an $n\times p$ matrix ($n$samples, $p$features) and $X_{num}$ is an $n\times m$ matrix ($n$samples, $m$features), then the spliced feature matrix $X$ is an $n\times \left(p+m\right)$ matrix is given as follows:

$$X=\left[\begin{array}{cc}{X}_{text}& X_{num}\end{array}\right]$$

(4)

Each row of $x$ corresponds to a complete feature vector for one sample; the first column $p$ denotes the $TF-IDF$ value of the text feature, and the last column $m$ denotes the numerical feature. The final feature matrix $x$contains all the features used for training, which can be used as the input for model training and prediction, enabling each neural network to learn from multimodal data representations. This matrix combines textual and numerical information to better represent the characteristics of a sample.

The structure diagram of each neural network is shown in Figure 3.

MLP: This model is divided into input, dense, and output layers. The neural network is configured with a single dense layer containing 10 neurons, trained using the Adam optimizer with a learning rate of 0.0001 over 200 epochs. To enhance generalization, a dropout rate of 0.2 is applied, and training is conducted in batches of 100 samples. For reproducibility, a random seed of 42 is fixed during data splitting and SMOTE resampling.

The data collection model parameters encompass the preprocessing pipeline for both numerical features (original weight, germination rate, moisture content, cleanliness, purity, shelf life, seeding rate per mu, and price) and textual features (maturity period). Numerical features were normalized using MinMaxScaler, while textual features were processed into 50-dimensional TF-IDF vectors through TfidfVectorizer. To address class imbalance, SMOTE oversampling was applied, ultimately generating a 58-dimensional hybrid feature vector (8 numerical + 50 textual).

For validation protocols, stratified sampling was employed to partition the dataset into training (80%), validation (10%), and test sets (10%). A custom callback function was implemented to monitor F1-score, recall, and other metrics in real-time during training. The model’s performance was rigorously evaluated through confusion matrix analysis and four key metrics (accuracy, precision, recall, and F1-score) on the independent test set.

FCNN: The model architecture consists of a fully connected neural network with carefully tuned hyperparameters to optimize performance. Key configuration parameters include a learning rate of 0.001 for the Adam optimizer, training over 200 epochs with a batch size of 64 samples, and a dropout rate of 0.1 to prevent overfitting. To ensure reproducible results, all random processes, including data splitting and SMOTE-based class balancing, are initialized with a fixed seed value of 42.

The data collection model parameters incorporate eight numerical agricultural metrics (original weight, germination rate, moisture content, cleanliness, purity, shelf life, seeding rate per mu, and price) along with the textual maturity period feature, which undergoes comprehensive preprocessing including MinMax normalization for numerical data and TF-IDF vectorization (50-dimensional representation) for textual features, resulting in a 58-dimensional hybrid feature space after concatenation, with missing values addressed through complete-case analysis and class imbalance mitigated via SMOTE oversampling.

For validation protocols, the framework implements stratified data partitioning (80% training, 10% validation, 10% test sets) with rigorous evaluation through a custom callback mechanism that tracks key performance metrics (accuracy, precision, recall, F1-score) at strategic training intervals, while the model architecture itself—a hybrid FCNN structure combining 1D convolutional layers with dense connections and dropout regularization (rate = 0.1)—undergoes comprehensive assessment through epoch-wise loss visualization and final confusion matrix analysis on the held-out test set, ensuring robust evaluation of both training dynamics and ultimate predictive performance.

RNN: This recurrent neural network implementation utilizes a sequential architecture beginning with two SimpleRNN layers (64 and 32 units, respectively), where the first layer returns full sequences to maintain temporal information for subsequent processing. The network then transitions to a dense layer with 16 units employing ReLU activation, with strategic dropout regularization (rates of 0.3 and 0.2) implemented after each recurrent layer to enhance model generalization. Training employs the Adam optimizer configured with a conservative learning rate of 0.0005 to ensure stable convergence, optimizing binary cross-entropy loss across 200 epochs of batch processing (size = 32). All stochastic processes, including data partitioning and SMOTE-based class balancing, maintain strict reproducibility through a fixed random seed (42).

The data collection model parameters encompass eight core agricultural metrics (original weight, germination rate, moisture content, cleanliness, purity, shelf Life, seeding rate per mu, price) along with textual maturity period data, which undergo rigorous preprocessing including MinMax normalization for numerical features and TF-IDF vectorization (50-dimensional representation) for textual data, yielding a 58-dimensional hybrid feature space. Missing values are addressed through complete-case analysis, while class imbalance is mitigated via SMOTE oversampling (random state = 42).

For validation protocols, the study implements a stratified data partitioning scheme (80% training, 10% validation, 10% test) and employs a stacked RNN architecture with two recurrent layers (64/32 units) and dropout regularization (0.3/0.2 rates). Model performance is rigorously evaluated through a comprehensive validation framework incorporating epoch-wise tracking of multiple metrics (accuracy, precision, recall, F1-score), detailed loss curve analysis, and final assessment via confusion matrix on the held-out test set, with all processes standardized using fixed random seeds (42) to ensure reproducibility. The Adam optimizer (lr = 0.0005) with binary cross-entropy loss drives model training over 200 epochs (batch size = 32), with validation metrics monitored through custom callbacks at strategic intervals to ensure robust evaluation of both training dynamics and predictive performance.

LSTM: The model features a dual-layer structure with 64 units in the initial LSTM layer (configured to return full sequences) and 32 units in the subsequent layer, creating a hierarchical temporal feature extraction pipeline. To combat overfitting, the architecture incorporates L2 regularization (λ = 0.01) coupled with a 0.3 dropout rate, while batch normalization following the first LSTM layer ensures stable gradient flow during training. The output layer employs a single sigmoid-activated neuron for probabilistic binary classification. All random processes maintain strict reproducibility through a fixed seed value (42), including data partitioning and SMOTE-based class balancing operations.

The data collection model parameters incorporate eight quantitative seed metrics (original weight, germination rate, moisture content, cleanliness, purity, shelf life, seeding rate, and price), alongside textual maturity period data processed through TF-IDF vectorization (50-dimensional representation). Numerical features undergo MinMax scaling before weight application, with missing values handled through complete-case analysis and class imbalance addressed via SMOTE oversampling.

The validation protocol employs stratified data partitioning (80% training, 10% validation, 10% test) with comprehensive monitoring through early stopping (patience = 5), learning rate reduction (factor = 0.5), and a custom callback tracking F1-score. Model architecture features two LSTM layers (64/32 units) with L2 regularization (λ = 0.01), batch normalization, and dropout (0.3), trained using Adam optimization (lr = 0.001) over 200 epochs (batch size = 128). Final evaluation incorporates multiple metrics (accuracy, precision, recall, F1, AUC-ROC) on the independent test set, with all preprocessing objects and model weights preserved for deployment. The framework demonstrates particular attention to feature engineering through differential weighting and maintains rigorous validation standards through its callback system and comprehensive metric suite.

2.3. Deployment and Integration of Neural Networks and Blockchains

Considering the operational characteristics of the seed source traceability business and aligning with actual production needs, herein, we designed a source traceability system, as shown in Figure 4. Seed companies are divided into five departments: breeding, propagation, processing, transportation, and sales. Each department is assigned distinct login credentials to ensure role-based access control. Upon logging into the system using department-specific credentials, authorized users can input seed information. This data is first evaluated by the neural network model to assess its authenticity. Verified data is subsequently stored in the traceability system and securely saved in the blockchain ledger. When a company department initiates a query, the system interacts with the blockchain to retrieve endorsed transactions, enabling access to perform a traceability information query. When logging in with a consumer identity, traceability information is screened to display only public information from the processing and sales departments, thereby safeguarding sensitive operational data while maintaining transparency for end users.

When the breeding department inputs breeding information, the neural network assesses the authenticity of the input information and assigns an accuracy label. A unique traceability code is then generated using the Snowflake algorithm. This code is associated with the information record and remains consistent across all subsequent inputs from different departments. The information input by each user type matches the traceability code. The breeding department transfers the seeds to the propagation department, and the propagation department inputs the propagation information according to the actual planting and propagation situation. Thereafter, the propagation seeds are handed over to the processing department to process the seeds and input the processing information according to the processing situation. Subsequently, the processed seeds are transported by the transportation department, and the transportation department inputs the transportation information according to the actual situation. When the seeds are transported to the sales department, the sales information is input according to the sales situation. For the breeding process, seeds cultivated in the laboratory are recorded, and the breeding department logs into the system with their corresponding identity, inputs the corresponding information, and submits it. The system automatically generates a unique traceability code using the Snowflake algorithm. All subsequent information corresponding to this source is traced through this traceability code. However, for seed companies, for the benefit of the department, consumers do not know all their traceability information. Therefore, some public information is screened for traceability and query by consumers; the public information is stored in a QR code, and the QR code is printed and posted on the seed wrapping paper. Consumers can scan QR codes to query for public information.

The front-end portal of the system is based on the TCP/IP protocol and developed using JavaScript, HTML, and CSS, with Vue.js templates used to create the browser visualization operation window of the system. Through the front-end data management website of the browser, the API interface provided by the call server is connected to the back-end database to perform user registration, permission management, information upload, and query functions. All users trigger user registration commands on the registration and login interface of the public front-end of the website, send requests, and receive the blockchain address along with public and private keys sent by the back-end.

The workflow diagram of the system is shown in Figure 5. The neural network layer verifies the on-chain information. The neural network model is deployed on Flask and communicates with the front-end platform through an API. After logging in with a specific account type, inputting the corresponding information, and uploading it, the information from the blockchain front-end data is transmitted to the locally deployed neural network Flask interface through the API interface. The neural network model in Flask automatically reads the form data and serves as input to the network, automatically determines authenticity, and outputs authenticity indicators. Finally, the output authenticity indicators are transmitted back to the blockchain front-end through the API interface. After receiving the data, the front end inputs it into the blockchain network through the back end and saves the accuracy labels in the blockchain.

3. Results

In evaluating the performance of the neural network models described in Section 2.2 Neural Network Architecture, we selected the following four widely used classification metrics: accuracy, precision, recall, and F1-score. These metrics are based on four basic values in the confusion matrix, which includes the following elements: true positive (TP), the number of samples correctly predicted by the model as a positive class; true negative (TN), the number of samples correctly predicted by the model as a negative class; false positive (FP), the number of samples incorrectly predicted by the model as a positive class (actually a negative class); and false negative (FN), the number of samples incorrectly predicted by the model as a negative class (actually a positive class).

Accuracy (5) is the proportion of the number of correct samples to the total number of samples predicted by the model. This measures the overall prediction ability of the model, and the range is [0, 1]. The closer the value is to 1, the better the overall prediction effect of the model. Precision (6) is the proportion of samples correctly predicted by the model as a positive class, which measures the prediction accuracy of a model, and the range is [0, 1]. The closer the value is to 1, the more accurate the prediction of the positive class by the model. The recall rate (7) is the proportion of samples predicted by the model as a positive class, measuring the coverage ability of the model. The range is [0, 1]. The closer the value is to 1, the stronger the coverage ability for the positive class. The F1-score (8) is the harmonic average of accuracy and recall, which is used to balance the accuracy and recall rate. The range is [0, 1]. The closer the value is to 1, the better the comprehensive performance of the model. The formula is as follows:

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(5)

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(6)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(7)

$$F1-score=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(8)

However, none of the models achieved perfect (100.00%) accuracy. Therefore, in the actual seed traceability process, the authenticity of seed quality information cannot be regarded as a binary classification problem. If some seeds are identified as false information by the neural networks owing to their unique biochemical indicators, they will not meet the actual requirements. Therefore, the output indicator is designated as a quantitative indicator, such that 90.00% represents real information and 3.00% represents false information, which can provide users with a quantitative determination standard, significantly avoiding the consequences of model accuracy caused by the determination error. During the evaluation stage, the accuracy labels that are more than 50.00% of the model are classified as samples predicted as correct, the accuracy labels that are less than 50.00% of the model are classified as samples predicted as incorrect, and the final statistical prediction success rate is calculated.

We searched and recorded the basic information of 30,000 real seeds in the seed online store https://wx.zhongziku.cc/b2b2c/goods_list.html?is_buy=1#3 (accessed on 11 December 2024). In parallel, we randomly generated the information of 30,000 fake seeds according to these data ranges and combined the 60,000 pieces of data into a training dataset. Regarding the construction of the training dataset, the authentic samples were collected directly from reputable online marketplaces, whereas synthetic fraudulent samples were generated by introducing randomised perturbations within the statistically valid ranges of authentic data parameters. This generation strategy was designed to simulate sophisticated commercial fraud patterns, as the synthetic samples exhibit such high fidelity that they are virtually indistinguishable from genuine entries by human evaluators, thereby accurately replicating the deceptive practices employed by dishonest merchants when uploading falsified product information. The detailed specifications of the collected seed information, along with the corresponding parameter ranges used in generating synthetic fraudulent samples, are systematically presented in

Table 2. Summary of dataset information.

Seed Information Fields	Synthetic Fraudulent Data Generation Ranges	Seed Information Fields	Synthetic Fraudulent Data Generation Ranges
Seed Name	Random selection from all available seed names	Germination Rate (%)	80–95
Original Weight (g)	1–100	Moisture Content (%)	1–12
Seeds per Bag	50–2000	Shelf Life (years)	1–2
Maturity Period	Random selection from all available maturity periods	Origin	Random selection from all available origins
Purity (%)	92–98	Seeding Rate per Mu (bag)	1–100
Cleanliness (%)	92–98	Price (RMB)	1–300

.

The calculation formula for the actual test accuracy is as follows

$$Actualtestaccuracy=X/Y$$

(9)

• X—Number of correctly predicted samples by the model
• Y—Total number of samples predicted by the model

Our experimental framework was implemented using Python (3.8.19) on the PyCharm (2024.2.2) platform with a Conda virtual environment, where we developed all model architectures mentioned in Section 2.2. The seed dataset was systematically organized into Excel spreadsheets and automatically processed through our Python pipeline, which handled data loading, preprocessing, training, validation, and testing phases. Based on the training results shown in Figure 6, the multimodal feature fusion model based on the LSTM architecture exhibits the best performance in processing seed traceability information. Specifically, the model yields an accuracy of 0.9780, precision of 0.9964, recall of 0.9945, F1-score of 0.9955, and actual test accuracy of 0.9672. These results demonstrate that, compared with other models, LSTM can more effectively capture the relationships between various seed characteristic features and better predict the authenticity of seed information. From the perspective of the loss function trends shown in Figure 7, the LSTM model demonstrates significantly better convergence compared with the other models. Its final training and validation loss values at the end of the training are 0.0016 and 0.0007, respectively. Based on the above training results, it can be observed that LSTM-based neural networks have better processing effects on this problem.

The seed traceability platform was constructed using a comprehensive full-stack architecture. The technology stack of the system included FabricV2.5 for blockchain infrastructure, the Gin framework for the server-side back end, Vue.js for the front-end interface, MySQL for relational data storage, and Flask for serving machine learning models. After deploying the platform code to the designate server environment, the system was launched via the command line to complete the deployment process. Figure 8 shows the user interface of the deployed platform. The online mall collected 300 real seed data points and used the random function of Excel to build 300 fake datasets. Seed information was manually entered into the system individually by the seed processing department, and each entry was evaluated in real-time by the integrated neural network module. The accuracy labels returned by the neural network were calculated as real data, and accuracy labels that were less than 50.00% were labelled as fraudulent. Finally, accuracy statistics were obtained for the actual entry situation, and the accuracy obtained by the final test was 90.65%.

After validating the accuracy of the system, we proceeded to evaluate the performance of the blockchain network by testing it under various simulated conditions and workload intensities. A primary objective was to measure transaction processing times, which helps determine the scalability and operational efficiency of the system across different scenarios. These evaluations yielded critical insights into how the architecture performs under different demand levels. To benchmark the performance of the Hyperledger Fabric’s network, we employed the Caliper benchmarking tool [37]. Caliper enables performance testing by generating diverse workload simulations and operational scenarios [38]. It tracks multiple performance indicators, including transaction processing speed (TPS), system throughput, and network latency [39]. Throughput quantifies the rate of successfully executed transactions or query operations each second. Network latency, on the other hand, represents the total duration for transaction completion, encompassing consensus mechanisms, validation procedures, and final confirmation [40]. These metrics are calculated as follows:

$$\mathrm{Throughput}(\mathrm{TPS})={\displaystyle \frac{\mathrm{Number}\mathrm{of}\mathrm{Transactions}}{(\mathrm{last}\mathrm{committing}\mathrm{time}-\mathrm{first}\mathrm{submitting}\mathrm{time})}}$$

(10)

$$\mathrm{Send}\mathrm{Rate}(\mathrm{TPS})={\displaystyle \frac{\left(\mathrm{Succ}+\mathrm{Fail}\right)}{(\mathrm{last}\mathrm{submitting}\mathrm{time}-\mathrm{first}\mathrm{submitting}\mathrm{time})}}$$

(11)

$$\mathrm{Transaction}\mathrm{Latency}={\displaystyle \frac{\mathrm{Number}\mathrm{of}\mathrm{Transactions}}{\mathrm{Confirmation}\mathrm{time}-\mathrm{submit}\mathrm{time}}}$$

(12)

This performance test was conducted on the Tencent Cloud server, which has a server configuration of a 2-core CPU, 4 GB of memory, and the system environment is Ubuntu 20.04. To test the stability of the system against transaction loads, the study conducted experiments with different transaction volumes, specifically 10, 30, 50, and 100 transactions, which were sent to the network at a rate of 10 transactions per second (TPS). The corresponding results are shown in

Table 3. Summary of performance metrics.

Name	Succ	Fail	Send Rate (TPS)	Max Latency (s)	Min Latency (s)	Avg Latency (s)	Throughput (TPS)
Create a commodity with 10 tx.	10	0	10.8	0.5	0.1	0.3	10.0
Create a commodity with 30 tx.	30	0	10.6	0.8	0.2	0.4	9.8
Create a commodity with 50 tx.	50	0	10.3	1.5	0.3	0.7	9.5
Create a commodity with 100 tx.	100	0	10.1	2.8	0.4	1.2	9.1

. In all test scenarios, the system maintains throughput close to the target send rate (10 TPS), demonstrating good stability at low to moderate loads (10–50 transactions). At 100 transactions, the throughput drops slightly to 9.1 TPS, indicating that the system may experience a slight performance degradation at high transaction volumes but still maintain high processing power. In all tests, the number of failures is 0, indicating that the system is highly reliable under a load of 10 TPS and is able to process all submitted transactions in their entirety. The system exhibits good stability and reliability under a constant load of 10 TPS and is able to efficiently handle the size of 10–50 transactions with reasonable latency. However, at 100 transactions, the maximum latency rises significantly, indicating that the system may face performance challenges under higher loads or longer transaction durations. This latency trend suggests that future optimizations should consider increasing memory capacity or optimizing the transaction caching mechanism to improve latency performance in high-load scenarios.

4. Discussion

In this study, we implemented a blockchain–neural network integration approach to enhance seed traceability. Traditional traceability systems typically operate under centralized architectures and rely heavily on physical record-keeping. However, such centralized models may expose traceability processes to the risk of data tampering or data loss. By contrast, the blockchain architecture employed in this research leveraged distributed data storage, and this fundamental change substantially improved system traceability and security. On top of the secure blockchain foundation, neural network algorithms were employed to verify the authenticity of input data. This experiment used multi-feature fusion networks based on MLP, RNN, FCNN, and LSTM to verify the authenticity of seed source information. Experimental results using the same dataset showed classification accuracies of 89.05% for MLP, 89.34% for RNN, 94.28% for FCNN, and 96. 72% for LSTM. These results suggested that the LSTM model performed best in identifying authentic seed information. This performance advantage can be attributed to the architectural design of the LSTM, which was significantly better than those of the MLP, RNN, and FCNN in capturing long-term dependencies, mitigating gradient vanishing issues, and dynamic feature selection. In contrast, the FCNN and MLP architectures independently connected each input node to an output node in fully connected layers. With this design, the parameter quantity increased with the square order of the input dimension, making these models prone to overfitting and less effective in modeling sequential dependencies due to the absence of structural priors. LSTM networks, by comparison, dynamically adjusted the information flow through a gating mechanism, which can also learn more complex input–output mapping relationships and have stronger nonlinear expression capabilities. These results are consistent with those of Al-Selwi et al., who demonstrated that with the introduction of a gated mechanism, LSTM effectively solved the gradient disappearance and explosion problems encountered by traditional RNNs in long sequence training [41]. Furthermore, our observations align with the results of Das et al., who used different models to compare the final accuracy, precision, and F1 scores for varying percentages of noisy data in their study. They found that as the noise level increased, the prediction accuracy provided by the CNN and MLP algorithms decreased, whereas LSTM models maintained over 80% accuracy even at a 10% noise level [42]. We believe that these results are a significant success, demonstrating the substantial potential of multi-feature fusion models based on LSTM in detecting information anomalies. In particular, the proposed approach aims to authenticate blockchain data through the support of neural network technology, further ensuring the security, transparency, and authenticity of the traceability system to foster stronger confidence among stakeholders in the agricultural seed supply chain and trust of participants. This study presents significant opportunities to transform agricultural practices and inform policy improvements. For agricultural production, this enhanced traceability framework can help combat the widespread issue of counterfeit or low-quality seeds, which undermine crop yields and farmer livelihoods. At the policy level, the immutable records generated by this system could strengthen regulatory oversight by providing auditable evidence for seed certification and quality control. Agricultural ministries could leverage such data to identify regions with recurring seed quality issues, enabling targeted interventions. The transparency afforded by blockchain-based traceability also aligns with global agricultural trade requirements, potentially facilitating smoother export processes for seed producers. By reducing information asymmetries across the seed supply chain, this technology integration contributes to fairer markets, more effective agricultural extension services, and ultimately, more resilient food systems.

While this research proposes an innovative and reliable solution for seed traceability, several limitations warrant consideration and suggest avenues for future improvement. First, the current neural network model was trained on datasets collected from online marketplaces, where synthetic anomalies were generated based on statistical patterns derived from commercial seed data. We acknowledge the importance of collaborating with seed supply chain partners to obtain authentic, enterprise-level datasets. Access to such real-world data would enable model retraining and facilitate practical performance evaluation in real-world traceability processes.

Second, the available dataset was limited to breeding-stage parameters. Extending the verification capability of the system to encompass the entire supply chain would enable more effective utilization of the temporal processing strengths of LSTM and potentially enhance overall accuracy. However, acquiring comprehensive, multi-stage agricultural datasets remains a significant challenge.

In terms of technical improvements, we plan to explore alternative neural architectures, such as transformer-based models, through rigorous benchmarking against our current LSTM implementation. To improve the system’s scalability, we intend to explore incorporating specialized models tailored to different supply chain segments while preserving the integrity of blockchain-based verification.

Regarding data privacy, we recognize data privacy protection as a crucial area requiring further improvement in our seed traceability system. Although the current blockchain-based framework provides robust data integrity verification, its inherent transparency poses challenges in protecting sensitive business information across the agricultural supply chain. Our research will focus on developing a novel privacy-preserving paradigm that harmonizes the conflicting demands of traceability and confidentiality. This involves creating adaptive cryptographic protocols capable of dynamically adjusting privacy safeguards in response to data sensitivity levels, without compromising the core verification capabilities of the system.

5. Conclusions

This study presented an innovative system for determining the authenticity of seed traceability information by integrating blockchain with deep learning technology. By combining the immutable nature of distributed ledgers with the intelligent decision-making capabilities of neural networks, the proposed system ensured both secure data storage and reliable authenticity verification across the entire seed lifecycle. A multimodal neural network architecture was used to verify the authenticity of the seed traceability information manually filled dialectically. Among the evaluated models, the experimental results showed that the verification model based on LSTM exhibited optimal performance, with an authenticity determination accuracy of 90.65%. The findings prove that neural networks, particularly LSTM-based architectures, have significant research value and potential for determining the authenticity of information on blockchains. This integration offers a promising pathway for enhancing the reliability, transparency, and trustworthiness of seed traceability systems, thereby contributing a novel technological solution to the agricultural supply chain. The integration of blockchain and deep learning technologies for seed traceability verification holds substantial promise for transforming agricultural systems through multiple dimensions. By ensuring the authenticity and reliability of seed provenance data, this approach directly contributes to enhanced agricultural productivity by mitigating the risks associated with counterfeit or misrepresented seeds that often lead to suboptimal crop performance. The technological framework also fosters more equitable market conditions by empowering farmers with verifiable information about seed quality and characteristics while providing regulators with robust tools for quality control and policy implementation. These combined effects ultimately contribute to strengthening food system resilience and supporting the transition toward more sustainable intensification of agricultural production.