Cross-chain transaction tracking method and system based on large language model

Abstract

The invention discloses a cross-chain transaction tracking method and system based on a large language model, and solves the technical problem of low accuracy of withdrawal transaction pairing caused by an existing cross-chain transaction tracking method. The method comprises the following steps: acquiring original data of a plurality of source chain transactions, and decoding the original data of each source chain transaction to generate basic structure information of each source chain transaction; carrying out deposit transaction identification on the basic structure information of each source chain transaction by adopting a large language model for deposit transaction identification, and outputting a plurality of source chain deposit transactions; based on the basic structure information corresponding to each source chain deposit transaction, determining a target chain withdrawal transaction candidate set corresponding to each source chain deposit transaction; and performing withdrawal transaction pairing by adopting a large language model for withdrawal transaction pairing according to each source chain deposit transaction and the target chain withdrawal transaction candidate set corresponding to each source chain deposit transaction, and outputting a target chain withdrawal transaction matching item.

Description

Technical Field

[0001] This invention relates to the field of transaction detection technology, and in particular to a cross-chain transaction tracking method and system based on a large language model. Background Technology

[0002] With the booming development of DeFi (Decentralized Finance), diverse asset ecosystems have emerged across different public blockchains, making the flow of assets between chains a core need. To achieve this goal, cross-chain bridges have gradually become the most critical infrastructure in multi-chain ecosystems, functioning to securely and efficiently transfer assets between different blockchains. Currently, mainstream cross-chain bridges include CelercBridge, Multichain, and Polygon Bridge.

[0003] Cross-chain operations typically follow a two-step model: "deposit / lock on the source chain" and "withdraw / mint / release on the target chain." This means cross-chain transfers are completed by locking assets on the source chain and releasing or minting mapped assets on the target chain. However, the core challenge of cross-chain tracing lies in the high heterogeneity between multi-chain systems: different bridging protocols have vastly different contract structures, event names, and parameter semantics; there is no native hash reference between the source and target chains during the cross-chain process; transaction confirmation suffers from time delays, and on-chain logs often suffer from missing fields and non-standardized formats.

[0004] These factors render cross-chain pairing methods that rely solely on field matching or fixed rules inaccurate and unreliable in practical applications. Meanwhile, with the increasing scale and security risks of cross-chain assets, security audits, compliance regulations, and source tracing analysis place higher demands on the traceability of cross-chain activities. For example, accurate reconstruction of cross-chain paths is crucial for identifying cross-chain arbitrage, fraud, and money laundering. Therefore, cross-chain tracking methods must not only possess accuracy but also robustness and explainability.

[0005] The withdrawal matching process in existing cross-chain transaction tracking methods is essentially still a mechanical logic of "feature field → rule comparison". Once a key field is missing (such as the source chain transaction missing dstChainId, the target chain transaction missing amount), or the log structure changes due to protocol upgrades (such as Multichain changing the "anySwapOut" event to "anySwapOutUnderlying"), its preset field matching rules become invalid, resulting in low accuracy of withdrawal transaction matching. Summary of the Invention

[0006] This invention provides a cross-chain transaction tracking method and system based on a large language model, which solves the technical problem of low accuracy in matching withdrawal transactions caused by existing cross-chain transaction tracking methods.

[0007] The first aspect of this invention provides a cross-chain transaction tracking method based on a large language model, comprising:

[0008] Obtain the raw data of multiple source chain transactions, decode the raw data of each source chain transaction, and generate the basic structure information of each source chain transaction;

[0009] A large language model for deposit transaction identification is used to identify deposit transactions based on the basic structural information of each source chain transaction, and outputs multiple source chain deposit transactions.

[0010] Based on the basic structural information corresponding to each source chain deposit transaction, a candidate set of target chain withdrawal transactions corresponding to each source chain deposit transaction is determined.

[0011] A large language model for withdrawal transaction matching is used to match withdrawal transactions based on the source chain deposit transactions and the corresponding target chain withdrawal transaction candidate set, and outputs the target chain withdrawal transaction matching item.

[0012] Optionally, the step of using a large language model for deposit transaction identification to identify deposit transactions based on the basic structural information of each source chain transaction, and outputting multiple source chain deposit transactions, including:

[0013] The basic structural information of each source chain transaction is converted into English prompts according to a predefined fixed prompt template.

[0014] Input the English prompts corresponding to each of the source chain transactions into the large language model used for deposit transaction identification, and output the judgment results corresponding to each of the source chain transactions;

[0015] Based on the judgment results corresponding to each source chain transaction, deposit transactions are screened for each source chain transaction to identify multiple source chain deposit transactions.

[0016] Optionally, determining the target chain withdrawal transaction candidate set corresponding to each source chain deposit transaction based on the infrastructure information corresponding to each source chain deposit transaction includes:

[0017] Based on the basic structure information corresponding to each source chain deposit transaction, obtain the target chain address information corresponding to each source chain deposit transaction;

[0018] Each source chain deposit transaction is associated with a target chain address and a withdrawal transaction is crawled to generate a candidate set of target chain withdrawal transactions for each source chain deposit transaction.

[0019] Optionally, the target chain withdrawal transaction candidate set includes multiple target chain withdrawal transactions; the use of a large language model for withdrawal transaction matching performs withdrawal transaction matching based on each source chain deposit transaction and the target chain withdrawal transaction candidate set corresponding to each source chain deposit transaction, outputting target chain withdrawal transaction matching items, including:

[0020] The original data of multiple target chain withdrawal transactions corresponding to each source chain deposit transaction are decoded, and the basic structure information of multiple target chain withdrawal transactions corresponding to each source chain deposit transaction is output.

[0021] Based on the basic structural information of each source chain deposit transaction and the basic structural information of multiple target chain withdrawal transactions corresponding to each source chain deposit transaction, multiple transaction splicing texts are determined.

[0022] The large language model used for withdrawal transaction matching outputs the target chain withdrawal transaction matching item based on the concatenated text of multiple transactions.

[0023] Optionally, the step of determining multiple transaction concatenation texts based on the basic structure information of each source chain deposit transaction and the basic structure information of multiple target chain withdrawal transactions corresponding to each source chain deposit transaction includes:

[0024] Select a single candidate source chain deposit transaction from each of the aforementioned source chain deposit transactions;

[0025] Based on the basic structural information of a single candidate source chain deposit transaction, multiple target chain candidate withdrawal transactions are selected from the multiple target chain withdrawal transactions corresponding to the single candidate source chain deposit transaction.

[0026] According to the predefined English prompt word template, the basic structure information of the single candidate cross-chain deposit transaction is concatenated and combined with the basic structure information of the corresponding candidate withdrawal transaction of the target chain to generate multiple transaction concatenation texts.

[0027] Optionally, the step of using the large language model for withdrawal transaction matching to output target chain withdrawal transaction matching items based on multiple transaction concatenation texts includes:

[0028] Each of the concatenated transaction texts is input into the large language model used for withdrawal transaction matching, and the matching score corresponding to each of the concatenated transaction texts is output.

[0029] Among the matching scores corresponding to each of the transaction concatenation texts, the target chain candidate withdrawal transaction with the highest matching score is selected as the target chain withdrawal transaction matching item corresponding to the single candidate cross-chain deposit transaction.

[0030] A second aspect of this invention provides a cross-chain transaction tracking system based on a large language model, comprising:

[0031] The acquisition module is used to acquire the raw data of multiple source chain transactions, decode the raw data of each source chain transaction, and generate the basic structure information of each source chain transaction.

[0032] The deposit transaction identification module is used to identify deposit transactions by using a large language model for deposit transaction identification to identify the basic structural information of each source chain transaction and output multiple source chain deposit transactions.

[0033] The determination module is used to determine a candidate set of target chain withdrawal transactions corresponding to each of the source chain deposit transactions based on the infrastructure information corresponding to each of the source chain deposit transactions.

[0034] The withdrawal transaction matching module is used to match withdrawal transactions based on the source chain deposit transactions and the target chain withdrawal transaction candidate set corresponding to each source chain deposit transaction using a large language model for withdrawal transaction matching, and outputs the target chain withdrawal transaction matching item.

[0035] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the cross-chain transaction tracking method based on the large language model as described above.

[0036] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the cross-chain transaction tracking method based on a large language model as described above.

[0037] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the cross-chain transaction tracking method based on the large language model as described above.

[0038] As can be seen from the above technical solutions, the present invention has the following advantages:

[0039] The above-mentioned technical solution of the present invention provides a cross-chain transaction tracking method based on a large language model. It acquires raw data from multiple source chain transactions, decodes the raw data of each source chain transaction to generate basic structural information of each source chain transaction, and uses a large language model for deposit transaction identification to identify deposit transactions based on the basic structural information of each source chain transaction, outputting multiple source chain deposit transactions. Based on the basic structural information corresponding to each source chain deposit transaction, it determines a candidate set of target chain withdrawal transactions corresponding to each source chain deposit transaction. Using a large language model for withdrawal transaction matching, it performs withdrawal transaction matching based on each source chain deposit transaction and the corresponding candidate set of target chain withdrawal transactions, outputting target chain withdrawal transaction matching items. Based on the above solution, the present invention leverages the powerful semantic understanding and generalization capabilities of the large language model, eliminating the need for preset fixed field matching rules. Even when faced with missing key fields or log structure changes due to protocol upgrades, it can still achieve effective matching by capturing the essential semantic relationships of cross-chain transactions, significantly improving the accuracy of withdrawal transaction matching. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating the steps of a cross-chain transaction tracking method based on a large language model, as provided in Embodiment 1 of the present invention;

[0042] Figure 2 This is an overall framework diagram of a cross-chain transaction tracking method based on a large language model provided in Embodiment 1 of the present invention;

[0043] Figure 3 This is a schematic diagram of the prompt word input content for a large language model for deposit transaction recognition provided in Embodiment 1 of the present invention;

[0044] Figure 4 This is a schematic diagram of the prompt word input content for a large language model used for matching withdrawal transactions, provided in Embodiment 1 of the present invention.

[0045] Figure 5 This is an overall framework diagram of cross-chain transaction pairing provided in Embodiment 1 of the present invention;

[0046] Figure 6 This is a structural block diagram of a cross-chain transaction tracking system based on a large language model, provided in Embodiment 2 of the present invention. Detailed Implementation

[0047] This invention provides a cross-chain transaction tracking method and system based on a large language model, which solves the technical problem of low accuracy in matching withdrawal transactions caused by existing cross-chain transaction tracking methods.

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that in the optional embodiments of the present invention, the object information and other related data involved require the permission or consent of the object when the embodiments of the present invention are applied to specific products or technologies, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. That is to say, if the embodiments of the present invention involve data related to the object, it needs to be obtained with the authorization and consent of the object, the authorization and consent of the relevant departments, and in compliance with the relevant laws, regulations, and standards of the country and region. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.

[0049] Terminology Explanation:

[0050] LLM (Large Language Model): A natural language processing model based on deep learning, with strong generalization, contextual reasoning and semantic understanding capabilities, which can be used for cross-domain semantic matching, classification and knowledge extraction tasks.

[0051] LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning technique that only updates the low-rank matrix parameters in the model, significantly reducing memory and computational requirements. It is suitable for rapid customized training of large language models with limited resources.

[0052] FP16 (16-bit floating-point): A half-precision floating-point format that consumes less computation and is often used in mixed-precision training and inference to improve speed and reduce memory consumption.

[0053] INT8 (8-bit integer): A quantization representation that further reduces storage space and accelerates model inference by compressing model parameters from floating-point numbers to 8-bit integers.

[0054] ABI (Application Binary Interface): A standard for describing functions and parameters of blockchain smart contracts, used for on-chain calls and event decoding, facilitating the parsing of transaction function signatures and log structures.

[0055] Cross-chain refers to the process of transferring assets, messages, or data between different blockchain networks. It typically involves "lock / deposit transactions" on the source chain and "release / withdrawal transactions" on the target chain, which are accomplished through cross-chain bridge protocols.

[0056] Cross-Chain Bridge: A relay system or set of contracts that connects different blockchains to enable cross-chain asset transfers, such as Celer cBridge, Multichain, and Polygon Bridge.

[0057] Multi-Chain: Refers to a network environment composed of multiple independent blockchain ecosystems (such as Ethereum, BSC, and Polygon). Data structures, contract interfaces, and transaction events vary between different chains.

[0058] A transaction is a state-changing operation initiated by a user or contract on the blockchain, which can include token transfers, contract calls, etc. In cross-chain scenarios, a complete cross-chain transaction typically consists of two parts: a source chain transaction and a target chain transaction.

[0059] Smart Contracts: Automated execution programs deployed on the blockchain, used to define and manage cross-chain bridge logic, event triggering, and fund locking rules.

[0060] Event Log: Event information recorded during the execution of a smart contract, including event name, parameters, and block timestamp, used to track transaction behavior and analyze cross-chain activities.

[0061] Semantic Matching: By understanding function names, parameter names, and contextual relationships through the model, it determines whether different fields or events represent the same meaning (such as "recipient" and "receiver") from a semantic level.

[0062] Transaction pairing: In cross-chain tracking, semantic comparison between source chain transactions and target chain candidate transactions determines whether they constitute the same cross-chain behavior.

[0063] Please see Figure 1 , Figure 1This is a flowchart illustrating the steps of a cross-chain transaction tracking method based on a large language model, as provided in Embodiment 1 of the present invention.

[0064] This invention provides a cross-chain transaction tracking method based on a large language model, comprising:

[0065] Step 101: Obtain the raw data of multiple source chain transactions, decode the raw data of each source chain transaction, and generate the basic structure information of each source chain transaction.

[0066] Raw data of source chain transactions: refers to the unprocessed transaction records captured from the source chain blockchain network. The core includes transaction input and event log, which is the original data source for extracting basic transaction structure information.

[0067] It should be noted that the raw data of multiple source chain transactions is obtained by using the Blockchainspider tool to connect to the source chain's RPC (Remote Procedure Call) interface for capture. This data includes transaction input (function call binary data) and event log (contract execution event log). The raw data of each source chain transaction is then decoded. The decoding process relies on the ABI (Application Binary Interface) of the smart contract related to the cross-chain bridge. The ABI is used to parse the function call logic of the transaction input and the event record content of the event log, extracting the function name, parameter fields (such as token address, amount, receiver, dstChainId, etc.), contract events and event parameters, transaction hash, block timestamp, and other core content of each source chain transaction to generate the basic structural information of each source chain transaction.

[0068] Furthermore, such as Figure 2 As shown, in the overall framework of the cross-chain transaction tracking method based on a large language model proposed in this invention, to overcome the problems of undisclosed bridge information, heterogeneous bridging structures, and difficulty in generalizing rules, this invention proposes an on-chain transaction modeling process based on semantic structure extraction, and introduces a large language model (LLM) as a cross-chain behavior discriminator with strong generalization ability. Based on this, a two-stage large language model inference process is designed and implemented: the first stage is used to identify whether a transaction belongs to cross-chain deposit behavior, and the second stage is used to match the target chain withdrawal transaction with the corresponding transaction in the candidate set. This invention aims to combine structured on-chain data with the semantic understanding capability of a language model, improving the intelligence level of cross-chain transaction identification and matching while maintaining data interpretability and source verifiability. Compared with methods based on fixed rules, this method has stronger flexibility and semantic fault tolerance, and is suitable for general transaction tracking tasks in multi-bridging protocol and multi-chain scenarios.

[0069] Step 102: Use a large language model for deposit transaction identification to identify deposit transactions based on the basic structural information of each source chain transaction, and output multiple source chain deposit transactions.

[0070] The large language model for deposit transaction identification refers to a dedicated model based on the basic large language model, which is efficiently fine-tuned by LoRA parameters. It is specifically designed to determine whether a source chain transaction is a cross-chain deposit and has the ability to identify the core semantic features of cross-chain deposit behavior.

[0071] It should be noted that a large language model for deposit transaction identification is used to identify deposit transactions based on the basic structural information of transactions from various source chains. This large language model for deposit transaction identification is based on the Qwen2.5-1.5B Large Language Model (LLM), and uses LoRA (Low-Rank Adaptation) parameter fine-tuning technology. Combined with the LLaMA-Factory training framework and labeled training dataset (positive examples are explicit cross-chain deposit transactions, and negative examples are ordinary transfers, non-cross-chain contract calls, etc.) to complete targeted training. Before identification, the basic structural information of each source chain transaction needs to be standardized, and then converted into English prompt words according to a predefined fixed English prompt word template. Subsequently, the English prompt words corresponding to each source chain transaction are input into the model. The model outputs the "whether it is a cross-chain deposit" judgment result (Yes / No) for each source chain transaction. Finally, based on the judgment result, transactions judged as Yes are selected, and multiple source chain deposit transactions are output.

[0072] Specifically, step 102 may include the following sub-steps:

[0073] S21. Convert the basic structure information of each source chain transaction into English prompts according to the predefined fixed prompt template;

[0074] S22. Input the English prompts corresponding to each source chain transaction into the large language model used for deposit transaction recognition, and output the judgment results corresponding to each source chain transaction.

[0075] S23. Based on the judgment results corresponding to each source chain transaction, perform deposit transaction screening on each source chain transaction to determine multiple source chain deposit transactions.

[0076] The predefined fixed prompt word template refers to a unified English expression template designed specifically for source chain deposit transaction identification tasks. It consists of three parts: "task instruction (clearly determine whether the transaction is a cross-chain deposit) + structured field labels (annotate the semantics of the core parameters of the transaction) + output requirements (limited to outputting only Yes or No)". It can transform the basic structural information of source chain transactions into an input format that can be understood by the large language model used for deposit transaction identification.

[0077] The judgment result refers to the output of the large language model used for deposit transaction identification on the source chain transaction. It only includes two categories: Yes (meaning the transaction is a cross-chain deposit transaction) and No (meaning the transaction is not a cross-chain deposit transaction), and is the direct basis for screening source chain deposit transactions.

[0078] Source chain deposit transactions refer to source chain transactions that are judged as Yes by the large language model used for deposit transaction identification. They are the starting point of cross-chain transactions and provide benchmark data for matching subsequent target chain withdrawal transactions.

[0079] It should be noted that this step aims to determine whether an on-chain transaction is the initiation of a cross-chain action, i.e., a deposit transaction initiated by a user to the target chain via a bridging protocol. Since different bridging protocols differ significantly in contract call structures, function names, and event designs, traditional rule-based methods are difficult to adapt uniformly. Therefore, this invention employs a large language model to model this task as a semantic binary classification problem. Prompt words guide the model to understand the transaction behavior, thereby determining whether it constitutes a cross-chain deposit.

[0080] Specifically, the basic structural information of each source chain transaction is converted into English prompts according to predefined fixed prompt word templates, which serve as input to a large language model for deposit transaction recognition. The input content of this large language model is as follows: Figure 3 As shown.

[0081] Furthermore, based on this prompt, the model outputs a judgment result (Yes or No), thereby completing the identification of the deposit transaction. This invention uses the Qwen2.5-1.5B model as the base model and uses the LoRA (Low-Rank Adaptation) strategy for efficient parameter fine-tuning. The training data consists of a large number of real on-chain transactions, where positive examples are explicitly marked cross-chain deposit transactions, and negative examples are ordinary transfers, contract calls, or non-cross-chain transactions. All training data are written using English prompts.

[0082] The model training employs the LLaMA-Factory framework, built upon a standard JSONL corpus. It incorporates different field combinations, expressions, and few-shot guidance strategies, designed using prompt word engineering, to enhance the model's understanding of cross-chain structures. Experiments show that the finely tuned model maintains stable recognition accuracy even when facing complex transaction structures such as function mutations, event field changes, and missing logs, significantly outperforming the untuned model or rule-based methods.

[0083] The transactions identified in this stage will be input as candidate deposit transactions into the next stage for matching determination on the target chain.

[0084] Specifically, the basic structural information of each source chain transaction is converted into English prompts according to a predefined fixed prompt template. This predefined fixed prompt template adopts a unified format of "task instruction + structured field label + output requirements," clearly annotating the semantic anchors of core fields such as token address, amount, and dstChainId. Simultaneously, the template content adapts to the semantic understanding logic of the Large Language Model (LLM). The English prompts corresponding to each source chain transaction are then input into the Large Language Model used for deposit transaction recognition. This model is based on the Qwen2.5-1.5B Large Language Model and utilizes LoRA (Low-Rank)... Adaptation (low-rank adaptive fine-tuning) is a high-efficiency parameter fine-tuning technique used to complete targeted training. Its training data includes clearly labeled positive examples of cross-chain deposits and negative examples such as ordinary transfers and non-cross-chain contract calls. After receiving the input, the model outputs the judgment result of "whether it is a cross-chain deposit" for each source chain transaction. The judgment result only includes two categories: Yes and No. Based on the judgment result of each source chain transaction, the deposit transaction is screened for each source chain transaction. Source chain transactions with a judgment result of Yes are selected, and multiple source chain deposit transactions are identified. These types of transactions will serve as the core benchmark objects for subsequent target chain withdrawal transaction pairing.

[0085] Step 103: Based on the basic structure information corresponding to each source chain deposit transaction, determine the target chain withdrawal transaction candidate set corresponding to each source chain deposit transaction.

[0086] It should be noted that, based on the basic structural information corresponding to each source chain deposit transaction (the source chain deposit transaction is identified as a deposit transaction by the large model for recognizing deposit transactions, and its basic structural information comes from the input data after decoding the source chain transaction and the event log obtained from the decoding), the target chain address information is first extracted from it. Then, the Blockchainspider tool is used to crawl the target chain address for transactions. At the same time, core screening dimensions such as dstChainId (target chain ID), transaction time, token address, and cross-chain bridge contract address are further extracted. The target chain network corresponding to dstChainId is first identified, and then the ABI (Application Blockchain Interface) of the target chain obtained through transaction crawling is retrieved. The target chain transaction infrastructure information is decoded using a binary interface (BIN). Then, precise filtering is performed on the target chain transaction infrastructure information according to preset filtering rules. These rules include: the transaction time being within a ±30-minute time window of the source chain deposit transaction time; the token address being consistent with or related to the token address of the source chain deposit transaction through on-chain mapping; and the transaction triggering function or event being a withdrawal type (such as Withdraw, anySwapIn, mint, etc.) and involving a corresponding cross-chain bridge contract. Through this multi-dimensional targeted filtering, a candidate set of target chain withdrawal transactions corresponding to each source chain deposit transaction is determined. This candidate set provides targeted and effective candidate objects for subsequent precise matching of cross-chain transactions.

[0087] Specifically, step 103 may include the following sub-steps:

[0088] S31. Based on the basic structure information corresponding to each source chain deposit transaction, obtain the target chain address information corresponding to each source chain deposit transaction;

[0089] S32. For each source chain deposit transaction, crawl the withdrawal transaction of the target chain address associated with the target chain address information, and generate a candidate set of target chain withdrawal transactions for each source chain deposit transaction.

[0090] Target chain address information refers to address-related data extracted from the basic structural information of source chain deposit transactions that points to the target chain. The core of this data includes the target chain receiving address and the target chain identifier, which is the core basis for determining the scope of target chain crawling.

[0091] Cash withdrawal transaction crawling refers to the process of using the Blockchainspider tool to capture on-chain transaction data related to cash withdrawal behavior within a specific time window for a target chain address associated with the target chain address information, ensuring the targeting and effectiveness of the crawled data.

[0092] The target chain withdrawal transaction candidate set refers to the transaction set formed after crawling transaction data from the target chain address, decoding it using ABI, and filtering it based on conditions. All transactions in the set have the potential to be paired with the corresponding source chain deposit transactions, providing targeted candidate objects for subsequent accurate matching.

[0093] It should be noted that, based on the basic structural information corresponding to each source chain deposit transaction, key content such as the receiver (receiving address) and the target chain identifier associated with dstChainId are extracted from the decoded inputdata (function call data) and eventlog (contract execution event log) included in this basic structural information, and integrated to obtain the target chain address information corresponding to each source chain deposit transaction; then, the Blockchainspider tool is used to crawl withdrawal transactions of the target chain addresses associated with the target chain address information corresponding to each source chain deposit transaction. The crawling scope is limited to the transaction data of the target chain address within a time window of ±30 minutes of the source chain deposit transaction time. The crawled content covers transaction input and eventlog, and then the ABI (Application Binary) of the cross-chain bridge smart contract is used. The interface (using binary interfaces) decodes the crawled transaction data, extracts core information such as function name, token address, amount, and srcChainId for each transaction, and further filters out transactions that trigger withdrawal functions or events (such as Withdraw and anySwapIn) and whose token addresses match the source chain deposit transactions, generating a candidate set of target chain withdrawal transactions corresponding to each source chain deposit transaction.

[0094] Step 104: Using a large language model for withdrawal transaction matching, perform withdrawal transaction matching based on the candidate set of withdrawal transactions on the target chain corresponding to each source chain deposit transaction and output the target chain withdrawal transaction matching item.

[0095] The target chain withdrawal transaction candidate set includes multiple target chain withdrawal transactions.

[0096] It should be noted that after identifying a cross-chain deposit transaction, the second stage aims to find the corresponding withdrawal transaction within the candidate transaction set of the target chain. Since there are no hash references between chains, withdrawal transactions cannot be directly uniquely identified by fields. Therefore, this invention designs the pairing task as a method where "multiple candidate transactions are combined with source chain transactions to form binary inputs," and the pairing relationship is determined through multiple rounds of reasoning.

[0097] Specifically, step 104 may include the following sub-steps:

[0098] S41. Decode the original data of multiple target chain withdrawal transactions corresponding to each source chain deposit transaction, and output the basic structure information of multiple target chain withdrawal transactions corresponding to each source chain deposit transaction.

[0099] S42. Based on the basic structural information of deposit transactions on each source chain and the basic structural information of withdrawal transactions on multiple target chains corresponding to deposit transactions on each source chain, determine multiple transaction splicing texts.

[0100] Further, step S42 may include the following sub-steps:

[0101] S421. Select a single candidate source chain deposit transaction from each source chain deposit transaction;

[0102] S422. Based on the basic structure information of a single candidate source chain deposit transaction, select multiple target chain candidate withdrawal transactions from multiple target chain withdrawal transactions corresponding to the single candidate source chain deposit transaction.

[0103] S423. According to the predefined English prompt word template, the basic structure information of a single candidate cross-chain deposit transaction is concatenated and combined with the basic structure information of each corresponding target chain candidate withdrawal transaction to generate multiple transaction concatenation texts.

[0104] A single candidate source chain deposit transaction refers to a single transaction selected from multiple source chain deposit transactions. It serves as the benchmark for cross-chain transaction pairing, and its basic structural information provides the core basis for the screening and pairing of candidate withdrawal transactions on the target chain.

[0105] Target chain candidate withdrawal transactions refer to transactions with cross-chain matching potential selected from withdrawal transactions of the corresponding target chain based on key screening dimensions of deposit transactions of a single candidate source chain. These transactions must meet multiple matching conditions such as time, token, and chain ID.

[0106] The predefined English prompt template refers to a unified English expression template designed specifically for cross-chain transaction pairing. It includes pairing task instructions, transaction core field labels, and output requirements, which can transform the basic structural information of two transactions into a comparison input format that the pairing model can understand.

[0107] Transaction splicing text refers to the text formed by splicing together the basic structural information of a single candidate source chain deposit transaction and a single target chain candidate withdrawal transaction according to a predefined template. It is the direct data source for inputting the large language model used for withdrawal transaction pairing.

[0108] It should be noted that a single candidate source chain deposit transaction is selected from each source chain deposit transaction, and this transaction serves as the benchmark for cross-chain pairing. Based on the basic structural information of this single candidate source chain deposit transaction, key screening dimensions such as transaction time, token address, dstChainId, receiver, and cross-chain bridge contract address are extracted. Among the multiple target chain withdrawal transactions corresponding to this single candidate source chain deposit transaction, those with transaction times within a 30-minute window of the benchmark transaction time, triggering withdrawal functions / events such as Withdraw or anySwapIn, and token addresses matching the benchmark transaction are selected. Transactions that are consistent or can be associated through on-chain mapping and whose srcChainId matches the source chain identifier are used as candidate withdrawal transactions for multiple target chains. Following a predefined English prompt template, the basic structural information of a single candidate source chain deposit transaction is concatenated with the basic structural information of each corresponding candidate withdrawal transaction on the target chain. This template uses "determine if they are the same cross-chain pairing transaction" as its task instruction, and clearly labels the core fields of the two transactions, such as amount, receiver, and chain ID, to strengthen semantic association. Finally, multiple structurally unified transaction concatenation texts are generated to prepare for subsequent input into the pairing model. The prompt input content for the large language model used for withdrawal transaction pairing is as follows: Figure 4 As shown.

[0109] S43. Using a large language model for withdrawal transaction matching, output the target chain withdrawal transaction matching item based on the concatenated text of multiple transactions.

[0110] Furthermore, step S43 may include the following sub-steps:

[0111] S431. Input each transaction concatenation text into the large language model used for withdrawal transaction matching, and output the matching score corresponding to each transaction concatenation text.

[0112] S432. Among the matching scores corresponding to each transaction concatenation text, select the target chain candidate withdrawal transaction corresponding to the highest matching score as the target chain withdrawal transaction matching item corresponding to a single candidate cross-chain deposit transaction.

[0113] It should be noted that, as Figure 5 As shown, each set of transaction concatenation text consists of a source chain deposit transaction (query) and a target chain candidate transaction (target). The structured fields, after expansion, are input into the model as part of the prompt words. For example:

[0114] The model classifies each pair of samples and outputs a "match" answer. Simultaneously, this invention records the probability value of the "Yes" token generated by the model as a pairing score, sorts all candidates, and selects the transaction with the highest score as the final match (target chain withdrawal transaction match).

[0115] Specifically, the concatenated text of each transaction is input into a large language model used for withdrawal transaction pairing. This model is based on the Qwen2.5-1.5B Large Language Model (LLM) and uses LoRA (Low-Rank Adaptation) parameter fine-tuning technology for targeted training. Its training data includes positive examples of real pairings of "source chain deposit - target chain withdrawal" and negative examples of non-pairings within the same time window. A pairwise evaluation mechanism is used to avoid insufficient GPU memory. After receiving the input, the model outputs the pairing score corresponding to each concatenated text of the transaction. This score is represented by the probability value of the token generated by the model to make a "Yes" judgment. The higher the value, the higher the degree of matching between the two transactions as the same cross-chain behavior. Among the pairing scores corresponding to each concatenated text of the transaction, it is necessary to first confirm whether the largest pairing score reaches a preset threshold (such as 80%) to ensure matching accuracy. Then, the target chain candidate withdrawal transaction corresponding to the largest pairing score that meets the threshold condition is selected as the target chain withdrawal transaction matching item corresponding to a single candidate cross-chain deposit transaction, thus completing the deposit and withdrawal pairing of a single cross-chain transaction.

[0116] It's worth noting that this step also uses English prompts for fine-tuning training. During data construction, positive samples are real paired trading pairs, while negative samples are erroneous candidates within the same time window. To improve the model's understanding, explicit labels are used for fields in the prompts to avoid interference from ambiguous information. For example, "receiver" is used instead of "address," and "dstChainId" and "token address" are clearly marked, ensuring that the model establishes correct field mapping logic in semantic judgment.

[0117] When the candidate set is large, it is impossible to concatenate all candidates at once, which can cause insufficient video memory. To address this, this invention employs a pairwise evaluation mechanism, inputting only one query and one target at a time, thereby achieving batch judgment under memory-constrained conditions.

[0118] For comparison of technical effectiveness, existing technologies can be used as a reference. Mainstream cross-chain transaction tracking methods can be roughly divided into two categories:

[0119] 1) Rule / Field Matching Methods: These methods perform rigorous comparisons and time window filtering using fields such as event name, function signature, amount, receiver, and dstChainId. The advantages are interpretability and high accuracy on structurally sound samples; the disadvantages are extreme invulnerability to missing fields, log distortions, protocol upgrades, and function renaming, and high maintenance costs. These methods are often used in single-chain or centralized cross-chain scenarios. A typical example is the CONNECTOR tool proposed by Lin et al., which achieves cross-chain pairing through static rule mapping, but lacks automated learning and adaptive capabilities. Furthermore, the core idea of ​​the existing CONNECTOR framework is to rely solely on provable on-chain data (contract calls + event logs), without relying on bridge-side private APIs or intermediary services. However, this approach is extremely invulnerable to missing fields / log distortions (rule-based methods can experience complete mismatches and accuracy drops to 0% under abnormal samples); rules need to be manually maintained for each bridge protocol / contract version, resulting in high maintenance costs when facing protocol upgrades or new bridges; and "zero hits" can lead to no matching output, affecting audit coverage and the availability of statistical indicators.

[0120] 2) Semantic Structure / Graph Modeling Improvements: These methods enhance the ability to identify heterogeneous structures by parsing the contract ABI, constructing contract call graphs or fund flow graphs, extracting semantic fields, and performing structured encoding (e.g., converting function signatures and key parameters into semantic vectors, then combining them with graph structure features). While these methods can alleviate problems caused by naming variations and parameter position variations to some extent, they still face the vulnerability of "missing fields / zero hits" if strict field matching is still relied upon for final pairing. In recent years, researchers have further introduced named entity recognition and information retrieval technologies, such as the ABCTRACER framework proposed by Lin et al., which combines event log semantic mining and implicit clue encoding to achieve bidirectional cross-chain tracking, significantly improving automation and generalization capabilities under multiple bridging protocols.

[0121] Based on the above, the shortcomings of existing technologies are roughly as follows:

[0122] (1) Strong heterogeneity of transactions: Cross-chain transactions involve two transactions on the source chain and the target chain. Different bridge protocols have huge differences in contract structure, function calls and event fields, and lack a unified identifier, making it difficult to pair them directly.

[0123] (2) Field missing and log non-standardization: Real on-chain transactions often have field missing, log distortion, time delay, etc. Rule-based methods become completely ineffective once the fields are incomplete, and the accuracy drops sharply.

[0124] (3) Poor adaptability of rule methods: Traditional rules rely on fixed field matching, which is insufficient for bridge protocol upgrades, version changes or new protocols, resulting in high maintenance costs and poor generalization.

[0125] (4) Increased security and auditing requirements: Cross-chain bridges have risks such as contract vulnerabilities and verification failures. Hacking, fraud and money laundering incidents occur frequently, and there are higher requirements for the transparency and traceability of cross-chain transactions. However, existing methods cannot provide robust tracking capabilities.

[0126] (5) Insufficient real-time performance and robustness: When faced with multi-bridge protocols and complex on-chain structures, traditional methods are difficult to maintain accurate and stable pairing capabilities in near real-time scenarios.

[0127] Therefore, this invention provides a cross-chain transaction tracking method based on a large language model, the core problem of which is:

[0128] (1) Overcome the heterogeneity of different bridge protocol contract structures and log events, and realize the universal identification and unified matching of cross-chain transactions.

[0129] (2) To solve the problem that existing rules and methods are completely ineffective in complex scenarios such as missing fields, non-standard logs, and cross-chain transaction time delays, and to improve the fault tolerance and robustness of the system.

[0130] (3) Provide generalization capabilities that can adapt to multi-bridge protocols, multi-chain structures and future new protocols, avoiding the high maintenance costs caused by manually writing rules.

[0131] (4) Meet the needs of cross-chain transactions in security audit and compliance tracking, and improve the transparency and traceability of cross-chain behavior.

[0132] (5) Enables rapid adaptation and localized deployment of large models with low resource consumption, while maintaining high accuracy and having real-time analysis capabilities and engineering feasibility.

[0133] Specifically, this invention replaces traditional rule engines for cross-chain transaction tracking through a two-stage semantic reasoning process: "deposit transaction identification → withdrawal transaction matching." For heterogeneous fields in different bridge protocols (such as Celer's dstChainId and Multichain's chainId), the first stage models transaction classification as a semantic binary classification task. The second stage determines the pairing relationship through comparative reasoning of "source chain transactions - target chain candidate transactions," overcoming the limitations of hard-coded rule adaptability. Secondly, for the Qwen2.5 model, LoRA low-rank adaptive fine-tuning is used, training only the low-rank matrix parameters of the attention layer. With LoRA-rank=8 and a learning rate of 1e-4, only 0.1% of the parameters of the original model are updated. Through FP16 mixed-precision training and gradient accumulation optimization, the memory requirements for fine-tuning the 1.5B model are controlled within 6GB, and for the 3B model within 8GB, reducing resource requirements by more than 60% compared to full fine-tuning (requiring 16GB+ memory). Simultaneously, standardized prompt word templates with field labels are designed for cross-chain transaction scenarios. The seven template types, structured as "task instruction + transaction details + output requirements," clearly define the semantic anchors of core fields such as "token," "amount," and "dstChainId," resolving ambiguity in field naming (e.g., standardizing the use of "receiver / to / recipient"). Experiments have verified that this design improves the accuracy of deposit transaction recognition to 100%, providing a stable semantic input paradigm for the model.

[0134] Furthermore, this invention employs a "single query - single candidate" pairwise input mode, combined with batch inference to optimize memory usage. By controlling the KVCache size for a single inference (2 × batch size × sequence length × number of layers × hidden dimension), the memory usage of a single GPU is stabilized below 6GB during 32 transactions / batch processing, ensuring continuous operation in high-frequency trading scenarios. In addition, this invention achieves low memory usage deployment based on FP16 precision optimization and quantization technology. The memory requirement for 1.5B model inference is controlled at 4-5GB, and for 3B model at 7-8GB (including 20-30% additional overhead). Localized deployment avoids the risk of data leakage, reducing response latency to the second level. Simultaneously, support for INT8 quantization further compresses memory requirements to 25% of the theoretical value, adapting to diverse scenarios in consumer and industrial computing devices.

[0135] Compared to existing technologies, the core logic of the mainstream improvement scheme of the existing Connector is as follows: by parsing the cross-chain bridge contract ABI file, constructing the contract call graph (such as the call path of user-bridge contract-token contract), extracting semantic features such as function names / key parameters and converting them into semantic vectors, and combining the graph structure adjacency relationship to complete the identification of cross-chain deposit transactions; in the withdrawal transaction matching stage, it still relies on the "strict matching of cross-chain metadata" strategy - that is, comparing fields such as amount (allowing ±3% error), receiver, token address, dstChainId, etc. one by one, and filtering candidate transactions through a time window (such as 30 minutes) to finally determine the matching relationship. While this existing technology alleviates the sensitivity issues of traditional rule-based methods to function naming variations and parameter position adjustments, it still has two major limitations: First, the final pairing depends on the completeness of key fields. If there are missing fields in the on-chain logs (such as the target chain transaction not recording srcChainId) or log distortions (such as abnormal event parameter encoding), the pairing will fail due to a "zero hit". Second, it has poor adaptability to cross-chain bridge protocol upgrades. If the new bridge protocol adds function types or changes event fields (such as changing "receiver" to "recipient"), the semantic feature extraction logic and field mapping rules need to be redesigned, resulting in high maintenance costs and difficulty in quickly adapting to the dynamic ecosystem of multiple bridges and multiple chains.

[0136] Compared to existing technologies, this invention achieves significant breakthroughs in robustness, generalization, and engineering feasibility through its core solution of "two-stage inference based on Large Language Model (LLM) + efficient fine-tuning of LoRA parameters." Its specific advantages are reflected in the following two aspects:

[0137] First, this invention completely eliminates the reliance on "strict field matching" in existing technologies, solving the matching failure problem in scenarios of "missing fields / deformed logs" through the semantic understanding capabilities of LLM. The withdrawal matching process in existing technologies is essentially still a mechanical logic of "feature fields → rule comparison". Once a key field is missing (such as the source chain transaction lacking dstChainId, the target chain transaction lacking amount), or the log structure changes due to protocol upgrades (such as Multichain changing the "anySwapOut" event to "anySwapOutUnderlying"), its preset field matching rules become invalid. In experiments, the accuracy of existing technologies in this scenario plummeted to 0%. This invention models deposit identification as a "semantic binary classification task" and withdrawal matching as a "cross-transaction semantic comparison task." Through structured prompt word engineering, it guides the Qwen2.5 model to understand the essential logic of cross-chain transactions. For example, the model can automatically establish mapping relationships between different fields through semantic associations such as "_bridge address identifies the cross-chain protocol" and "_dstAddress corresponds to the target link recipient" (e.g., identifying "lockedAmount" of PolygonBridge and "_amount" of Celer as the same semantic amount field). Even if 1-2 key fields are missing (e.g., covering amount or removing dstChainId), it can still complete the judgment based on contextual reasoning. In experiments, the accuracy of this invention in such scenarios remains above 70%, completely solving the core pain point of "fragile field dependency" in existing technologies.

[0138] Secondly, this invention significantly improves the generalization ability for multi-bridge protocols and protocol upgrade scenarios through a highly efficient LoRA parameter fine-tuning strategy, while reducing engineering maintenance costs and resource consumption. Existing technologies require separate design of semantic feature extraction rules for different cross-chain bridge protocols (e.g., Celer cBridge needs to parse "Send" and "Withdraw" events, while Multichain needs to parse "anySwapOut" and "anySwapIn" events). If a new bridge protocol (such as the newly added LayerZero compatible bridge) adopts a completely new contract call structure, the contract call graph needs to be reconstructed and the semantic vector encoding logic adjusted. The entire adaptation process requires several weeks of feature engineering iterations. This invention, based on the Qwen2.5 model, uses LoRA technology for targeted fine-tuning. The training data only needs to include a few hundred real transaction samples from the New Bridge protocol (the training dataset "pairwise_train" in the document contains standardized cross-chain transaction pairs), without requiring model architecture reconstruction or retraining of the backbone network. Furthermore, the memory consumption during fine-tuning is controllable; the memory required for fine-tuning based on the Qwen2.5-3B model is kept below 8GB, far lower than the resource requirements for full model training, shortening the adaptation cycle to new protocols to 1-2 days. In addition, existing technologies require separate field verification logic for each chain to support multi-chain (such as ETH, BSC, Polygon) tracking, while the LLM of this invention can automatically adapt to the differences in transaction structures of different chains through the semantic identifier "chainId" in the prompt words. In experiments, it maintained a pairing accuracy of over 86.5% in cross-chain tracking tasks for Celer, Multichain, and Polygon Bridge protocols, demonstrating significantly better generalization ability than existing technologies.

[0139] In this embodiment of the invention, a cross-chain transaction tracking method based on a large language model is provided. The method acquires raw data from multiple source chain transactions and decodes the raw data to generate basic structural information for each source chain transaction. A large language model for deposit transaction identification is used to identify deposit transactions based on the basic structural information of each source chain transaction, outputting multiple source chain deposit transactions. Based on the basic structural information corresponding to each source chain deposit transaction, a candidate set of target chain withdrawal transactions corresponding to each source chain deposit transaction is determined. A large language model for withdrawal transaction matching is used to match withdrawal transactions based on each source chain deposit transaction and the corresponding candidate set of target chain withdrawal transactions, outputting target chain withdrawal transaction matching items. Based on the above scheme, this invention leverages the powerful semantic understanding and generalization capabilities of the large language model, eliminating the need for pre-setting fixed field matching rules. Even when faced with missing key fields or log structure changes due to protocol upgrades, it can still achieve effective matching by capturing the essential semantic relationships of cross-chain transactions, significantly improving the accuracy of withdrawal transaction matching.

[0140] Please see Figure 6 , Figure 6 This is a structural block diagram of a cross-chain transaction tracking system based on a large language model, provided in Embodiment 2 of the present invention.

[0141] This invention provides a cross-chain transaction tracking system based on a large language model, comprising:

[0142] The acquisition module 601 is used to acquire the raw data of multiple source chain transactions, decode the raw data of each source chain transaction, and generate the basic structure information of each source chain transaction.

[0143] The deposit transaction identification module 602 is used to identify deposit transactions by using a large language model for deposit transaction identification to identify the basic structural information of each source chain transaction and output multiple source chain deposit transactions.

[0144] Module 603 is used to determine the candidate set of target chain withdrawal transactions corresponding to each source chain deposit transaction based on the infrastructure information corresponding to each source chain deposit transaction.

[0145] The withdrawal transaction matching module 604 is used to match withdrawal transactions based on the candidate set of withdrawal transactions on the target chain corresponding to each source chain deposit transaction and the large language model used for withdrawal transaction matching, and output the target chain withdrawal transaction matching item.

[0146] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0147] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the cross-chain transaction tracking method based on a large language model as described in the above embodiments.

[0148] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the cross-chain transaction tracking method based on a large language model as described in the above embodiments.

[0149] This invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the cross-chain transaction tracking method based on a large language model as described in the above embodiments.

[0150] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0151] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0152] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0153] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0154] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for cross-chain transaction tracking based on a large language model, characterized in that, The method comprises the following steps: obtaining raw data of a plurality of source chain transactions, and decoding the raw data of each of the source chain transactions to generate basic structure information of each of the source chain transactions; performing deposit transaction identification on the basic structure information of each of the source chain transactions by using a large language model for deposit transaction identification, and outputting a plurality of source chain deposit transactions; based on the basic structure information corresponding to each of the source chain deposit transactions, determining a candidate set of target chain withdrawal transactions corresponding to each of the source chain deposit transactions; performing withdrawal transaction pairing on each of the source chain deposit transactions and the candidate set of target chain withdrawal transactions corresponding to each of the source chain deposit transactions by using a large language model for withdrawal transaction pairing, and outputting a target chain withdrawal transaction matching item. 2.The method of claim 1, wherein, The method comprises the following steps: converting the basic structure information of each of the source chain transactions into English prompt words according to a predefined fixed prompt word template; inputting the English prompt words corresponding to each of the source chain transactions into the large language model for deposit transaction identification, and outputting a determination result corresponding to each of the source chain transactions; based on the determination result corresponding to each of the source chain transactions, performing deposit transaction screening on each of the source chain transactions to determine a plurality of source chain deposit transactions. 3.The method of claim 1, wherein, The method comprises the following steps: based on the basic structure information corresponding to each of the source chain deposit transactions, obtaining target chain address information corresponding to each of the source chain deposit transactions; performing withdrawal transaction crawling on the target chain addresses associated with the target chain address information corresponding to each of the source chain deposit transactions to generate a candidate set of target chain withdrawal transactions corresponding to each of the source chain deposit transactions. 4.The method of claim 1, wherein, The candidate set of target chain withdrawal transactions comprises a plurality of target chain withdrawal transactions. The method comprises the following steps: decoding the raw data of the plurality of target chain withdrawal transactions corresponding to each of the source chain deposit transactions to output basic structure information of the plurality of target chain withdrawal transactions corresponding to each of the source chain deposit transactions; based on the basic structure information of each of the source chain deposit transactions and the basic structure information of the plurality of target chain withdrawal transactions corresponding to each of the source chain deposit transactions, determining a plurality of transaction splicing texts; 5. The method of claim 4, wherein the method further comprises: outputting a target chain withdrawal transaction matching item by using the large language model for withdrawal transaction pairing according to the plurality of transaction splicing texts. The method comprises the following steps: selecting a single candidate source chain deposit transaction from each of the source chain deposit transactions; based on the basic structure information of the single candidate source chain deposit transaction, selecting a plurality of target chain candidate withdrawal transactions from the plurality of target chain withdrawal transactions corresponding to the single candidate source chain deposit transaction; The basic structure information of the single candidate cross-chain deposit transaction is respectively spliced and combined with the basic structure information of each corresponding target chain candidate withdrawal transaction according to a predefined English prompt word template, to generate a plurality of transaction splicing texts. 6.The method of claim 5, wherein, The large language model for withdrawal transaction pairing is used to output target chain withdrawal transaction matching items according to the plurality of transaction splicing texts, including: Each of the transaction splicing texts is input into the large language model for withdrawal transaction pairing, and a pairing score corresponding to each of the transaction splicing texts is output; In the pairing scores corresponding to each of the transaction splicing texts, the target chain candidate withdrawal transaction corresponding to the maximum pairing score is selected as the target chain withdrawal transaction matching item corresponding to the single candidate cross-chain deposit transaction. 7.A cross-chain transaction tracking system based on a large language model, characterized in that, Including: An acquisition module is configured to acquire original data of a plurality of source chain transactions, and decode the original data of each of the source chain transactions to generate basic structure information of each of the source chain transactions; A deposit transaction identification module is configured to use a large language model for deposit transaction identification to identify the basic structure information of each of the source chain transactions as deposit transactions, and output a plurality of source chain deposit transactions; A determination module is configured to determine a candidate set of target chain withdrawal transactions corresponding to each of the source chain deposit transactions based on the basic structure information corresponding to each of the source chain deposit transactions A withdrawal transaction pairing module is configured to use a large language model for withdrawal transaction pairing to pair withdrawal transactions according to each of the source chain deposit transactions and the candidate set of target chain withdrawal transactions corresponding to each of the source chain deposit transactions, and output target chain withdrawal transaction matching items.

8. An electronic device, comprising: A computer program is stored in a memory and executed by a processor, and when the computer program is executed, the processor executes the steps of the cross-chain transaction tracking method based on a large language model according to any one of claims 1-6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed to implement the cross-chain transaction tracking method based on a large language model according to any one of claims 1-6.

10. A computer program product, characterised in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer executes the cross-chain transaction tracking method based on a large language model according to any one of claims 1-6.

Images