Blockchain oracle off-chain data authenticity verification method and system based on parallel deepening reasoning

Abstract

The application proposes a blockchain oracle off-chain data authenticity verification method and system based on parallel deepening reasoning, and constructs a complete off-chain data verification mechanism. The scheme first uniformly encapsulates a structured evidence package containing multi-dimensional information such as the statement to be verified, signature evidence, cross-source observation, etc. Then, multiple parallel verification perspectives such as time, cross-source, and semantics are generated to construct an initial reasoning path with a traceable evidence pointer. The path is quantitatively evaluated by three-dimensional indicators and Top-K pruning is used to control the computational overhead. The reasoning is iteratively deepened until the depth or evidence sufficiency condition is met. The assumption conflict frequency of all iterative reasoning paths is summarized, and the data normal / abnormal attack label and confidence are output. The application supplements the semantic level verification based on digital signature, can identify semantic attacks such as time sequence manipulation, caliber fraud, and pseudo multi-source, and the reasoning is traceable and the computing power is controllable, which improves the off-chain data chaining credibility of the oracle.

Description

Technical Field

[0001] This invention belongs to the fields of blockchain security, oracle security, off-chain data authenticity verification, and artificial intelligence reasoning enhancement technology, specifically involving a method and system for verifying the authenticity of off-chain data in a blockchain oracle based on parallel deep reasoning. Background Technology

[0002] Blockchain systems provide a strong and reliable foundation for on-chain state transfers through distributed consensus and immutable ledger mechanisms. Smart contracts can automatically execute business logic such as finance, insurance, supply chain, and data transactions on-chain according to preset rules. However, since smart contracts themselves typically cannot directly access information from the off-chain world, such as market prices, logistics status, weather data, real-world events, identity authentication results, or the status of external business systems, practical applications require the reliance on blockchain oracles to bring off-chain data into the on-chain environment.

[0003] Oracles act as a bridge between on-chain smart contracts and off-chain data sources, and their security directly impacts the correctness of smart contract execution results. Generally, an oracle system involves off-chain data generation, data collection, data processing, authentication and encapsulation, transaction submission, and on-chain contract consumption. During this process, if off-chain data is manipulated, delayed, replaced, or misinterpreted before entering the blockchain, even if the on-chain smart contract executes correctly according to the protocol, economic losses or incorrect business outcomes may occur based on erroneous or misleading data.

[0004] Existing oracle security mechanisms typically emphasize data source authentication, transmission integrity, and signature verification. For example, digital signatures, hash verification, trusted execution environment proofs, metadata authentication, or multi-node aggregation mechanisms can verify whether data originates from the expected entity or whether the data has been tampered with at the bit level during transmission. While these mechanisms play a crucial role in confirming data source and transmission integrity, they primarily focus on "whether the data was submitted by a specific entity" and "whether the data remains consistent during transmission," and cannot fully guarantee "whether the data semantically reflects the off-chain facts."

[0005] Specifically, in many real-world attack scenarios, attackers do not need to compromise the signature mechanism or directly tamper with on-chain transactions. Instead, they can create misleading data through semantic manipulation. For example, attackers can exploit short-term price fluctuations to pollute the oracle's sampling window, making the reported data valid in terms of signature but misleading in terms of timing; they can change units, statistical definitions, or field meanings to make the data format appear legitimate but the actual semantics deviate; they can create multiple seemingly independent data sources that actually originate from the same upstream dependency, thus creating false cross-source consistency; and they can construct situations where off-chain narratives are inconsistent with the on-chain transaction context, causing smart contracts to execute based on data that is formally valid but semantically inaccurate.

[0006] Therefore, verifying the authenticity of off-chain data is not a task based on a single signal, but requires comprehensive analysis of various heterogeneous pieces of evidence. The verification system needs to examine not only signatures and metadata, but also raw observational data, cross-source reference data, historical trajectories, on-chain transaction records, smart contract states, and off-chain contextual information to determine whether these pieces of evidence support each other and whether there are anomalies such as temporal inconsistencies, semantic discrepancies, cross-source evidence conflicts, or source dependency contamination.

[0007] In recent years, large language models have demonstrated strong capabilities in natural language understanding, complex reasoning, code generation, and multi-step decision-making tasks. Chain-based reasoning methods enable models to generate intermediate reasoning steps, thereby improving the interpretability of complex problem handling. Tree-based, graph-based, and multi-agent discussion methods further attempt to enhance the model's decision-making ability through branch search or multi-path reasoning. However, off-chain data authenticity verification tasks have strong structured constraints, requiring models to perform traceable, constrained, and verifiable reasoning across multiple evidence channels. If only ordinary chain-based reasoning is used, the model may miss key evidence or introduce single-path bias; if unconstrained tree-based or graph-based extended reasoning is used, it may generate too many redundant branches, increase computational overhead, and introduce reasoning content unrelated to the current evidence, thus reducing decision stability.

[0008] Therefore, the existing technology has at least the following problems:

[0009] First, existing oracle verification methods based on signature, hash, or authentication mechanisms mainly ensure the integrity of the source and transmission, but have difficulty identifying off-chain data with valid signatures but distorted semantics.

[0010] Second, existing rule-based anomaly detection methods rely on manually set rule templates, which are difficult to cover different protocols, different data types, and unknown semantic attack variants, and have weak generalization and transferability.

[0011] Third, if existing large language model reasoning methods lack constraints oriented towards evidence structure, they are prone to problems such as single reasoning path, insufficient evidence coverage, branch expansion, redundant reasoning, or unstable conclusions.

[0012] Fourth, the existing oracle verification process lacks a reasoning framework that can comprehensively analyze temporal consistency, semantic consistency, cross-source consistency, source independence, and on-chain context consistency while controlling computational costs. Summary of the Invention

[0013] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for verifying the authenticity of off-chain data using a blockchain oracle based on parallel in-depth reasoning, thereby overcoming the aforementioned deficiencies in existing technologies.

[0014] To achieve the above objectives, the present invention provides the following technical solution:

[0015] A blockchain oracle-based method for verifying the authenticity of off-chain data using parallel in-depth reasoning includes the following steps:

[0016] The evidence acquisition step involves acquiring a structured evidence package corresponding to the oracle off-chain data authenticity verification task to be verified. The structured evidence package includes at least the statement to be verified and multi-dimensional related evidence associated with the statement to be verified.

[0017] The parallel perspective generation step involves determining the number of initial verification perspectives based on the structured evidence package and combining the evidence dimension features, and generating a corresponding number of parallel verification perspectives. Each parallel verification perspective is used to independently analyze the semantic authenticity of the off-chain data from different evidence levels.

[0018] The initial path construction step generates an initial candidate inference path with an independent verification target for each parallel verification perspective, forming an initial candidate inference path set. Each initial candidate inference path contains a local hypothesis and a corresponding evidence reference pointer that supports the hypothesis.

[0019] The path evaluation and screening steps involve quantitatively evaluating and calculating the logical correlation parameters between each candidate reasoning path in the current candidate reasoning path set and the structured evidence package at the current reasoning depth, and then performing fixed-width truncation screening based on the calculation results to select a preset number of candidate reasoning paths as the retained reasoning path set for the current depth.

[0020] The one-to-one deepening update step involves extracting complementary or conflicting features from the structured evidence package for each retained reasoning path in the retained reasoning path set, performing a single downward reasoning to generate candidate reasoning paths corresponding to the next reasoning depth, ensuring that the total number of paths after each round of reasoning deepening is not greater than the preset number; and in each iteration, the above path evaluation and screening steps and the one-to-one deepening update step are repeated sequentially until the preset termination constraint condition is met.

[0021] The integrated judgment and output steps summarize the explored reasoning paths accumulated in each iteration to form a judgment set. Based on the judgment set, the overall authenticity status of off-chain data is identified, and the corresponding off-chain data authenticity judgment result and judgment confidence level are output.

[0022] Preferably, the multidimensional related evidence in the structured evidence package is stored and encapsulated in a unified format, which includes independent data modality fields. These data modality fields cover at least one or more of the following: forensic proof information, original observation data, cross-source evidence information, historical context information, and on-chain context information. Specifically, the statement to be verified represents the off-chain data content submitted to the smart contract, the forensic proof information represents the signature or metadata integrity verification result, and the cross-source evidence information represents independent reference data from multiple external data sources.

[0023] Preferably, in the parallel perspective generation step, the multiple parallel verification perspectives include a time perspective, a cross-source perspective, a semantic perspective, and a mechanism perspective; the time perspective is configured to check whether the statement to be verified meets the event triggering order or time window constraints; the cross-source perspective is configured to check whether there is data feature overlap and causal dependency relationship between multiple heterogeneous evidence sources; and the semantic perspective is configured to check the matching degree between the data entity content and the underlying business logic.

[0024] Preferably, the specific implementation method of the path evaluation and screening step includes:

[0025] At the d-th round of reasoning depth, the current set of candidate reasoning paths is evaluated and a path evaluation score is calculated.

[0026] Based on the path evaluation scores, each candidate reasoning path is sorted in descending order, and the top k candidate reasoning paths are extracted to form the retained reasoning path set for the d-th round; its mathematical expression is as follows:

[0027] ;

[0028] in, Let d represent the set of candidate reasoning paths in round d. This represents an evaluation function that quantitatively calculates the matching degree between candidate reasoning paths and evidence packages, where k represents a preset path retention constant. Let d represent the set of retained reasoning paths obtained from the d-th round of selection.

[0029] The quantitative evaluation calculation process in the path evaluation and screening step generates a total path evaluation score based on three independent constraint dimensions. These independent constraint dimensions include: the content overlap ratio between the candidate reasoning path and the core evidence elements in the structured evidence package; the completeness of the deduction nodes between the local assumptions within the candidate reasoning path and the existing evidence reference pointers; and the number of abnormal labels captured by the candidate reasoning path for phenomena such as time inversion, statistical caliber shift, or source dependency overlap between heterogeneous evidence.

[0030] Preferably, in the one-to-one deepening update step, the retained inference path set is... The specific execution methods for performing a single downward inference include:

[0031] For sets For each path in the process, retrieve the supplementary evidence node data that has not yet been covered and perform single-step logical deduction to generate the first... Round candidate inference path set Its mathematical expression is:

[0032] ;

[0033] Where Deepen(·) represents the one-to-one deepening inference function executed on a single preserved inference path p under the constraints of the structured evidence package, and E represents the structured evidence package and the updated path set. The total number of elements in the set The total number of elements must be kept consistent.

[0034] Preferably, in the one-to-one deepening update step, the preset termination constraint includes two triggering mechanisms: one is that the iteration depth variable reaches the set maximum preset reasoning depth threshold; the other is that the quantitative evaluation score of the current retained reasoning path is higher than the preset confidence threshold and meets the predetermined evidence sufficiency condition, which is limited to the complete extraction of key evidence elements, the complete qualitative characterization of the main contradiction and conflict evidence, and the absence of anomaly verification of the time sequence constraint.

[0035] Preferably, the specific logic of the comprehensive judgment and output step is as follows: perform discrete classification mapping on the judgment set according to the set rules; assign normal state label or abnormal attack label to the off-chain data based on the local hypothesis conflict frequency and its accumulated evidence support parameters of all path records in the judgment set; wherein, the abnormal attack label indicates that the corresponding off-chain data content has structural conflict data blocks that violate the protocol specifications even if it has a compliant digital signature, in terms of internal time mechanism, business semantic expression or cross-source relationship.

[0036] Preferably, in the one-to-one deepening update step, when reasoning downwards on a single retained reasoning path, a data retrieval query request needs to be generated based on the verification characteristics of the current perspective. The auxiliary observation node data that has a positive support correlation with the current perspective and the abnormal log node data that has a logically exclusive correlation are dynamically extracted. The two types of node data are combined with the existing local assumptions to complete the state iteration of the current reasoning branch.

[0037] Preferably, in the initial path construction step, the constructed local hypothesis representation generates a temporary hypothesis proposition for the content to be verified at a single verification level. The evidence reference pointer records the evidence source locator and structured data fragment offset address used to support the temporary hypothesis proposition, ensuring that the multi-level reasoning process derived from the candidate reasoning path can be traced back to the source by extracting the corresponding offset address.

[0038] The blockchain oracle-based off-chain data authenticity verification system based on parallel in-depth reasoning includes the following steps:

[0039] The evidence acquisition module acquires a structured evidence package corresponding to the oracle off-chain data authenticity verification task to be verified. The structured evidence package includes at least the statement to be verified and multi-dimensional related evidence related to the statement to be verified.

[0040] The parallel perspective generation module determines the number of initial verification perspectives based on the structured evidence package and the evidence dimension features, and generates a corresponding number of parallel verification perspectives. Each parallel verification perspective is used to independently analyze the semantic authenticity of the off-chain data from different evidence levels.

[0041] The initial path construction module generates initial candidate inference paths with independent verification objectives for each parallel verification perspective, forming an initial candidate inference path set. Each initial candidate inference path contains a local hypothesis and corresponding evidence reference pointers that support the hypothesis.

[0042] The path evaluation and screening module performs quantitative evaluation and calculation based on the logical association parameters between each candidate reasoning path in the current candidate reasoning path set and the structured evidence package at the current reasoning depth. Based on the calculation results, it performs fixed-width truncation screening and selects a preset number of candidate reasoning paths as the retained reasoning path set for the current depth.

[0043] The one-to-one deepening update module extracts complementary or conflicting features from the structured evidence package for each retained reasoning path in the retained reasoning path set and performs a single downward reasoning to generate candidate reasoning paths corresponding to the next reasoning depth, so that the total number of paths after each round of reasoning deepening is not greater than the preset number; and in each iteration, the above path evaluation and screening steps and the one-to-one deepening update steps are repeated sequentially until the preset termination constraint condition is reached.

[0044] The integrated judgment and output module summarizes the explored reasoning paths accumulated in each round of iteration to form a judgment set. Based on the judgment set, it identifies the overall authenticity status of off-chain data and outputs the corresponding off-chain data authenticity judgment result and judgment confidence level.

[0045] The beneficial effects of this invention are as follows: It can supplement the shortcomings of existing signature verification mechanisms. This invention does not replace digital signatures, hash verification, or authentication mechanisms, but rather further verifies the semantic authenticity of off-chain data, enabling it to identify data with valid signatures but anomalies in terms of time, semantics, cross-source, or context; the evidence coverage is more comprehensive. This invention uniformly organizes the statements to be verified, forensic proofs, original observation data, cross-source evidence, historical context, on-chain context, and off-chain context through structured evidence packages, allowing the verification process to comprehensively consider multiple heterogeneous evidences rather than relying on a single indicator; the inference path is more stable. This invention adopts a parallel perspective generation mechanism, enabling the verification process to analyze off-chain data from multiple angles such as time consistency, cross-source consistency, semantic consistency, causal mechanisms, and on-chain context, reducing the risk of missing key evidence in single-path inference; the computational cost is more controllable. This invention employs a top-k path filtering and one-to-one deepening mechanism, ensuring that the number of inference paths after each round of deepening does not exceed the number of retained paths, avoiding excessive inference overhead caused by unconstrained branch expansion; the judgment results have stronger interpretability. This invention retains local hypotheses and evidence citation information in each candidate reasoning path, enabling the final judgment to be traced back to specific evidence sources and reasoning paths. This helps oracle systems, smart contract developers, or auditors understand why off-chain data is judged as normal or abnormal; it is applicable to various semantic attack scenarios. This invention can be used to identify off-chain data authenticity anomalies such as time mechanism manipulation, semantic ambiguity deception, cross-evidence conflicts, and source dependency pollution, helping to improve the security and reliability of blockchain oracles in complex real-world scenarios. Attached Figure Description

[0046] Figure 1 This is an overall flowchart of the blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning provided in the embodiments of the present invention;

[0047] Figure 2 This is a schematic diagram of the blockchain oracle off-chain data authenticity verification framework based on parallel deep reasoning provided in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the structure of the blockchain oracle off-chain data authenticity verification device based on parallel deep reasoning provided in an embodiment of the present invention. Detailed Implementation

[0049] 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, and 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.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0051] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings:

[0052] Blockchain oracles are critical infrastructure connecting on-chain smart contracts with off-chain external information. In practical applications such as decentralized finance, insurance settlement, and supply chain automation, the execution logic of smart contracts heavily relies on off-chain observation data provided by oracles. However, this dependence constitutes a severe security vulnerability: if off-chain observation data is tampered with before being uploaded to the blockchain, even if the smart contract executes correctly at the protocol level, it may produce erroneous results at the economic semantic level.

[0053] The core of existing defense mechanisms is verification based on digital signatures. This involves verifying whether the submitted record originates from the intended entity and has not been modified during transmission to ensure source credibility and bit-level integrity. However, digital signatures only guarantee the source proof and transmission integrity of data; they cannot guarantee the authenticity of reported content in terms of timing constraints, cross-source consistency, and on-chain context semantics. In real-world attack scenarios, data may remain cryptographically valid but semantically misleading, for example, through manipulation of timing windows, unit substitution, delayed source synchronization, or source dependency pollution. Rule-based detectors also have limitations. They rely on manually designed rule templates and are significantly inadequate in terms of protocol portability and robustness against unknown attack variants.

[0054] To address the aforementioned technological gaps, this invention proposes a parallel in-depth reasoning framework, hereinafter referred to as the PDR framework. This framework is designed for structured evidence verification scenarios and organically integrates parallel perspective generation, iterative evaluation, screening, in-depth updating, and comprehensive judgment and decision-making to achieve systematic verification of the semantic authenticity of off-chain data.

[0055] This invention formalizes the verification of off-chain data authenticity into a binary judgment problem for structured evidence packages. Let the evidence package be:

[0056] ;

[0057] in, This represents the off-chain data content submitted to the smart contract, i.e., the statement to be verified; It includes proof of origin and cryptographic validity signals, covering digital signature verification and metadata integrity verification results, i.e., evidence verification information; Describes the raw observation data transmitted through the oracle pipeline; It provides independent reference data from multiple external data sources, i.e., cross-source evidence information; and Separately encode the temporal context and the on-chain execution context, corresponding to historical context information and on-chain context information; This is optional off-chain context metadata. All of the above components together constitute the multidimensional related evidence in the structured evidence package, stored and encapsulated using a unified format, with each related evidence component existing as an independent data modality field.

[0058] Based on evidence package E, verify the model output labels. and confidence score :

[0059] ;

[0060] in, It is a binary set consisting of normal state labels or abnormal attack labels. Belongs to the interval This characterizes the degree to which the current evidence package supports the predicted label under the consistency check within each evidence channel of the model. It is important to note that the prediction of the anomaly attack label does not necessarily mean a failure at the cryptographic level, but rather indicates that the semantic interpretation of the reported content is inconsistent at the cross-source, temporal, or on-chain context levels. That is, even if the corresponding off-chain data content has a compliant digital signature, there are structurally conflicting data blocks that violate the protocol specifications in terms of internal timing mechanisms, business semantic expression, or cross-source relationships.

[0061] The complete execution flow of the method of this invention includes six core steps in sequence: evidence acquisition step, parallel perspective generation step, initial path construction step, path evaluation and screening step, one-to-one deepening update step, and comprehensive judgment and output step. The steps are connected sequentially through clear logical relationships to form a complete reasoning and verification closed loop.

[0062] The first step is to perform the evidence acquisition step. In this step, the system acquires the structured evidence package corresponding to the oracle off-chain data authenticity verification task to be verified. The structured evidence package includes at least the statement to be verified and multi-dimensional correlation evidence related to the statement to be verified.

[0063] The structured evidence package serves as the unified input carrier for the verification process of this invention. Its design principle is to organize all heterogeneous evidence signals required for verifying the authenticity of off-chain data into a controllable structured form, thereby supporting independent extraction and correlation analysis of different evidence dimensions in subsequent reasoning steps. The multidimensional correlated evidence in the structured evidence package is stored and encapsulated using a unified format, containing independent data modality fields. These data modality fields cover at least one or more of the following: forensic proof information, original observation data, cross-source evidence information, historical context information, and on-chain context information. The statement to be verified represents the off-chain data content submitted to the smart contract; the forensic proof information represents the signature or metadata integrity verification result; and the cross-source evidence information represents independent reference data from multiple external data sources.

[0064] The evidence information field stores the signature validity verification result and metadata integrity markers to confirm the credibility of the data source and the integrity of the transmission. However, this field can only exclude transport layer attacks through effective verification, and cannot exclude semantic-level deception. The original observation data field stores the first-hand observations transmitted through the oracle pipeline as a benchmark for subsequent semantic verification. The cross-source evidence information field aggregates independent reference data from multiple heterogeneous external sources, providing a data foundation for cross-source consistency analysis. The historical context information field records historical time-series trajectories, supporting the verification of time-series anomalies and trend deviations. The on-chain context information field encapsulates on-chain transaction records and contract execution status, used to verify the causal matching relationship between off-chain claims and on-chain behaviors.

[0065] After evidence acquisition is completed, the system proceeds to the parallel perspective generation step. In this step, the system determines the number of initial verification perspectives based on the structured evidence package and the dimensional features of the evidence, and generates a corresponding number of parallel verification perspectives. Each parallel verification perspective is used to independently analyze the semantic authenticity of the off-chain data from different evidence levels.

[0066] The parallel perspective generation step corresponds to the core operation of Phase One in the PDR framework. Given an evidence package E, the system adaptively determines the initial number of perspectives N and the corresponding initial set of candidate reasoning paths, which can be formally expressed as:

[0067] ;

[0068] ;

[0069] in, This represents a function that adaptively selects the number of initial viewpoints based on the features of the evidence package. The initial number of viewpoints is determined by this function. Indicates by The set of the 0th layer reasoning frontier, consisting of the initial candidate reasoning paths. Indicates the 0th layer. Initial candidate reasoning paths, Indicates based on evidence package Number of viewpoints A generator function that generates the initial set of multi-view candidate paths.

[0070] The generated parallel verification perspectives include a temporal perspective, a cross-source perspective, a semantic perspective, and a mechanism perspective. The temporal perspective is configured to examine whether the statement to be verified meets the event triggering order or time window constraints. Specifically, this perspective is responsible for identifying structural contradictions in the reported narrative at the temporal level, such as missing causal premises, inverted event sequences, or abnormal trigger windows, which violate the execution sequence. Such temporal manipulation attacks are common in on-chain financial scenarios such as flash loan pollution attacks. The cross-source perspective is configured to examine whether multiple heterogeneous evidence sources have overlapping data features and causal dependencies, focusing on identifying whether multi-source consistency is an apparent phenomenon rather than genuine independence, especially distinguishing between feedback loops and pseudo-independent sources. The semantic perspective is configured to examine the matching degree between the data entity content and the underlying business logic, focusing on examining attack methods such as semantic contradictions within fields, replacement of unit of measurement, and template retention but semantic payload replacement. The mechanism perspective, starting from protocol constraints and business causal constraints, examines whether the declared operation sequence conforms to the corresponding protocol execution premises and mechanism dependencies. In addition to the four perspectives mentioned above, the system can also adaptively introduce supplementary perspectives such as statistical and causal perspectives based on the dimensional characteristics of the specific evidence package, so as to ensure that each data modality field in the multidimensional correlation evidence can be independently and fully covered and analyzed.

[0071] After generating parallel perspectives, the system proceeds to the initial path construction step. In this step, for each parallel verification perspective, the system generates an initial candidate inference path with an independent verification objective, forming an initial candidate inference path set. Each initial candidate inference path contains a local hypothesis and corresponding evidence reference pointers supporting that hypothesis.

[0072] In the initial path construction step, the constructed local hypothesis representation generates temporary hypothesis propositions for the content of the statement to be verified at a single verification level. These temporary hypothesis propositions are not the final judgment conclusion, but rather a preliminary speculative representation of the statement's truthfulness from a specific verification perspective, used to guide the direction of subsequent in-depth reasoning and evidence extraction strategies. The evidence reference pointer records the evidence source locator and the structured data fragment offset address used to support the temporary hypothesis proposition. The evidence source locator identifies the data modality field category to which the evidence fragment belongs, while the structured data fragment offset address precisely locates the storage position of the evidence fragment within the corresponding field. Through this dual recording mechanism, it is ensured that the multi-level reasoning deduction process derived from this candidate reasoning path can be traced back to its source by extracting the corresponding offset address, thereby providing a complete reasoning chain tracing capability for the final judgment result.

[0073] The size of the initial candidate inference path set is consistent with the number of parallel perspectives N. Each path is anchored to a specific verification perspective, carrying the local hypothesis and evidence reference pointers corresponding to that perspective, which serve as the original input for subsequent iterative evaluation and deepening update operations.

[0074] After the initial path construction is completed, the system proceeds to the path evaluation and screening step. In this step, at the current inference depth, the system performs quantitative evaluation calculations based on the logical correlation parameters between each candidate inference path in the current candidate inference path set and the structured evidence package. Based on the calculation results, a fixed-width truncation screening is performed to select a preset number of candidate inference paths as the retained inference path set for the current depth.

[0075] The specific execution method of the path evaluation and screening steps is as follows: At the d-th round of reasoning depth, the current candidate reasoning path set is evaluated and a path evaluation score is calculated; based on the path evaluation score, each candidate reasoning path is sorted in descending order, and the top k candidate reasoning paths are extracted to form the retained reasoning path set for the d-th round; its mathematical expression is:

[0076] ;

[0077] in, Indicates the first Round candidate inference path set, denoted as the evaluation function that quantitatively calculates the matching degree between candidate reasoning paths and evidence packages, k represents the preset path retention constant, and S(d) represents the set of retained reasoning paths obtained in the d-th round of screening. This represents a truncation and filtering operation that extracts the top k paths from the score ranking.

[0078] The quantitative evaluation calculation process in the path evaluation and screening step generates a total path evaluation score based on three independent constraint dimensions. These independent constraint dimensions include: the content overlap ratio between the candidate reasoning path and the core evidence elements in the structured evidence package; the completeness of the deduction nodes between the local assumptions within the candidate reasoning path and the existing evidence reference pointers; and the number of abnormal labels captured by the candidate reasoning path for phenomena such as time inversion, statistical caliber shift, or source dependency overlap between heterogeneous evidence.

[0079] The first constraint dimension measures the extent to which candidate paths cover core evidentiary elements, preventing paths from being deduced solely from prior knowledge without the support of key evidence. The second constraint dimension measures the logical consistency of the path's internal reasoning structure, ensuring clear connections between local hypotheses and cited evidence, rather than abrupt assertions. The third constraint dimension measures the path's sensitivity to capturing anomalous signals among evidence, specifically quantifying three typical semantic attack signals: time sequence reversal, caliber shift, and overlapping source dependencies. The scores from these three dimensions are weighted and combined to form the final path evaluation score, retaining the set of reasoning paths. The path consists of the top k paths with the highest scores. Paths that are not selected are pruned and discarded at the current depth, thereby achieving centralized allocation of computing resources under fixed-width truncation filtering.

[0080] After completing path evaluation and screening, the system proceeds to the one-to-one deepening update step. In this step, for each retained reasoning path in the retained reasoning path set, the system extracts complementary or conflicting features from the structured evidence package for a single downward reasoning, generating candidate reasoning paths corresponding to the next reasoning depth, ensuring that the total number of paths after each round of reasoning deepening does not exceed the preset number; and in each iteration, the above path evaluation and screening steps and the one-to-one deepening update step are repeated sequentially until the preset termination constraint condition is met.

[0081] The one-to-one deepening update step is the core design feature that distinguishes the PDR framework from tree-based and graph-based search paradigms. This step targets the set of preserved inference paths. The specific steps for performing a single downward inference are as follows: For a set For each path in the process, retrieve the supplementary evidence node data that has not yet been covered and perform single-step logical deduction to generate the first... Round candidate inference path set Its mathematical expression is:

[0082] ;

[0083] Where Deepen(·) represents a single preserved reasoning path under the constraints of the structured evidence package. The one-to-one in-depth inference function is executed. This refers to the structured evidence package. express The single preserved reasoning path in This represents the set of retained reasoning paths obtained from the d-th round of selection. The updated set of paths... The total number of elements in the set The total number of elements is kept consistent, meaning that each retained path generates exactly one deepening path, and the total number of paths remains constant throughout the iteration process. This ensures that the search space is controlled and eliminates the problem of computational overhead caused by unconstrained bifurcation.

[0084] When reasoning downwards along a single retained reasoning path, a data retrieval query request needs to be generated based on the verification characteristics of the current perspective. This dynamically extracts auxiliary observation node data that has a positive supporting correlation with the current perspective, as well as abnormal log node data that has a logically contradictory correlation. The acquired two types of node data are then combined with existing local hypotheses to complete the state iteration of the current reasoning branch. Auxiliary observation node data is used to strengthen the evidence supporting the current local hypothesis, while abnormal log node data is used to introduce potential contradictory signals, thereby triggering hypothesis correction or a reduction in path score. This two-way dynamic evidence extraction mechanism ensures that each step of in-depth reasoning accumulates supporting evidence while capturing adversarial signals, avoiding confirmation bias caused by one-way evidence accumulation.

[0085] In the one-to-one deepening update step, the preset termination constraint includes two triggering mechanisms: First, the iteration depth variable reaches the set maximum preset inference depth threshold. At this point, regardless of the scoring status of the currently retained path, the iteration is forcibly terminated. Second, the quantitative evaluation score of the currently retained inference path is higher than a preset confidence threshold and a predetermined evidence sufficiency condition is met. The predetermined evidence sufficiency condition is limited to the complete extraction of key evidence elements, the complete qualitative characterization of the main conflicting evidence, and the absence of anomalies in the time sequence constraint verification. The second triggering mechanism constitutes an early termination mechanism. When the inference has formed a high-confidence judgment with sufficient evidence, the system can terminate the iteration early without waiting for the depth limit, thereby effectively reducing the computational overhead of inference while ensuring the quality of the judgment. There is a logical OR relationship between the two triggering mechanisms; termination is triggered when either condition is met.

[0086] The path evaluation and selection steps, along with the one-to-one deepening update steps, are executed alternately in each iteration, forming a complete iterative closed loop in Phase Two of the PDR framework. Each iteration first performs a quantitative evaluation of the current candidate path set and retains the top k paths. Then, a one-to-one deepening process is performed on each retained path to generate the next depth candidate set, until the termination constraint is triggered. This design organically unifies the control of exploration breadth with the advancement of exploration depth, achieving a balance between the sufficiency of evidence coverage and the controllability of computational overhead in the reasoning process.

[0087] After the iteration terminates, the system proceeds to the comprehensive judgment and output step. In this step, the system summarizes the explored inference paths accumulated in each iteration to form a judgment set. Based on the judgment set, the overall authenticity status of the off-chain data is identified, and the corresponding off-chain data authenticity judgment result and judgment confidence level are output.

[0088] The comprehensive judgment and output steps correspond to the comprehensive decision-making operation in Phase 3 of the PDR framework, and their formal expression is as follows:

[0089] ;

[0090] in, This represents the set of explored paths, which aggregates all explored reasoning path records across depths. This represents a decision function that performs a comprehensive assessment based on the set of explored paths. For the final output data authenticity label, This represents the corresponding confidence score.

[0091] The specific logic of the comprehensive judgment and output steps is as follows: A discrete classification mapping is performed on the judgment set according to the set rules; based on the local hypothesis conflict frequency and accumulated evidence support parameters of all path records in the judgment set, a normal state label or an abnormal attack label is assigned to the off-chain data. The abnormal attack label indicates that even if the corresponding off-chain data content has a compliant digital signature, it contains structurally conflicting data blocks that violate protocol specifications in terms of internal time mechanisms, business semantic expression, or cross-source relationships.

[0092] In the comprehensive judgment process, the judgment set not only gathers the final retained paths but also includes candidate path records that were pruned and discarded during intermediate depth iterations but have already been scored, thus making full use of all the reasoning information accumulated during the iteration process. The discrete classification mapping rule is based on the joint analysis of conflict frequency statistics and evidence support parameters: when the dominant local hypothesis direction in the judgment set points to a normal semantic interpretation that is fully supported by evidence, a normal state label is output; when multiple path records in the judgment set present high-frequency conflict markers regarding temporal anomalies, caliber shifts, or source dependency contamination, and the cumulative evidence support parameters exceed the anomaly judgment threshold, an anomaly attack label is output, and a corresponding confidence score c is output simultaneously. This score quantifies the degree of consistency of the current judgment conclusion with the evidence presented by all explored paths.

[0093] This invention also provides a blockchain oracle off-chain data authenticity verification system based on parallel in-depth reasoning, comprising the following modules: evidence acquisition module, parallel perspective generation module, initial path construction module, path evaluation and screening module, one-to-one in-depth update module, and comprehensive judgment and output module. The modules are executed in a logical order through data flow transmission.

[0094] The evidence acquisition module is responsible for acquiring the structured evidence package corresponding to the oracle off-chain data authenticity verification task to be verified. The structured evidence package includes at least the statement to be verified and multi-dimensional correlation evidence related to the statement. As the system input interface, the evidence acquisition module encapsulates heterogeneous evidence signals in a unified format mode to form a structured evidence package containing independent data modality fields, and then passes it to the parallel perspective generation module.

[0095] The parallel perspective generation module, based on the structured evidence package and combined with the dimensional features of the evidence, determines the initial number of verification perspectives and generates a corresponding number of parallel verification perspectives. Each parallel verification perspective is used to independently analyze the semantic authenticity of the off-chain data from different evidentiary levels. The parallel perspective generation module adaptively determines the optimal number of perspectives N by analyzing the dimensional features of each data modality field in the evidence package, and assigns a dedicated evidence focus area and verification target to each perspective, ensuring that different perspectives are independent of each other and that the evidence coverage is complementary.

[0096] The initial path construction module generates initial candidate inference paths with independent verification objectives for each parallel verification perspective, forming an initial candidate inference path set. Each initial candidate inference path contains a local hypothesis and corresponding evidence reference pointers supporting that hypothesis. The initial path construction module synchronously allocates local hypotheses and evidence reference pointers to each initial candidate inference path. The evidence reference pointers ensure the integrity and traceability of the inference process through a dual recording mechanism of evidence source locators and structured data fragment offset addresses.

[0097] The path evaluation and filtering module performs quantitative evaluation calculations based on the logical correlation parameters between each candidate inference path in the current candidate inference path set and the structured evidence package at the current inference depth. Based on the calculation results, it performs a fixed-width truncation filtering, selecting a preset number of candidate inference paths as the retained inference path set for the current depth. The path evaluation and filtering module generates a total path evaluation score based on three independent constraint dimensions: content overlap ratio, inference node completeness, and the number of anomaly tag captures. It then performs a TopK truncation filtering operation and outputs the retained inference path set. .

[0098] The one-to-one deepening update module extracts complementary or conflicting features from the structured evidence package for each retained reasoning path in the retained reasoning path set, performing a single downward reasoning to generate candidate reasoning paths for the next reasoning depth, ensuring that the total number of paths after each round of reasoning deepening does not exceed the preset number; and in each iteration, the above path evaluation and screening steps and the one-to-one deepening update steps are repeated sequentially until the preset termination constraint is met. The one-to-one deepening update module dynamically extracts auxiliary observation node data and abnormal log node data by executing data retrieval query requests for each retained path, and completes the state iteration of the current reasoning branch through combined operations to generate a set of candidate paths for the next depth. and maintain the total number of paths with The mandatory consistency constraint.

[0099] The comprehensive judgment and output module summarizes the explored inference paths accumulated in each iteration to form a judgment set. Based on this judgment set, it identifies the overall authenticity status of off-chain data and outputs the corresponding off-chain data authenticity judgment result and judgment confidence level. Based on the local hypothesis conflict frequency and accumulated evidence support parameters of all path records in the judgment set, the comprehensive judgment and output module performs discrete classification mapping, and finally outputs a normal state label or an abnormal attack label and the corresponding confidence score, completing the judgment output process of the entire verification system.

[0100] The method and system described in this invention can cover four major attack groups in off-chain data authenticity attacks: timing mechanism manipulation attacks, semantic caliber deception attacks, cross-evidence conflict attacks, and source dependency pollution attacks. Attacks targeting temporal order and mechanism consistency include subtypes such as flash loan pollution, causal chain breakage, and counterfactual failure. The off-chain claims constructed by these attacks appear reasonable on the surface, but structural contradictions violating execution timing constraints or business causal premises can be found once the timeline and context are jointly examined. Semantic deception attacks change semantic meaning while preserving syntactic correctness, encompassing three methods: field semantic conflicts, unit of measurement substitution, and template migration. The deception of these attacks lies in semantic interpretation rather than format validity. Cross-evidence conflict attacks emphasize the inconsistency between heterogeneous evidence channels. Core claims cannot be stably supported when jointly verified by on-chain logs, off-chain data feeds, and auxiliary modal signals, or key supporting signals are abnormally missing. Source dependency pollution attacks challenge the assumption that multi-source consistency represents reliability. In these attacks, source diversity is an apparent phenomenon rather than true independence. Specific forms include feedback loops where downstream outputs recursively inject upstream references and pseudo-independent sources where multiple reference sources share implicit upstream dependencies. This invention ensures that all dimensions of evidence involved in the above-mentioned types of attacks are covered by independent verification perspectives through parallel perspective design, and ensures that cross-perspective evidence conflict signals are analyzed in detail layer by layer through iterative deepening mechanism, thereby achieving systematic detection of semantic-level attacks.

[0101] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning, characterized in that, Includes the following steps: The evidence acquisition step involves acquiring a structured evidence package corresponding to the oracle off-chain data authenticity verification task to be verified. The structured evidence package includes at least the statement to be verified and multi-dimensional related evidence associated with the statement to be verified. The parallel perspective generation step involves determining the number of initial verification perspectives based on the structured evidence package and combining the evidence dimension features, and generating a corresponding number of parallel verification perspectives. Each parallel verification perspective is used to independently analyze the semantic authenticity of the off-chain data from different evidence levels. The initial path construction step generates an initial candidate inference path with an independent verification target for each parallel verification perspective, forming an initial candidate inference path set. Each initial candidate inference path contains a local hypothesis and a corresponding evidence reference pointer that supports the hypothesis. The path evaluation and screening steps involve quantitatively evaluating and calculating the logical correlation parameters between each candidate reasoning path in the current candidate reasoning path set and the structured evidence package at the current reasoning depth, and then performing fixed-width truncation screening based on the calculation results to select a preset number of candidate reasoning paths as the retained reasoning path set for the current depth. The one-to-one deepening update step involves extracting complementary or conflicting features from the structured evidence package for each retained reasoning path in the retained reasoning path set, performing a single downward reasoning to generate candidate reasoning paths corresponding to the next reasoning depth, ensuring that the total number of paths after each round of reasoning deepening is not greater than the preset number; and in each iteration, the above path evaluation and screening steps and the one-to-one deepening update step are repeated sequentially until the preset termination constraint condition is met. The integrated judgment and output steps summarize the explored reasoning paths accumulated in each iteration to form a judgment set. Based on the judgment set, the overall authenticity status of off-chain data is identified, and the corresponding off-chain data authenticity judgment result and judgment confidence level are output.

2. The blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning according to claim 1, characterized in that, The multidimensional related evidence in the structured evidence package is stored and encapsulated in a unified format, and contains independent data modality fields. The data modality fields cover at least one or more of the following: evidence-gathering proof information, original observation data, cross-source evidence information, historical context information, and on-chain context information. Among them, the statement to be verified represents the off-chain data content submitted to the smart contract, the evidence-gathering proof information represents the signature or metadata integrity verification result, and the cross-source evidence information represents independent reference data from multiple external data sources.

3. The blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning according to claim 1, characterized in that, In the parallel perspective generation step, the multiple parallel verification perspectives include a time perspective, a cross-source perspective, a semantic perspective, and a mechanism perspective. The time perspective is configured to check whether the statement to be verified meets the event triggering order or time window constraints. The cross-source perspective is configured to check whether there is overlap in data features and causal dependency among multiple heterogeneous evidence sources. The semantic perspective is configured to check the matching degree between the data entity content and the underlying business logic.

4. The blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning according to claim 1, characterized in that, The specific implementation methods of the path evaluation and screening steps include: At the d-th round of reasoning depth, the current set of candidate reasoning paths is evaluated and a path evaluation score is calculated. Based on the path evaluation scores, each candidate reasoning path is sorted in descending order, and the top k candidate reasoning paths are extracted to form the retained reasoning path set for the d-th round; its mathematical expression is as follows: ； in, Let d represent the set of candidate reasoning paths in round d. This represents an evaluation function that quantitatively calculates the matching degree between candidate reasoning paths and evidence packages, where k represents a preset path retention constant. Let d represent the set of retained reasoning paths obtained from the d-th round of selection.

5. The blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning according to claim 4, characterized in that, In the one-to-one deepening update step, for the set of retained inference paths The specific execution methods for performing a single downward inference include: For sets For each path in the process, retrieve the supplementary evidence node data that has not yet been covered and perform single-step logical deduction to generate the first... Round candidate inference path set Its mathematical expression is: ； Where Deepen(·) represents the one-to-one deepening inference function executed on a single preserved inference path p under the constraints of the structured evidence package, and E represents the structured evidence package and the updated path set. The total number of elements in the set The total number of elements must be kept consistent.

6. The method for verifying the authenticity of off-chain data using a blockchain oracle based on parallel deep reasoning according to claim 1, characterized in that, In the one-to-one deepening update step, the preset termination constraint includes two triggering mechanisms: one is that the iteration depth variable reaches the set maximum preset inference depth threshold. Secondly, the quantitative evaluation score of the current retained reasoning path is higher than the preset confidence threshold and meets the predetermined evidence sufficiency conditions. The predetermined evidence sufficiency conditions are limited to the complete extraction of key evidence elements, the complete qualitative characterization of the main contradiction and conflict evidence, and the absence of abnormal verification of the time sequence constraints.

7. The blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning according to claim 1, characterized in that, The specific logic of the comprehensive judgment and output step is as follows: perform discrete classification mapping on the judgment set according to the set rules; based on the local hypothesis conflict frequency and its accumulated evidence support parameters of all path records in the judgment set, assign normal state label or abnormal attack label to the off-chain data; wherein, the abnormal attack label indicates that the corresponding off-chain data content has structural conflict data blocks that violate the protocol specifications even if it has a compliant digital signature, in terms of internal time mechanism, business semantic expression or cross-source relationship.

8. The blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning according to claim 1, characterized in that, In the one-to-one deepening update step, when reasoning downwards on a single retained reasoning path, a data retrieval query request needs to be generated based on the verification characteristics of the current perspective. The auxiliary observation node data that has a positive support correlation with the current perspective and the abnormal log node data that has a logically exclusive correlation are dynamically extracted. The two types of node data are combined with the existing local assumptions to complete the state iteration of the current reasoning branch.

9. The blockchain oracle off-chain data authenticity verification method based on parallel deep reasoning according to claim 1, characterized in that, In the initial path construction step, the constructed local hypothesis representation generates temporary hypothesis propositions for the content to be verified at a single verification level. The evidence reference pointer records the evidence source locator and structured data fragment offset address used to support the temporary hypothesis proposition, ensuring that the multi-level reasoning process derived from the candidate reasoning path can be traced back to the source by extracting the corresponding offset address.

10. A blockchain oracle-based off-chain data authenticity verification system based on parallel deep reasoning, characterized in that, Includes the following steps: The evidence acquisition module acquires a structured evidence package corresponding to the oracle off-chain data authenticity verification task to be verified. The structured evidence package includes at least the statement to be verified and multi-dimensional related evidence related to the statement to be verified. The parallel perspective generation module determines the number of initial verification perspectives based on the structured evidence package and the evidence dimension features, and generates a corresponding number of parallel verification perspectives. Each parallel verification perspective is used to independently analyze the semantic authenticity of the off-chain data from different evidence levels. The initial path construction module generates initial candidate inference paths with independent verification objectives for each parallel verification perspective, forming an initial candidate inference path set. Each initial candidate inference path contains a local hypothesis and corresponding evidence reference pointers that support the hypothesis. The path evaluation and screening module performs quantitative evaluation and calculation based on the logical association parameters between each candidate reasoning path in the current candidate reasoning path set and the structured evidence package at the current reasoning depth. Based on the calculation results, it performs fixed-width truncation screening and selects a preset number of candidate reasoning paths as the retained reasoning path set for the current depth. The one-to-one deepening update module extracts complementary or conflicting features from the structured evidence package for each retained reasoning path in the retained reasoning path set and performs a single downward reasoning to generate candidate reasoning paths corresponding to the next reasoning depth, so that the total number of paths after each round of reasoning deepening is not greater than the preset number; and in each iteration, the above path evaluation and screening steps and the one-to-one deepening update steps are repeated sequentially until the preset termination constraint condition is reached. The integrated judgment and output module summarizes the explored reasoning paths accumulated in each round of iteration to form a judgment set. Based on the judgment set, it identifies the overall authenticity status of off-chain data and outputs the corresponding off-chain data authenticity judgment result and judgment confidence level.

Images