Event authenticity verification method, device, equipment and medium

Abstract

The present application relates to the technical field of image processing, which can be applied to business scenarios such as financial technology, and discloses an event authenticity verification method, device, equipment and medium, comprising: based on the event description text information and the on-site certificate media file in the transaction declaration request, analyzing the claimed space-time information and extracting the actual collection space-time information, and performing physical authenticity verification. In the case of passing the verification, the environment feature of the on-site certificate media file is analyzed and the historical environment reference data is obtained, the event description text information, the environment feature information and the historical environment reference data are input into a multi-modal reasoning model, a logical consistency analysis result is generated and a corresponding processing operation is performed. The present application introduces a multi-modal consistency analysis mechanism to jointly verify text, image and objective environment information, thereby improving the accuracy of abnormal transaction identification, reducing the need for manual review, and improving the efficiency and reliability of automatic processing of financial technology business.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, device, and medium for verifying the authenticity of an event. Background Technology

[0002] In the fintech sector, the underwriting, claims, and compensation processes for motor vehicle-related matters are continuously evolving towards online and self-service models. With the widespread adoption of mobile devices and network communication capabilities, these processes are gradually shifting from traditional manual processing, on-site verification, and offline review to a model where users submit transaction requests via mobile devices, along with text descriptions and on-site supporting documentation. This model offers significant advantages in improving response speed and reducing labor costs, and has become the mainstream operating mode for current financial service platforms.

[0003] However, existing online processing systems still primarily rely on rule-driven or manual review mechanisms to verify the authenticity of transactions. Most systems focus on verifying the formal completeness of the submitted information, such as whether fields are missing or whether the time falls within a valid range. They lack a systematic and scalable verification capability to ensure the content of the submission itself aligns with objective facts. In this situation, the textual descriptions in the transaction request and the on-site supporting documents are often treated as independent data carriers, failing to form an effective cross-verification relationship, making it difficult to identify false or inconsistent submissions in a timely manner.

[0004] Furthermore, existing technologies, when processing on-site credential media files, primarily focus on analyzing the visual content itself, such as identifying the presence of specific targets or damage features in the image, while paying insufficient attention to the authenticity of the spatiotemporal attributes of the media file, such as its generation time and acquisition location. Due to the lack of effective verification of the consistency between the claimed spatiotemporal information in the declaration and the actual spatiotemporal information of the media file acquisition, existing systems struggle to identify false declarations formed by reusing historical data, tampering with source information, or misappropriating materials from non-current transactions, thus creating potential risks in the business processing stage.

[0005] Furthermore, in the transaction reporting process, textual descriptions are typically presented in natural language, containing a large amount of implicit temporal, spatial, and environmental semantics, while on-site credential media files reflect the objective environmental conditions in the form of images or videos. Existing processing mechanisms generally lack the ability to uniformly analyze this heterogeneous information, making it difficult to determine whether there are logical conflicts between the textual descriptions and the visible environmental characteristics. For example, when there is a significant inconsistency between descriptive information and the environmental conditions at a basic cognitive level, the system often cannot form an effective judgment and can only rely on manual review as a fallback, resulting in decreased processing efficiency. Summary of the Invention

[0006] The main objective of this invention is to provide a method, apparatus, device, and storage medium for verifying the authenticity of events. This invention aims to solve the technical problem that existing technologies rely solely on a single information dimension for formal verification, lacking comprehensive verification of the consistency between the declaration text, on-site credentials, and objective environmental reference information. This results in the difficulty of timely and accurately identifying false or abnormal transaction declaration requests that are inconsistent with the declared content.

[0007] To achieve the above objectives, the present invention provides a method for verifying the authenticity of an event, comprising: Receive pending transaction reporting requests, wherein the transaction reporting requests include event description text information and on-site credential media files; Parse the claimed spatiotemporal information of the event description text, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata; The actual spatiotemporal information collected is compared with the claimed spatiotemporal information, and a physical authenticity verification operation is performed based on the comparison result and a preset spatiotemporal verification strategy. If the physical authenticity verification operation fails, a first verification conclusion is generated and a first type of processing operation is performed; If the physical authenticity verification operation passes, then the visual feature analysis of the on-site credential media file is performed to obtain environmental feature information; Based on the claimed spatiotemporal information, obtain the corresponding historical environmental reference data; The event description text information, the environmental feature information, and the historical environmental reference data are input into the multimodal reasoning model to obtain the logical consistency analysis results. A second verification conclusion is generated based on the logical consistency analysis results, and a second type of processing operation is performed based on the second verification conclusion.

[0008] Furthermore, to achieve the above objectives, the present invention provides an event authenticity verification device, comprising: The transaction access module is used to receive pending transaction declaration requests, which include event description text information and on-site credential media files. The spatiotemporal information parsing module is used to parse the claimed spatiotemporal information from the event description text information, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata; The spatiotemporal consistency verification module is used to compare the actual collected spatiotemporal information with the claimed spatiotemporal information, and to perform physical authenticity verification operations based on the comparison results and a preset spatiotemporal verification strategy. The first type of decision-making module is used to generate a first verification conclusion and execute a first type of processing operation if the physical authenticity verification operation fails. The environmental visual analysis module is used to perform visual feature analysis on the on-site credential media file to obtain environmental feature information if the physical authenticity verification operation passes. The historical environment data acquisition module is used to acquire corresponding historical environment reference data based on the claimed spatiotemporal information. The multimodal logical reasoning module is used to input the event description text information, the environmental feature information, and the historical environmental reference data into the multimodal reasoning model to obtain the logical consistency analysis results. The second type of decision-making module is used to generate a second verification conclusion based on the logical consistency analysis results, and to perform a second type of processing operation according to the second verification conclusion.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and an event authenticity verification program stored in the memory and executable on the processor, wherein when the event authenticity verification program is executed by the processor, it implements the steps of the event authenticity verification method as described above.

[0010] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an event authenticity verification program, which, when executed by a processor, implements the steps of the event authenticity verification method as described above.

[0011] Beneficial Effects: This invention relates to the field of image processing technology and can be applied to business scenarios such as fintech. It discloses a method, apparatus, device, and medium for verifying the authenticity of events, including: receiving a transaction declaration request containing event description text information and on-site credential media files; parsing the event description text information to obtain the claimed spatiotemporal information of the event; extracting the actual collection spatiotemporal information from the metadata of the on-site credential media files; and performing a physical authenticity verification operation based on the comparison result of the two. If the verification passes, visual feature analysis is performed on the on-site credential media files to obtain environmental feature information, and historical environmental reference data corresponding to the claimed spatiotemporal information of the event is obtained. The event description text information, environmental feature information, and historical environmental reference data are input into a multimodal inference model to generate a logical consistency analysis result and perform corresponding processing operations accordingly. This invention conducts multimodal consistency analysis by integrating text description, on-site credential, and objective environmental reference information. It introduces a cross-information dimension joint verification mechanism on the basis of traditional spatiotemporal verification, thereby improving the accuracy of identifying false or abnormal transaction declarations, reducing the cost of manual intervention, and improving the overall reliability of the automatic review process for fintech businesses. Attached Figure Description

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for an event authenticity verification method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the event authenticity verification method of the present invention; Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the event authenticity verification device of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0013] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0014] The event authenticity verification method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can receive transaction declaration requests containing event description text information and on-site credential media files from the client, parse the event description text information to obtain the claimed spatiotemporal information, and extract the actual collection spatiotemporal information from the metadata of the on-site credential media files. Based on the comparison results of the two, a physical authenticity verification operation is performed. If the verification passes, the on-site credential media files are subjected to visual feature analysis to obtain environmental feature information, and historical environmental reference data corresponding to the claimed spatiotemporal information is obtained. The event description text information, environmental feature information, and historical environmental reference data are input into a multimodal inference model to generate logical consistency analysis results and execute corresponding processing operations accordingly. This invention conducts multimodal consistency analysis by integrating text description, on-site credential, and objective environmental reference information, introducing a cross-information dimension joint verification mechanism on the basis of traditional spatiotemporal verification, thereby improving the accuracy of identifying false or abnormal transaction declarations, reducing the cost of manual intervention, and improving the overall reliability of the automatic review process of financial technology business. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The present invention will now be described in detail through specific embodiments.

[0015] Please see Figure 2 , Figure 2This is a flowchart illustrating an embodiment of the event authenticity verification method provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0016] like Figure 2 As shown, the event authenticity verification method proposed in this invention includes the following steps: S10, Receive a pending transaction declaration request, the transaction declaration request including event description text information and on-site credential media file; In this embodiment, receiving pending transaction declaration requests is used to establish a complete and associative data carrier on the system side, enabling subsequent processing to revolve around the same transaction. The transaction declaration request consists of event description text information and on-site credential media files. These two differ in their generation mechanisms, expression forms, and information attributes, therefore they need to be uniformly incorporated into the same transaction context during the receiving phase. The event description text information exists in natural language form, carrying the user's subjective description of the event's course, environmental state, and time location. Its sources include direct user input, template filling, or speech-to-text results. The on-site credential media files exist in binary data form, originating from terminal device acquisition, and their content includes visual image information and additional attributes generated by the acquisition device.

[0017] During the reception process, the system separately accesses text information and media files, establishing a binding relationship between them through a unified transaction identifier, making the text content and media content an inseparable combination within the system. Text information is cached as character data upon reception to maintain its original semantic expression, while media files are stored as raw files upon reception, avoiding compression, re-encoding, or format conversion during the reception phase. Through this process, the transaction request forms a basic data unit within the system, containing both subjective descriptive information and objectively collected information, providing stable data input conditions for subsequent processing.

[0018] This embodiment integrates the event description text information and the on-site credential media file into the same transaction declaration request during the receiving phase, and maintains their original state and association. This enables the system to completely save subjective descriptions and objectively collected information, avoiding information dispersion or mismatch, thereby providing a consistent and traceable data foundation for subsequent analysis.

[0019] S20, parse the claimed spatiotemporal information from the event description text information, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata; In this embodiment, the claimed spatiotemporal information is parsed from the event description text information. This process transforms the user's natural language description of time and space into a spatiotemporal expression that can be processed by the system. The event description text information originates from the transaction declaration request and may contain explicit time statements, incomplete time references, or implicit location descriptions. Therefore, the text needs to be structured during parsing to ensure that time-related and location-related statements can be identified and associated separately. The parsed claimed spatiotemporal information reflects the time range and spatial location of the event in the user's subjective statement; essentially, it is an abstract expression of the spatiotemporal meaning in the text semantics.

[0020] Simultaneously, the metadata of the on-site credential media files is parsed to obtain additional information automatically generated by the device during the acquisition phase. This metadata is embedded within the media file, existing independently of the visible content, and its recording method is automatically written by the acquisition device when the file is generated. By reading the metadata, information reflecting the time and location of the acquisition activity can be extracted and transformed into unified actual acquisition spatiotemporal information. This information reflects the objective acquisition state at the time the media file was generated, and while its source differs from the user's text description, it is relevant within the same transaction.

[0021] This embodiment obtains spatiotemporal expressions from different sources from text information and media files, enabling the system to simultaneously grasp the spatiotemporal occurrence of events as subjectively described by the user and the acquisition spatiotemporal of objective media records, providing a clear and distinguishable dual spatiotemporal information basis for subsequent processing.

[0022] S30, compare the actual collected spatiotemporal information with the claimed spatiotemporal information, and perform a physical authenticity verification operation based on the comparison result and a preset spatiotemporal verification strategy; In this embodiment, the actual spatiotemporal information of the data collection is compared with the claimed spatiotemporal information of the event to determine whether the spatiotemporal information of the media file is consistent with the spatiotemporal information of the event described in the text. The actual spatiotemporal information of the data collection originates from objective records during the media file generation stage and reflects the actual time and spatial location of the data collection activity; the claimed spatiotemporal information of the event originates from the event description text and reflects the applicant's statement on the time and place of the event. There should be a reasonable correspondence between the two within the same event.

[0023] The comparison process is based on the time and space dimensions, measuring the degree of difference between the two types of spatiotemporal information at the numerical level, and forming a quantifiable comparison result. This comparison result does not directly provide a judgment of authenticity, but rather serves as an input condition for subsequent verification judgments.

[0024] Based on this, a pre-defined spatiotemporal verification strategy is introduced to define what range of spatiotemporal deviations can be considered reasonable. The spatiotemporal verification strategy is manifested as constraint rules on time and space deviations, which can originate from business rules, risk control requirements, or historical statistical characteristics. Based on the comparison results and the spatiotemporal verification strategy, a physical authenticity verification operation is performed to determine whether the transaction achieves spatiotemporal consistency at the physical level, thus establishing a pass or fail verification status.

[0025] This embodiment quantifies and compares the spatiotemporal information in the text description with the spatiotemporal information of media collection, and makes judgments in conjunction with a clear verification strategy, enabling the system to identify at the physical level whether there is an unreasonable deviation between the declared content and the objective collection behavior.

[0026] S40, if the physical authenticity verification operation fails, a first verification conclusion is generated and a first type of processing operation is performed; In this embodiment, when the physical authenticity verification operation is determined to have failed, it means that there is a deviation beyond a reasonable range between the actual collected spatiotemporal information and the claimed spatiotemporal information, and this deviation cannot be interpreted as normal reporting behavior at the physical level. Based on this judgment result, it is necessary to form a clear stage judgment mark for the current transaction so that the subsequent system can take differentiated processing paths accordingly.

[0027] The first verification conclusion expresses the judgment result of the physical authenticity verification stage. Its content comes from the failure status itself and the corresponding deviation information, and is used to identify the authenticity risk level of the transaction in the current verification stage. The first verification conclusion does not make a final decision on the overall result of the transaction, but serves as an intermediate conclusion of the physical verification stage to drive subsequent processing logic.

[0028] The first type of processing operation is used to control and adjust the transaction flow after the first verification conclusion is generated. Its purpose is to prevent transactions that do not meet the requirements of physical authenticity from continuing to enter the subsequent high-cost or high-risk processing links, and at the same time to perform necessary status marking or flow adjustment on the transaction to ensure that abnormal transactions are handled independently.

[0029] In one implementation, the first verification conclusion is generated in the form of structured state information and associated with the current transaction to identify the judgment result of the transaction in the physical authenticity verification stage. Subsequently, the first type of processing operation controls the transaction processing path to prevent the transaction from entering subsequent processing flows based on visual or logical reasoning.

[0030] In another implementation, the first type of processing operation may also include updating the transaction state, causing the transaction to enter a processing queue awaiting further verification, thereby achieving isolation from the processing of normal transactions. Under different implementations, the expression of the first verification conclusion and the specific execution mechanism of the first type of processing operation may differ, but they all revolve around the fact that the physical authenticity verification failed.

[0031] This embodiment generates a verification conclusion and executes corresponding processing operations in real time when the physical authenticity verification fails, so that transactions that do not conform to basic physical consistency are effectively intercepted in the early stage and prevented from entering the subsequent complex processing flow.

[0032] S50, if the physical authenticity verification operation passes, then the visual feature analysis of the on-site credential media file is performed to obtain environmental feature information; In this embodiment, when the physical authenticity verification operation is successful, it means that the transaction declaration request does not have any obvious physical conflict in the time and space dimensions, and the on-site credential media file meets the prerequisites for entering the visual layer analysis. At this time, the purpose of visual feature analysis of the on-site credential media file is not to identify specific object categories, but to extract information that can reflect the objective state of the shooting environment in order to characterize the external environmental conditions at the time of the event.

[0033] On-site credential media files refer to image or video data collected by terminal devices and submitted along with the transaction declaration request. Their content includes observable information about the surrounding environment at the time of capture, such as lighting, surface conditions, and natural elements. Visual feature analysis is used to extract stable and comparable environmental cues from this type of media content, enabling the environmental state to be expressed in a structured form.

[0034] Environmental feature information is used to characterize the overall objective attributes of the shooting environment. Its source depends entirely on the visual information of the media content itself, without inferring from external data. The formation of environmental feature information is based on the analysis of the image area, light and shadow distribution, and the visual representation of environmental elements, and is used to provide visual factual evidence for subsequent consistency judgments.

[0035] This embodiment introduces visual feature analysis after the physical authenticity verification is passed, so that the environmental state implicit in the on-site credential media file can be expressed in a structured way, providing an intuitive and comparable environmental basis for subsequent consistency analysis.

[0036] S60, based on the claimed spatiotemporal information, obtain the corresponding historical environmental reference data; In this embodiment, the claimed spatiotemporal information is used to express the subjective statement of the time and location of an event in the transaction declaration request. It originates from the parsing results of the event description text information and has clear temporal and spatial orientation. The purpose of obtaining historical environmental reference data based on this spatiotemporal information is to construct an objective environmental background that matches the claimed spatiotemporal conditions, which can then serve as an external reference basis for subsequent consistency judgments.

[0037] Historical environmental reference data reflects the verifiable environmental state in the objective world under the claimed time and spatial conditions. It does not rely on on-site credential media files themselves, but rather originates from historical environmental records bound to time and space. By using the spatiotemporal information of the claimed occurrence as a retrieval index, the environmental state can be limited to a specific time and location, thus preventing environmental judgments from being divorced from the event context.

[0038] The core value of historical environmental reference data lies in providing environmental facts independent of the declared content, which are used to describe the objective conditions of the environment at that time in terms of meteorological conditions, geographical attributes, etc., so that subsequent analysis can make comparative judgments based on temporal and spatial consistency.

[0039] This embodiment obtains historical environmental reference data based on the spatiotemporal information of the claimed occurrence, so that the environmental background corresponding to the event declaration has a traceable and verifiable objective basis, providing environmental comparison conditions independent of the declaration content for subsequent consistency judgment.

[0040] S70, input the event description text information, the environmental feature information, and the historical environmental reference data into the multimodal reasoning model to obtain the logical consistency analysis results; In this embodiment, the event description text information is used to express the linguistic statement of the event's process, cause, and environment in the transaction declaration request. Its essence is unstructured semantic information, containing temporal references, spatial references, and subjective descriptions of the environmental state. Environmental feature information is used to express the objective visual environmental state parsed from the on-site credential media file, reflecting observable external conditions such as lighting, ground surface, and vegetation in the image. Historical environmental reference data is used to express verifiable historical environmental records in the objective world under the claimed time and space conditions, reflecting the objective state of the real environment under the same spatiotemporal constraints.

[0041] Multimodal reasoning models are used to jointly model heterogeneous information from multiple sources within a unified reasoning space, enabling textual semantics, visual environmental features, and historical environmental records to be analyzed in relation to each other within the same logical framework. By simultaneously inputting these three types of information into the model, it can perceive subjective descriptions, objective images, and external environmental facts during the reasoning process, thereby determining whether there are consistent or conflicting relationships between different pieces of information at the temporal, spatial, and environmental cognitive levels. The results of logical consistency analysis are used to quantify the degree of logical coordination among multi-source information, reflecting the matching between event descriptions and the objective environment.

[0042] This embodiment combines textual descriptions, visual environmental features, and historical environmental records for joint reasoning, enabling the logical consistency analysis results to comprehensively reflect the inherent relationships between multiple sources of information, thereby improving the ability to judge the authenticity of event descriptions.

[0043] S80, a second verification conclusion is generated based on the logical consistency analysis result, and a second type of processing operation is performed according to the second verification conclusion.

[0044] In this embodiment, the logical consistency analysis result reflects the overall logical coordination between event description text information, environmental feature information, and historical environmental reference data. Essentially, it is a quantified or hierarchical expression of the consistency state of multi-source information. The second verification conclusion is used to explicitly express this consistency state, enabling the reasoning result to be transformed into an executable judgment output. The second type of processing operation is used to further process the transaction declaration request based on the second verification conclusion, allowing the logical analysis result to directly participate in the processing flow control.

[0045] When generating the second verification conclusion, the logical consistency analysis results are processed to map continuous or discrete analysis results to conclusion states with clear meanings, thereby distinguishing whether logical relationships hold true. This mapping process eliminates the uncertainty of the model output, making the reasoning results interpretable and executable.

[0046] When performing the second type of processing operation, the corresponding processing path is triggered according to the second verification conclusion, so that the transaction declaration request enters different processing states under different conditions of logical consistency being established or not, thereby forming a closed loop of processing based on the reasoning result.

[0047] This embodiment transforms the logical consistency analysis results into a second verification conclusion and executes the corresponding processing operation accordingly, enabling the reasoning analysis results to directly participate in transaction processing control and reducing the risk of disconnect between logical judgment and execution.

[0048] In one embodiment, step S20 above includes: S201, perform text cleaning preprocessing on the event description text information, and extract the time description entity field and the location description entity field from the preprocessed event description text information through the named entity recognition module; S202, convert the time description entity field into a standard format timestamp value, and convert the location description entity field into a standard format latitude and longitude coordinate value through geocoding; S203, Based on the timestamp value and the latitude and longitude coordinate values, generate the spatiotemporal information of the claimed occurrence; S204, Read the binary header data of the on-site credential media file, locate and separate the interchangeable image file format data from the binary header data as metadata; S205, retrieve and decode the shooting date and time tag value and the global positioning system coordinate tag value from the metadata; S206, convert the shooting date and time label value and the GPS coordinate label value into a standard spatiotemporal format, and generate actual acquisition spatiotemporal information based on the converted standard spatiotemporal format data.

[0049] In this embodiment, text cleaning preprocessing is used to eliminate noisy expressions that interfere with time and location descriptions, ensuring consistent input boundaries for subsequent extraction operations. Text cleaning preprocessing may include character normalization to unify full-width and half-width characters, uppercase and lowercase letters, and punctuation; delimiter normalization to unify date delimiters and whitespace; redundant fragment pruning to remove repetitive paraphrases or templated prompts; numerical expression normalization to unify the representation of Chinese and Arabic numerals; and semantic classification of time-related words to reduce bifurcation caused by synonyms. After cleaning, the preprocessed event description text information is sequence-labeled or fragment-located using a named entity recognition module. The named entity recognition module uses time description entity fields and location description entity fields as extraction targets. Time description entity fields carry text fragments that point to a specific time or time period of occurrence, while location description entity fields carry text fragments that point to a specific location or region of occurrence. During the extraction process, entity boundaries are constrained to avoid erroneously incorporating non-time or non-location terms into the fields. Furthermore, multiple entities are grouped to ensure that time-description entity fields and location-description entity fields within the same semantic segment can maintain their association.

[0050] After obtaining the time description entity field, it is converted into a standard format timestamp value. This conversion maps the time expression in natural language to a calculable time value. During the mapping, missing items for year, month, day, hour, minute, and second are handled using completion rules. Relative time expressions are parsed based on a reference time, and time zone and daylight saving time offsets are standardized to ensure a consistent scale for the output timestamp values ​​in subsequent comparisons. After obtaining the location description entity field, it is converted into a standard format latitude and longitude coordinate value through geocoding. Geocoding maps location names, road descriptions, landmark descriptions, or administrative division descriptions to geographic coordinates. During the mapping, it can be parsed step-by-step based on the location hierarchy to reduce discrepancies caused by locations with the same name. When the location description entity field contains range expressions, the center point coordinates or representative coordinates can be output to meet the uniform format requirements. After generating the timestamp and latitude / longitude coordinates, the claimed spatiotemporal information is generated based on these values. This generation process binds time and space together, forming a composite structure that can be directly referenced by subsequent modules. This ensures that the claimed spatiotemporal information has a consistent set of fields, a unified numerical unit, and a stable data type.

[0051] After the on-site credential media file enters the metadata parsing process, the binary header data of the on-site credential media file is read first. This binary header data is used as the location basis to identify the file container structure and the position of metadata segments. Interchangeable image file format data is located and separated from the binary header data as metadata. The separation process determines the start position and length of the metadata based on fixed markers and offset fields, and this segment is treated as an independent parsable data object, thus enabling the extraction of time and location without decoding the image pixel content. After metadata separation, the shooting date and time tag value and the GPS coordinate tag value are retrieved and decoded from the metadata. The retrieval process locates the corresponding tag entries in the tag directory structure, and the decoding process converts the raw values ​​stored in the tag entries into computable intermediate values. For date and time tags, string encoding and format differences are handled; for coordinate tags, degrees, minutes, seconds, and fractions are represented, resulting in a computable representation of the shooting date and time tag value and the GPS coordinate tag value.

[0052] After extracting the tag values, the shooting date and time tag values ​​and the GPS coordinate tag values ​​are converted into a standard spatiotemporal format. The conversion process normalizes the shooting date and time tag values ​​into timestamps or unified time strings, and the GPS coordinate tag values ​​into latitude and longitude coordinate values. Coordinate symbols and hemispherical identifiers are also standardized to ensure that the output coordinates use the same coordinate representation system as the claimed spatiotemporal information. Based on the converted standard spatiotemporal format data, the actual acquisition spatiotemporal information is generated. This generation process also binds time and space, forming a reusable composite structure. This allows the actual acquisition spatiotemporal information to be directly aligned with the claimed spatiotemporal information at the format level, thus providing calculable input for subsequent difference calculations and threshold determination.

[0053] This embodiment cleanses, extracts entities, and converts numerical values ​​from the event description text to form a standardized spatiotemporal information claiming the occurrence of the event. It then uses binary header data location, metadata separation, tag decoding, and standardization conversion of the on-site credential media file to form a standardized spatiotemporal information of the actual collection. This ensures that the spatiotemporal expression on the text side and the spatiotemporal record on the media side are consistent and aligned in terms of data type, unit scale, and field structure. This reduces comparison deviations caused by ambiguity in natural language, format differences, and inconsistencies in coordinate representation, providing a stable input basis for subsequent consistency determination.

[0054] In one embodiment, step S30 above includes: S301, the actual time dimension parameter and the actual spatial coordinate dimension parameter are separated from the actual collected spatiotemporal information, and the claimed time dimension parameter and the claimed spatial coordinate dimension parameter are separated from the claimed spatiotemporal information. S302, determine the absolute value of the time difference between the actual time dimension parameter and the claimed time dimension parameter, and determine the geographical distance value between the actual spatial coordinate dimension parameter and the claimed spatial coordinate dimension parameter; S303, combine the absolute value of the time difference with the geographical distance value to obtain the comparison result; S304, Read the maximum allowed time deviation threshold and the maximum allowed space deviation threshold from the preset spatiotemporal verification strategy; S305, determine whether the absolute value of the time difference is less than the maximum time deviation threshold, and determine whether the geographical distance value is less than the maximum spatial deviation threshold; S306, if the absolute value of the time difference is less than the maximum time deviation threshold and the geographical distance value is less than the maximum spatial deviation threshold, then the physical authenticity verification operation is determined to be in a passed state; otherwise, it is determined to be in a failed state.

[0055] In this embodiment, the actual time dimension parameter and the actual spatial coordinate dimension parameter are separated from the actual collected spatiotemporal information. This separation process decomposes the time and spatial fields in the composite structure into independent parameters. The time field maintains a numerical scale consistent with a timestamp or a standard time format that can be equivalently mapped, while the spatial field maintains a numerical scale consistent with latitude and longitude coordinates or a standard coordinate format that can be equivalently mapped. Similarly, the claimed time dimension parameter and the claimed spatial coordinate dimension parameter are separated from the claimed spatiotemporal information. This separation process maintains the same field caliber as the aforementioned separation process, ensuring that subsequent calculations can be performed within the same unit system. The output of parameter separation is used to eliminate the risk of misuse caused by the mixing of fields within the composite structure, and simultaneously establishes a clear data dependency relationship for difference calculation. The actual time dimension parameter and the claimed time dimension parameter are used to form the absolute value of the time difference, while the actual spatial coordinate dimension parameter and the claimed spatial coordinate dimension parameter are used to form the geographical distance value.

[0056] The determination of the absolute value of the time difference takes the actual time dimension parameter and the claimed time dimension parameter as input. The determination action converts both to the same time scale before performing the difference calculation, and then takes the absolute value of the difference to eliminate the influence of the order on the sign of the result, so that the absolute value of the time difference directly represents the amount of time deviation. If the time dimension parameter includes time zone information, the determination action normalizes the time zone offset before calculating the difference; if there are differences in the precision of the time dimension parameters, the determination action aligns them to the same precision level to avoid introducing systematic bias due to inconsistent rounding rules. The determination of the geographic distance value takes the actual spatial coordinate dimension parameter and the claimed spatial coordinate dimension parameter as input. The determination action unifies the coordinate values ​​to the same coordinate reference and the same angle unit before performing the distance calculation. The distance calculation can be based on spherical distance or approximate planar distance. Spherical distance is used to cover cross-regional scenes to reduce the error caused by latitude changes, while approximate planar distance is used to cover small-scale scenes to reduce computational overhead. When the coordinate value is accompanied by a height or precision label, the geographic distance value determination action can retain this label for subsequent threshold selection or outlier filtering, but the output is still the geographic distance value as the core result.

[0057] After the absolute value of the time difference and the geographical distance are formed, they are combined to obtain the comparison result. The purpose of this combination is to encapsulate the time deviation and spatial deviation into a single output, enabling subsequent verification operations to use the comparison result as input to complete a consistent threshold reading and judgment process. The comparison result can store the absolute value of the time difference and the geographical distance in a structured field format, or it can store them in a fixed-order key-value pair format. During combination, the field naming and unit caliber remain unchanged to ensure stable location of the corresponding values ​​from the comparison result later.

[0058] The preset spatiotemporal verification strategy provides threshold caliber and judgment rules. When reading the maximum allowed time deviation threshold and the maximum allowed spatial deviation threshold, the reading action first locates the threshold item in the preset spatiotemporal verification strategy, and then converts the threshold item into a unit system consistent with the absolute value of the time difference and the geographical distance value. The maximum time deviation threshold can be in seconds, minutes, or hours, and the maximum spatial deviation threshold can be in meters or kilometers. The reading action ensures that the values ​​can be directly used for comparison when outputting the thresholds. The preset spatiotemporal verification strategy can support multiple sets of thresholds and output the maximum allowed time deviation threshold and the maximum allowed spatial deviation threshold in a conditional selection manner. For example, the threshold group can be selected based on the type label of the transaction declaration request, the time period label of the claimed spatiotemporal information, or the region label. However, the result of the threshold selection still falls on the two threshold outputs and does not change the input form of the subsequent judgment.

[0059] The judgment phase compares the absolute value of the time difference with the maximum time deviation threshold and the geographical distance value with the maximum spatial deviation threshold. The comparison uses a less-than relationship to indicate that the deviation is within the allowable range; an absolute value of the time difference less than the maximum time deviation threshold indicates that the time deviation meets the constraint, and a geographical distance value less than the maximum spatial deviation threshold indicates that the spatial deviation meets the constraint. To avoid uncertainty caused by boundary value processing, the comparison action can explicitly specify whether a value equal to the threshold is allowed in a preset spatiotemporal verification strategy, and select a strict less-than or less-than-equal comparison relationship accordingly. However, when the claim uses a less-than relationship as the criterion, the comparison action converges the boundary consistency processing to the threshold setting side, for example, by slightly adjusting the threshold according to the target rule to equivalently express boundary tolerance.

[0060] The status output of the physical authenticity verification operation is triggered by the logical conjunction of two comparison results. When the absolute value of the time difference is less than the maximum time deviation threshold and the geographical distance is less than the maximum spatial deviation threshold, the physical authenticity verification operation is considered passed, and the comparison result is interpreted as satisfying the spatiotemporal consistency constraint; otherwise, it is considered failed, and the comparison result is interpreted as at least one type of deviation exceeding the constraint. This judgment output provides a single criterion for subsequent branch processing. A passed state triggers the subsequent visual feature analysis link, while a failed state triggers the first verification conclusion and the first type of processing operation link.

[0061] This embodiment separates the actual collected spatiotemporal information from the claimed spatiotemporal information, determines the absolute value of the time difference and the geographical distance, and combines the absolute value of the time difference and the geographical distance into a comparison result. Then, a preset spatiotemporal verification strategy outputs the maximum allowable time deviation threshold and the maximum allowable spatial deviation threshold, and completes the dual threshold comparison judgment. This allows time deviation and spatial deviation to be quantified and constrained under the same caliber, reducing the unit inconsistency and unclear boundary problems caused by direct comparison of composite fields. It forms a reusable pass or fail status output, providing a stable trigger basis for subsequent process branches.

[0062] In one embodiment, step S40 above includes: S401, If ​​the physical authenticity verification operation fails, the comparison results are analyzed to determine the specific inconsistency dimension that caused the failure. S402, Generate corresponding abnormal feature labels based on the specific inconsistency dimension, and define the abnormal feature labels as the first verification conclusion; S403, trigger the automatic interception policy to terminate the transaction declaration request and proceed to the subsequent visual feature parsing process; S404, mark the transaction declaration request as pending review and route the transaction declaration request to a preset review queue.

[0063] In this embodiment, when the physical authenticity verification operation fails, the process enters the conclusion attribution and handling linkage link. The goal is to extract the reason for failure from the comparison results within the same data chain, solidify the reason in a structured form as the first verification conclusion, and simultaneously isolate the transaction declaration request from the subsequent visual feature analysis process and transfer it to a preset review queue. The determination result of the physical authenticity verification operation failing the status comes from the preceding threshold comparison process. Once this status is formed, it serves as the trigger condition for this link, so that subsequent actions do not rely on manual trigger signals but on status field or judgment mark trigger signals, reducing missed interceptions caused by status drift.

[0064] When analyzing and comparing results to determine the specific inconsistency dimension causing the failure status, the comparison result serves as the input carrier, containing at least two types of deviations: the absolute value of the time difference and the geographical distance value. The analysis action breaks down the comparison result into locatable deviation items and establishes a mapping relationship with the maximum time deviation threshold and the maximum spatial deviation threshold used for judgment in the spatiotemporal verification strategy. This determines whether the failure was triggered by the time dimension, the spatial dimension, or both. The specific inconsistency dimension is a structured expression of the reason for failure. It can be expressed using enumerated types such as time dimension, spatial dimension, or time and spatial dimensions, or using a multi-value set to simultaneously retain multiple triggering sources. The generation of the specific inconsistency dimension needs to be consistent with the field caliber in the comparison result, so that the subsequent tag generation driven by the specific inconsistency dimension has a stable input source. To avoid inconsistent attribution in multiple calculations of the same transaction declaration request, the analysis action can bind the generation time mark or verification version mark of the comparison result, so that the specific inconsistency dimension is consistent with the current judgment context.

[0065] When generating corresponding anomaly feature labels based on specific inconsistency dimensions, these labels transform the attribution results into searchable, aggregable, and routable identifiers. The generation process takes the specific inconsistency dimension as primary input and can combine it with a pre-defined label mapping table, rule template, or label dictionary to output anomaly feature labels. This ensures that different specific inconsistency dimensions are mapped to different anomaly feature labels, while maintaining consistency in the anomaly feature labels for the same specific inconsistency dimension across different transaction requests. This facilitates subsequent review queue routing based on labels. The granularity of the anomaly feature labels can cover both single-dimensional and dual-dimensional deviations. For example, different anomaly feature labels can be generated for time-dimensional and spatial-dimensional triggers, and composite anomaly feature labels can be generated for both time-dimensional and spatial-dimensional triggers. The generation process can also encode segmented deviation information into the anomaly feature labels, such as subdividing labels by absolute time difference intervals or geographical distance value intervals, to support priority sorting or clustering review in the review queue. However, the label output remains anomaly feature labels.

[0066] When an anomaly feature label is defined as the first verification conclusion, it serves as the output carrier for verifying branches that failed, ensuring the conclusion's transferability and reusability. The defined action binds the anomaly feature label to the transaction declaration request for storage. This binding can be achieved by writing the anomaly feature label into the status field, extended field, or conclusion field of the transaction declaration request, or by generating an independent conclusion record and associating it with the transaction declaration request through a transaction identifier. Regardless of the method, the first verification conclusion must be referenced by subsequent first-type processing operations to explain the interception reason, drive the review route, or generate a review context. The defined action also needs to ensure the first verification conclusion is traceable throughout the lifecycle of the same transaction declaration request, for example, by recording the conclusion generation time, the conclusion source field set, or a comparison result summary, so that the review queue can quickly locate the basis for the conclusion, without changing the fundamental expression that the first verification conclusion consists of anomaly feature labels.

[0067] When an automatic interception strategy is triggered to terminate a transaction declaration request from entering the subsequent visual feature analysis process, the automatic interception strategy acts as a process gate. The triggering action uses the physical authenticity verification operation failing or the first verification conclusion generating an event as the trigger signal, intercepting the transaction declaration request at the process orchestration layer, message passing layer, or task scheduling layer, preventing it from continuing into the visual feature analysis process. Interception can be implemented by setting the processing status of the transaction declaration request to "cannot continue," marking subsequent tasks as canceled, switching the task delivery channel to a rejection channel, or adding status verification at the executor entry point. The key is that the visual feature analysis process side can recognize the interception result and stop allocating computational resources, avoiding the continued execution of image processing and inference operations for transaction declaration requests with obvious spatiotemporal inconsistencies. The automatic interception strategy can also include idempotent control, ensuring that repeated triggering of the same transaction declaration request does not repeatedly create subsequent tasks or repeatedly release resources, thereby maintaining the stability of the processing chain.

[0068] When a transaction declaration request is marked as pending review and routed to a pre-defined review queue, the pending review status serves as a status identifier for the transaction processing phase, while the pre-defined review queue acts as a container for subsequent processing. The marking action writes the transaction declaration request into the pending review status, changing its visible state within the system from "ready to proceed automatically" to "pending review," preventing it from being mistakenly considered ready for further automatic processing in other processing stages. The routing action delivers the transaction declaration request to the pre-defined review queue. The delivered content can include the transaction declaration request's identification information, the first verification conclusion, a summary of the comparison results, specific inconsistencies, and other context required for review, enabling the consumer end of the review queue to obtain complete review materials in a single retrieval.

[0069] This embodiment generates specific inconsistency dimensions by comparing and performing attribution analysis on the results. These inconsistency dimensions are then mapped to abnormal feature labels to form the first verification conclusion. Simultaneously, an automatic interception strategy is triggered to terminate the transaction declaration request and enter the visual feature parsing process. The transaction declaration request is marked as pending review and then routed to a preset review queue. This ensures that the verification failure branch forms a continuous closed loop from cause extraction to disposal, reducing the increased review costs caused by the inability to structurally accumulate the reasons for failure, reducing the resource waste caused by failure requests continuing to occupy subsequent computing links, and improving the aggregation and diversion efficiency of abnormal requests in the pending review status and the preset review queue.

[0070] In one embodiment, step S50 above includes: S501, if the physical authenticity verification operation is passed, then perform image denoising and contrast enhancement preprocessing on the on-site credential media file to obtain a preprocessed image; S502, the preprocessed image is divided into a background environment region and a foreground subject region by the semantic segmentation module; S503, Analyze the road surface texture reflection features in the background environment area to extract the surface dry / wet state and cover properties; S504, Analyze the shadow shape cast by the foreground subject area in the background environment area to extract the light intensity level and the direction of light source projection; S505, Identify the color and morphological characteristics of plants in the background environment area to extract the vegetation growth season attribute; S506, integrate the surface dryness and wetness status and cover attributes, the light intensity level and light source projection direction, and the vegetation growth season attributes to obtain environmental characteristic information.

[0071] In this embodiment, when the physical authenticity verification operation passes, visual feature analysis is performed on the on-site credential media file to obtain environmental feature information. The prerequisite for this step is that the physical authenticity verification operation has completed its state determination. Passing the state provides the executable conditions for visual feature analysis, enabling the on-site credential media file to enter the image processing pipeline, while avoiding further consumption of image processing resources in failed branches. The on-site credential media file covers various formats such as still images, burst sequences, and short video keyframes. Visual feature analysis adopts a unified frame-level processing approach for these different carriers, ensuring that the subsequently generated environmental feature information has a consistent data structure and comparability.

[0072] When performing image denoising and contrast enhancement preprocessing on the on-site credential media files to obtain the preprocessed image, image denoising is used to suppress the interference of sensor noise, compression noise, and low-light noise on texture and edges. Denoising can be achieved based on spatial domain filtering, frequency domain suppression, and edge smoothing, prioritizing the preservation of road surface texture reflection details, shadow boundary contours, and vegetation edge structures. Contrast enhancement is used to expand the brightness dynamic range, improve the visibility of dark areas, and enhance the separation of local details. Contrast enhancement can be achieved based on global grayscale stretching, local contrast enhancement, and brightness channel remapping, making the differences in reflection, shadow, and color under different exposure conditions of the same image more stable. The preprocessed image serves as the direct input to the subsequent semantic segmentation module. The preprocessed image needs to be normalized in terms of spatial resolution, color space, and pixel value range to reduce the offset introduced by differences in acquisition from different terminals.

[0073] When the preprocessed image is divided into background environment regions and foreground subject regions by the semantic segmentation module, the semantic segmentation module is responsible for pixel-level region attribution determination, outputting region masks or region boundary sets for the background environment regions and foreground subject regions. The background environment regions are used to carry environmental composition information such as roads, sky, vegetation, and buildings, while the foreground subject regions are used to carry subject information that has occlusion and projection relationships with the environment. The division between the two provides a clear spatial source for subsequent feature extraction, avoiding misclassification of subject surface materials as road surface reflections or subject paint colors as vegetation colors. The semantic segmentation module can be implemented by a model inference unit, a post-processing unit, and a connected component correction unit. The model inference unit outputs pixel category distribution, the post-processing unit performs boundary smoothing and hole filling, and the connected component correction unit removes fragmented regions according to area thresholds and shape constraints, thereby obtaining stable background environment regions and foreground subject regions.

[0074] When analyzing the reflection features of road surface textures in the background environment to extract the surface wetness / dryness state and cover attributes, the road surface texture reflection features are derived from candidate sub-regions of the road surface within the background environment. These candidate sub-regions can be obtained directly through category masks or further refined through constraints such as horizon position, perspective gradient, and connected component morphology. The extraction of road surface texture reflection features emphasizes the quantification of visual quantities such as specular distribution, specular reflection intensity, texture contrast attenuation, and local bright spot connectivity. The surface wetness / dryness state is used to express whether there is a water film or changes in reflection patterns caused by dampness on the road surface. Cover attributes are used to express whether there are layers of water, snow, mud, fallen leaves, etc., causing changes in texture and color. The extraction process can first calculate the luminance and gradient statistical features of the candidate sub-regions of the road surface, then calculate the spatial clustering and directional consistency of the candidate specular pixels, and combine this with the texture energy distribution to determine changes in surface roughness, thereby outputting the surface wetness / dryness state and cover attributes. To improve cross-device consistency, the extraction process can calculate and normalize features based on the luminance and chrominance channels of the preprocessed image, ensuring that the surface wetness / dryness state and cover attributes remain comparable under different exposure conditions.

[0075] When analyzing the shadow patterns cast by the foreground subject area onto the background environment to extract the illumination intensity level and light source projection direction, the shadow pattern originates from the occlusion projection relationship between the foreground subject area and the background environment area. The shadow area can be determined through brightness abrupt changes, chromaticity shifts, and boundary geometric consistency. The illumination intensity level is used to express the difference in shadow sharpness and contrast caused by the difference in the proportion of direct light and diffused light, while the light source projection direction is used to express the incident light direction corresponding to the shadow extension direction. During the extraction process, possible shadow boundary bands can be searched near the boundary of the foreground subject area first, and the brightness ratio and gradient magnitude on both sides of the boundary can be calculated to quantify the shadow intensity. Then, the light source projection direction can be estimated based on the shadow principal axis direction and the geometric center position of the foreground subject area. Simultaneously, the illumination intensity level can be obtained using the sharpness of the shadow boundary and the brightness variance inside the shadow. To reduce the interference of road surface material on shadow determination, shadow extraction can be linked with the surface wetness and dryness state and the properties of the covering material. In the presence of strong specular reflection, the probability of false triggering of shadow boundaries by highlights can be reduced, making the illumination intensity level and light source projection direction more stable.

[0076] When identifying plant color and morphological features in a background environment to extract vegetation growth season attributes, these features are derived from candidate vegetation sub-regions within the background environment. These candidate sub-regions can be obtained through category masks or further filtered using color distribution and texture directionality constraints. The extraction of plant color and morphological features focuses on visual indicators such as leaf color distribution, the proportion of withered and yellowed areas, the degree of branch exposure, and canopy density. Vegetation growth season attributes are used to express whether the vegetation is in a vigorous growth phase, undergoing color decline, or dormant and sparse stage. The extraction process can begin by calculating the statistical distribution of hue and saturation in the candidate vegetation sub-regions to obtain the proportion of green and yellowish-brown areas. Then, morphological analysis of the edge structure, including elongated branches and reticulated textures, is performed to quantify leaf density and branch significance, thereby outputting the vegetation growth season attributes. To reduce the color cast caused by white balance differences, the extraction process can perform color constancy correction on the preprocessed image or use relative chromaticity features, making the vegetation growth season attributes more dependent on structural information and relative color relationships.

[0077] When integrating surface wetness / dryness, cover attributes, light intensity level and light source direction, and vegetation growth season attributes to obtain environmental feature information, the integration process encapsulates these three complementary but different environmental attributes into a unified environmental feature information. This ensures that the environmental feature information includes not only surface condition clues but also light and seasonal clues, facilitating subsequent reasoning and consistency judgments using a single input object. The integration can be implemented using a field-based structure, where surface wetness / dryness, cover attributes, light intensity level and light source direction, and vegetation growth season attributes are written into fixed fields, along with field confidence levels, source region identifiers, and preprocessed image version identifiers. Alternatively, a vectorized representation can be used, mapping the three types of attributes to feature vectors of a unified dimension and retaining a field index mapping table to support rapid calculation and interpretable backtracking. The generation of environmental feature information must ensure a one-to-one correspondence with on-site credential media files and be bound to the processing context of the transaction declaration request to avoid feature crosstalk between different transactions.

[0078] This embodiment obtains a preprocessed image by performing image denoising and contrast enhancement preprocessing on the on-site credential media file. Then, it obtains the background environment region and the foreground subject region based on the semantic segmentation module. Next, it extracts the surface dryness and wetness status, cover attributes and vegetation growth season attributes from the background environment region, and extracts the light intensity level and light source projection direction from the projection relationship between the foreground subject region and the background environment region. Finally, it integrates these to form environmental feature information, transforming the environmental information from unstructured media content into a structured and comparable representation, reducing the impact of differences in collection quality on environmental judgment, and improving the usability and stability of environmental clues in subsequent consistency analysis.

[0079] In one embodiment, step S70 above includes: S701, Input the event description text information into the text encoding module to obtain the text semantic feature vector; S702, The environmental feature information is input into the visual encoding module to obtain the visual semantic feature vector; S703, input the historical environmental reference data into the environmental data encoding module to obtain the environmental data feature vector; S704, input the text semantic feature vector, the visual semantic feature vector and the environmental data feature vector into the multimodal fusion module of the multimodal inference model to obtain the fused feature vector; S705, the fused feature vector is input into the cross-modal attention module of the multimodal inference model for logical association derivation, and the logical association derivation result is obtained; S706, Based on the logical association derivation result, a confidence score is generated by the score generation module as the logical consistency analysis result.

[0080] In this embodiment, event description text information, environmental feature information, and historical environmental reference data are input into a multimodal inference model to obtain logical consistency analysis results. The goal is to convert textual semantic cues, visual environmental cues, and external reference cues into alignable representations within the same inference space, and output quantifiable confidence scores through fusion and correlation derivation within the model. The event description text information comes from the natural language description in the transaction declaration request, and its information form may include unstructured content such as time expression, location expression, environmental description, and collision cause description; the environmental feature information comes from the visual feature analysis results of the on-site credential media documents, and its information form includes structured fields such as surface wetness and dryness status and cover attributes, light intensity level and light source projection direction, and vegetation growth season attributes; the historical environmental reference data comes from external reference records that match the claimed spatiotemporal information, and its information form includes structured fields such as precipitation type records, temperature value records, visibility value records, road attribute category data, and surrounding land feature distribution data. As a unified reasoning subject, the multimodal reasoning model needs to receive three types of heterogeneous inputs and output unified logical consistency analysis results. Therefore, a text encoding module, a visual encoding module, and an environmental data encoding module are introduced at the input end to form three-way feature representations. A multimodal fusion module is introduced in the middle of the model to form a fused feature vector. A cross-modal attention module is introduced at the end of the model to form a logical association derivation result. A score generation module is introduced at the output end to form a confidence score and use it as the logical consistency analysis result.

[0081] When the event description text is input into the text encoding module to obtain the text semantic feature vector, the text encoding module performs functions such as text cleaning, word segmentation or sub-word splitting, serialization representation, semantic context modeling, and vectorized output. Text cleaning is used to unify the character set, eliminate redundant whitespace and noisy symbols, and standardize the writing differences between time and location expressions, avoiding representational biases caused by synonyms in the text. Serialization representation maps the event description text information into an ordered sequence of tags, and then maps the tag sequence into an embedding representation. The embedding representation can contain word vectors, position vectors, and sentence vectors, thus preserving word order relationships and paragraph structure. Semantic context modeling is used to establish long-range dependencies within the tag sequence, enabling descriptions such as "slippery roads in rainy weather," "low visibility at night," and "highway sections" to influence the vector output in a context-consistent manner. As a unified semantic representation of the text, the text semantic feature vector needs to have fixed dimensions and a stable numerical range so that it can be integrated with the visual semantic feature vector and the environmental data feature vector into the same fusion interface. Therefore, the text encoding module usually sets a pooling or convergence mechanism at the output end to compress the sequence-level representation into a single vector. At the same time, it can be attached with text-side confidence or key label attention weights to facilitate the location of contribution sources in subsequent association derivation.

[0082] When environmental feature information is input into the visual encoding module to obtain a visual semantic feature vector, the visual encoding module undertakes the re-encoding and semantic compression of structured visual fields, mapping the environmental feature information from a set of fields into a continuous vector that can be aligned with the text. Environmental feature information may contain both discrete categorical fields and continuous numerical fields. For example, the dryness / wetness of the ground surface and the attributes of the cover may be expressed in an enumeration or multi-label manner; the light intensity level and the direction of light source projection may include level values ​​and directional angles; and the seasonal attributes of vegetation growth may be expressed in stage categories or probability distributions. The visual encoding module can perform embedding mapping on discrete fields, normalization and interval mapping on continuous fields, and set weights or gating for different fields to distinguish between fields that are more sensitive to logical consistency and fields with higher noise levels. Subsequently, multiple fields are embedded and integrated into a visual semantic feature vector through field-level aggregation. The aggregation process can employ weighted summation, attention aggregation, or gating fusion, ensuring that the visual semantic feature vector retains the combination relationship between ground surface state, light state, and vegetation state, while maintaining scale characteristics similar to the text semantic feature vector. Visual semantic feature vectors play the role of evidence channel for "visible environment" in subsequent fusion, and their encoding quality directly affects the sensitivity of cross-modal consistency inference to environmental conflicts.

[0083] When historical environmental reference data is input into the environmental data encoding module to obtain environmental data feature vectors, the environmental data encoding module undertakes the spatiotemporal consistency, field standardization, and evidence compression of external reference fields, enabling historical environmental reference data to participate in fusion inference with unified semantic coordinates. Historical environmental reference data typically includes both dynamic and static references. Dynamic references include records that change over time, such as precipitation type records, temperature records, and visibility records. Static references include records that change more slowly over time, such as road attribute category data and surrounding feature distribution data. The environmental data encoding module needs to establish a time alignment standard for dynamic reference fields, for example, mapping the record time window to the time point or time period statistics corresponding to the claimed spatiotemporal information, and converting the statistics into fixed-structure fields; and establish a spatial alignment standard for static reference fields, for example, mapping road attribute category data and surrounding feature distribution data to spatial neighborhood features centered on the claimed geographic coordinates, and retaining spatial scale parameters in the output. Field standardization is used to unify units, dimensions, and value ranges, such as unit normalization of temperature records, interval mapping of visibility records, and enumeration mapping of road attribute category data. The environmental data feature vector is obtained by aggregating the above standardized fields. It can carry the confidence level of the reference source or the missing marker, which can be used by the subsequent cross-modal attention module to adjust the dependence on the reference channel when there is evidence conflict.

[0084] When textual semantic feature vectors, visual semantic feature vectors, and environmental data feature vectors are input into the multimodal fusion module of a multimodal inference model to obtain a fused feature vector, the multimodal fusion module undertakes cross-channel alignment and unified representation construction. Before fusion, dimensional alignment and scale alignment need to be performed on the three types of vectors. Dimensional alignment maps the outputs of different encoding modules to the same dimension through linear mapping or projection layers. Scale alignment normalizes the numerical distribution of different channels to a comparable range to avoid excessive amplitude of one vector causing fusion bias. The fusion process needs to explicitly preserve channel identity so that the model can distinguish between textual evidence, on-site visual evidence, and external reference evidence. This can be achieved through channel embedding or gated vectors. The generation of fused feature vectors can be achieved through methods such as concatenated projection, weighted summation, and gated fusion. Among these, gated fusion can dynamically adjust channel contributions based on the input confidence and missing markers, thereby suppressing the influence of unstable channels on inference when historical environmental reference data is incomplete or environmental feature information is noisy. The fused feature vector, as a unified input for cross-modal inference, should simultaneously contain complementary relationships and potential contradictory clues among the three types of information, providing a decomposable evidentiary basis for subsequent logical association derivation.

[0085] When the fused feature vectors are input into the cross-modal attention module of the multimodal reasoning model for logical association derivation and to obtain the logical association derivation results, the cross-modal attention module undertakes evidence alignment, relationship modeling, and contradiction localization. Evidence alignment establishes correspondences between environmental terms in text descriptions and environmental fields in environmental feature information and reference fields in historical environmental reference data. For example, it aligns rainfall descriptions in text with precipitation type records, visibility descriptions in text with visibility numerical records, and road type descriptions in text with road attribute category data. Simultaneously, it establishes consistency constraints between surface wetness / dryness and cover attributes in environmental feature information and precipitation type and temperature numerical records. Relationship modeling constructs supportive and conflicting relationships among the aligned evidence. Supportive relationships enhance consistency tendencies, while conflicting relationships enhance contradiction tendencies. Contradiction localization focuses conflicts on specific field pairs or semantic fragments through attention weights or saliency distribution, thereby forming outputtable logical association derivation results. The logical association derivation results can include structured content such as cross-channel matching scores, conflict field sets, and conflict intensity statistics. They can also include intermediate vector representations used for calculation by the score generation module. The key is that they can be stably mapped to confidence scores by the score generation module and support backtracking of conflict sources when needed.

[0086] When a confidence score is generated from the logical association derivation results and used as the result of logical consistency analysis, the score generation module undertakes the mapping from the derivation state to a quantifiable output. The confidence score needs to express the degree of logical consistency. Interval mapping can be used to map consistency strength to a high score and inconsistency strength to a low score. Alternatively, multi-segment mapping can be used to set higher resolutions in the high-confidence and low-confidence intervals to enhance discriminative power. The input to the score generation module can be the consistency score and inconsistency score from the logical association derivation results, or their vector representations. The output confidence score usually requires calibration to ensure comparability of scores from different batches and data distributions. The logical consistency analysis results are output in the form of confidence scores, enabling subsequent processing links to perform threshold comparisons and triage control based on unified values, while maintaining a comprehensive reflection of textual, visual, and reference data evidence.

[0087] This embodiment converts event description text information into text semantic feature vectors through a text encoding module, environmental feature information into visual semantic feature vectors through a visual encoding module, and historical environmental reference data into environmental data feature vectors through an environmental data encoding module. After dimensional alignment and channel contribution adjustment are completed by the multimodal fusion module, a fused feature vector is formed. Then, the cross-modal attention module performs alignment and relationship modeling on the three sources of evidence to obtain the logical association inference result. Finally, the score generation module outputs the confidence score as the logical consistency analysis result. This allows text descriptions, on-site environmental clues, and external reference records to form a quantifiable consistency measure within the same reasoning framework, reducing the probability of misjudgment caused by a single evidence channel and improving the detectability and comparability of cross-information source contradictions.

[0088] In one embodiment, step S80 above includes: S801, Obtain the confidence score from the logical consistency analysis results; S802, compare the confidence score with a preset confidence threshold to obtain a confidence comparison result; S803, Based on the confidence comparison results, generate a second verification conclusion; S804, if the second verification conclusion is a logically consistent conclusion, then an automatic approval process for the transaction declaration request is triggered; S805, if the second verification conclusion is a logical contradiction, then the transaction declaration request is marked as a logical anomaly request, the logical anomaly request is automatically routed to a preset review queue, and a logical doubt report is generated based on the logical contradiction conclusion.

[0089] In this embodiment, the logical consistency analysis result includes a confidence score, which is used to express the degree of self-consistency between the event description text information, environmental feature information, and historical environmental reference data at the reasoning level. The second verification conclusion is used to solidify this quantitative expression into a routable, recordable, and verifiable discrete conclusion, so that subsequent processing does not depend on explanatory text but on deterministic branch conditions.

[0090] When retrieving the confidence score from the logical consistency analysis results, the system locates the confidence score field in the data structure of the logical consistency analysis results and reads it. This reading action must be consistent with the output format of the logical consistency analysis results to avoid field offsets or missing fields due to different output versions. After reading, the confidence score undergoes value validation, including numerical type validation and numerical range validation. Numerical type validation ensures that comparison operations can be performed directly, while numerical range validation ensures that the confidence score falls within the expected scale. If the logical consistency analysis results carry multiple inferences or multiple outputs, the reading action can be bound to the inference identifier of the same transaction declaration request to determine the unique source of the confidence score and avoid inconsistencies in subsequent conclusion generation.

[0091] When comparing the confidence score with a preset confidence threshold to obtain the confidence comparison result, the system reads the confidence threshold and performs a comparison operation with the confidence score. The comparison operation needs to clearly define the comparison relationship. For example, "greater than or equal to" corresponds to inclusion by the boundary, and "strictly greater than" corresponds to exclusion by the boundary. Once the comparison relationship is determined, it should be consistent in the conclusion generation. The confidence threshold, as a preset condition, usually exists in the form of a configuration item and is bound to the version information. The version information is used to trace the previous judgment basis after the threshold is adjusted, preventing the same confidence score from being interpreted as opposite conclusions under different threshold versions. The confidence comparison result, as the output of the comparison operation, can be expressed in Boolean form or enumeration form. Boolean form is used for fast branching, and enumeration form is used to expand boundary states such as missing and abnormal states, so that the conclusion generation covers more operating scenarios without introducing uncertain default values.

[0092] When generating the second verification conclusion based on the confidence comparison result, the system establishes a definite mapping from the confidence comparison result to the second verification conclusion. The mapping rule maps successful conclusions to logically consistent conclusions and unsuccessful conclusions to logically contradictory conclusions. The generation of the second verification conclusion not only produces the conclusion value but also simultaneously generates a conclusion record carrier. This record carrier can include a transaction identifier, confidence score, confidence threshold, comparison relationship, threshold version, and generation timestamp, used to solidify the conclusion from the runtime state into an auditable persistent structure. Since the second verification conclusion is subsequently used as a branch condition, the conclusion value should remain a closed set and use consistent naming throughout the entire chain to avoid confusion in branch judgments caused by near-synonyms such as "logically consistent" or "logically passed."

[0093] When the second verification conclusion is logically consistent, an automatic approval process for the transaction declaration request is triggered. This automatic approval process corresponds to the pass branch in the second type of processing operation. The triggering action can be manifested as submitting an approval task to the entry queue of the automatic approval process or calling the approval processing interface. The payload of the approval task must at least include the transaction identifier and the second verification conclusion, enabling downstream processing to complete the approval action without recalculating the confidence score. To avoid duplicate triggering and resulting duplicate approvals, the triggering action can introduce idempotency control. An idempotency key is constructed using the transaction identifier and approval type; if the idempotency key already exists, duplicate writing is rejected. During the execution of the approval process, the automatic approval process can simultaneously write approval trajectory records, establishing a one-to-one association between the approval action and the second verification conclusion at the data level, facilitating subsequent verification of the approval basis.

[0094] When the second verification conclusion is a logically contradictory conclusion, the marking, routing, and report generation processes are executed. These three actions together constitute the non-passing branch in the second type of processing operation. The action of marking a transaction declaration request as a logically abnormal request is used to transition the transaction declaration request's status from a normal pending state to a logically abnormal request state. An exception reason field can be written to bind the logically contradictory conclusion. This state transition allows any subsequent processing node to identify its handling path when reading the transaction declaration request, without needing to read the logical consistency analysis results again. The action of automatically routing logically abnormal requests to a preset review queue is used to deliver logically abnormal requests to the queue entity that carries review tasks. The preset review queue needs to have a queue identifier and consumption permission control. The queue identifier ensures the uniqueness of the routing target, and the consumption permission control ensures that only review roles or review services are allowed to read this type of task. The routing action can also introduce idempotency control to prevent logically abnormal requests from repeatedly entering the preset review queue in retry scenarios. The action of generating a logical inconsistency report based on logical contradiction conclusions is used to structurally encapsulate key information for contradiction determination into a readable, searchable, and archiveable report. The logical inconsistency report can include a transaction identifier, conclusion value, confidence score, confidence threshold, comparison relationship, threshold version, and generation timestamp. It can further include a set of contradiction-related fields associated with the logical consistency analysis results, enabling the review process to quickly locate the points of evidence conflict that need to be verified without re-executing the reasoning. The storage of logical inconsistency reports can be bound to the task payload of a preset review queue, or written to a separate report storage area and carried as a report index in the review queue task, thus achieving consistency between report reachability and task traceability during queue consumption.

[0095] This embodiment extracts a confidence score from the logical consistency analysis results and compares it with a confidence threshold to form a confidence comparison result. Then, a second verification conclusion is generated from the confidence comparison result, which drives the automatic approval process or the marking and routing of logical anomaly requests. At the same time, a logical doubt report is generated in the logical contradiction conclusion branch, so that the transaction declaration request has a clear handling entry and traceable conclusion basis in both logically consistent and logically contradictory situations. This reduces the risk of a break in the chain between conclusion generation and subsequent handling, and reduces the reliance on repeated reasoning and repeated manual search during review and location.

[0096] In one embodiment, an event authenticity verification device is provided, which corresponds one-to-one with the event authenticity verification methods described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the event authenticity verification device of the present invention. The modules include: transaction access module 10, spatiotemporal information parsing module 20, spatiotemporal consistency verification module 30, first-type handling decision module 40, environmental visual parsing module 50, historical environmental data acquisition module 60, multimodal logic reasoning module 70, and second-type handling decision module 80. Detailed descriptions of each functional module are as follows: Transaction access module 10 is used to receive pending transaction declaration requests, wherein the transaction declaration request includes event description text information and on-site certificate media file; The spatiotemporal information parsing module 20 is used to parse the claimed spatiotemporal information from the event description text information, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata; The spatiotemporal consistency verification module 30 is used to compare the actual collected spatiotemporal information with the claimed spatiotemporal information, and to perform physical authenticity verification operations based on the comparison results and the preset spatiotemporal verification strategy. The first type of decision-making module 40 is used to generate a first verification conclusion and execute a first type of processing operation if the physical authenticity verification operation fails. The environmental visual analysis module 50 is used to perform visual feature analysis on the on-site credential media file to obtain environmental feature information if the physical authenticity verification operation passes. The historical environment data acquisition module 60 is used to acquire corresponding historical environment reference data based on the claimed spatiotemporal information. The multimodal logic reasoning module 70 is used to input the event description text information, the environmental feature information, and the historical environmental reference data into the multimodal reasoning model to obtain the logic consistency analysis results; The second type of decision-making module 80 is used to generate a second verification conclusion based on the logical consistency analysis results, and to perform a second type of processing operation according to the second verification conclusion.

[0097] For specific limitations regarding the event authenticity verification device, please refer to the aforementioned limitations on the event authenticity verification method, which will not be repeated here. Each module in the aforementioned event authenticity verification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0098] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When executed by the processor, the computer program implements the functions or steps of an event authenticity verification method on the server side.

[0099] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the client-side functions or steps of an event authenticity verification method.

[0100] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Receive pending transaction reporting requests, wherein the transaction reporting requests include event description text information and on-site credential media files; Parse the claimed spatiotemporal information of the event description text, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata; The actual spatiotemporal information collected is compared with the claimed spatiotemporal information, and a physical authenticity verification operation is performed based on the comparison result and a preset spatiotemporal verification strategy. If the physical authenticity verification operation fails, a first verification conclusion is generated and a first type of processing operation is performed; If the physical authenticity verification operation passes, then the visual feature analysis of the on-site credential media file is performed to obtain environmental feature information; Based on the claimed spatiotemporal information, obtain the corresponding historical environmental reference data; The event description text information, the environmental feature information, and the historical environmental reference data are input into the multimodal reasoning model to obtain the logical consistency analysis results. A second verification conclusion is generated based on the logical consistency analysis results, and a second type of processing operation is performed based on the second verification conclusion.

[0101] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, performs the following steps: Receive pending transaction reporting requests, wherein the transaction reporting requests include event description text information and on-site credential media files; Parse the claimed spatiotemporal information of the event description text, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata; The actual spatiotemporal information collected is compared with the claimed spatiotemporal information, and a physical authenticity verification operation is performed based on the comparison result and a preset spatiotemporal verification strategy. If the physical authenticity verification operation fails, a first verification conclusion is generated and a first type of processing operation is performed; If the physical authenticity verification operation passes, then the visual feature analysis of the on-site credential media file is performed to obtain environmental feature information; Based on the claimed spatiotemporal information, obtain the corresponding historical environmental reference data; The event description text information, the environmental feature information, and the historical environmental reference data are input into the multimodal reasoning model to obtain the logical consistency analysis results. A second verification conclusion is generated based on the logical consistency analysis results, and a second type of processing operation is performed based on the second verification conclusion.

[0102] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0104] It should be noted that if any AI models, software tools, or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these 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, and should all be included within the protection scope of the present invention.

[0105] The user personal information involved in this application embodiment is all authorized (knowing and consenting) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various open, legal and compliant means. The collection, storage, use, processing, transmission, provision and disclosure of the information, data and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals.

Claims

1. A method for verifying the authenticity of an event, characterized in that, Includes the following steps: Receive pending transaction reporting requests, wherein the transaction reporting requests include event description text information and on-site credential media files; Parse the claimed spatiotemporal information of the event description text, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata; The actual spatiotemporal information collected is compared with the claimed spatiotemporal information, and a physical authenticity verification operation is performed based on the comparison result and a preset spatiotemporal verification strategy. If the physical authenticity verification operation fails, a first verification conclusion is generated and a first type of processing operation is performed; If the physical authenticity verification operation passes, then the visual feature analysis of the on-site credential media file is performed to obtain environmental feature information; Based on the claimed spatiotemporal information, obtain the corresponding historical environmental reference data; The event description text information, the environmental feature information, and the historical environmental reference data are input into the multimodal reasoning model to obtain the logical consistency analysis results. A second verification conclusion is generated based on the logical consistency analysis results, and a second type of processing operation is performed based on the second verification conclusion.

2. The event authenticity verification method as described in claim 1, characterized in that, Parse the claimed spatiotemporal information from the event description text, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata, including: The event description text information is subjected to text cleaning preprocessing, and the time description entity field and the location description entity field are extracted from the preprocessed event description text information by the named entity recognition module. The time description entity field is converted into a standard format timestamp value, and the location description entity field is converted into a standard format latitude and longitude coordinate value through geocoding; Based on the timestamp value and the latitude and longitude coordinate values, the spatiotemporal information of the claimed occurrence is generated; Read the binary header data of the on-site credential media file, locate and separate the interchangeable image file format data from the binary header data as metadata; The shooting date and time tag value and the GPS coordinate tag value are retrieved and decoded from the metadata; The shooting date and time tag value and the GPS coordinate tag value are converted into a standard spatiotemporal format. Based on the converted standard spatiotemporal format data, the actual acquisition spatiotemporal information is generated.

3. The event authenticity verification method as described in claim 1, characterized in that, The actual spatiotemporal information collected is compared with the claimed spatiotemporal information, and a physical authenticity verification operation is performed based on the comparison result and a preset spatiotemporal verification strategy, including: The actual time dimension parameter and the actual spatial coordinate dimension parameter are separated from the actual collected spatiotemporal information, and the claimed time dimension parameter and the claimed spatial coordinate dimension parameter are separated from the claimed spatiotemporal information. Determine the absolute value of the time difference between the actual time dimension parameter and the claimed time dimension parameter, and determine the geographical distance between the actual spatial coordinate dimension parameter and the claimed spatial coordinate dimension parameter; The comparison result is obtained by combining the absolute value of the time difference with the geographical distance value. Read the maximum allowed time deviation threshold and the maximum allowed space deviation threshold from the preset spatiotemporal verification strategy; Determine whether the absolute value of the time difference is less than the maximum time deviation threshold, and determine whether the geographical distance value is less than the maximum spatial deviation threshold; If the absolute value of the time difference is less than the maximum time deviation threshold and the geographical distance value is less than the maximum spatial deviation threshold, the physical authenticity verification operation is determined to be in a passed state; otherwise, it is determined to be in a failed state.

4. The event authenticity verification method as described in claim 1, characterized in that, If the physical authenticity verification operation fails, a first verification conclusion is generated and a first type of processing operation is performed, including: If the physical authenticity verification operation fails, the comparison results are analyzed to determine the specific inconsistency dimension that caused the failure. Based on the specific inconsistency dimension, a corresponding abnormal feature label is generated, and the abnormal feature label is defined as the first verification conclusion. The automatic interception policy is triggered to terminate the transaction declaration request and proceed to the subsequent visual feature parsing process; The transaction declaration request is marked as pending review, and the transaction declaration request is routed to a preset review queue.

5. The event authenticity verification method as described in claim 1, characterized in that, If the physical authenticity verification operation passes, then visual feature analysis is performed on the on-site credential media file to obtain environmental feature information, including: If the physical authenticity verification operation is passed, then image denoising and contrast enhancement preprocessing is performed on the on-site credential media file to obtain a preprocessed image; The preprocessed image is divided into a background environment region and a foreground subject region using a semantic segmentation module. Analyze the reflection features of the road surface texture in the background environment area to extract the surface dryness and wetness status and the properties of the covering material; The shadow shape cast by the foreground subject area on the background environment area is analyzed to extract the light intensity level and the direction of light source projection; Identify the color and morphological characteristics of plants in the background environment area to extract the seasonal attributes of vegetation growth; By integrating the surface dryness and wetness status and cover attributes, the light intensity level and light source projection direction, and the vegetation growth season attributes, environmental characteristic information is obtained.

6. The event authenticity verification method as described in claim 1, characterized in that, The event description text, environmental feature information, and historical environmental reference data are input into a multimodal reasoning model to obtain logical consistency analysis results, including: The event description text information is input into the text encoding module to obtain the text semantic feature vector; The environmental feature information is input into the visual encoding module to obtain a visual semantic feature vector; The historical environmental reference data is input into the environmental data encoding module to obtain the environmental data feature vector; The text semantic feature vector, the visual semantic feature vector, and the environmental data feature vector are input into the multimodal fusion module of the multimodal inference model to obtain the fused feature vector; The fused feature vector is input into the cross-modal attention module of the multimodal inference model for logical association derivation, and the logical association derivation result is obtained. Based on the logical association derivation results, a confidence score is generated by the score generation module as the result of the logical consistency analysis.

7. The event authenticity verification method as described in claim 1, characterized in that, A second verification conclusion is generated based on the logical consistency analysis results, and a second type of processing operation is performed according to the second verification conclusion, including: Obtain the confidence score from the logical consistency analysis results; The confidence score is compared with a preset confidence threshold to obtain a confidence comparison result; Based on the confidence level comparison results, a second verification conclusion is generated; If the second verification conclusion is a logically consistent conclusion, then an automatic approval process for the transaction declaration request is triggered; If the second verification conclusion is a logical contradiction, the transaction declaration request is marked as a logical anomaly request, the logical anomaly request is automatically routed to a preset review queue, and a logical doubt report is generated based on the logical contradiction conclusion.

8. An event authenticity verification device, characterized in that, The event authenticity verification device includes: The transaction access module is used to receive pending transaction declaration requests, which include event description text information and on-site credential media files. The spatiotemporal information parsing module is used to parse the claimed spatiotemporal information from the event description text information, parse the metadata of the on-site credential media file, and extract the actual collection spatiotemporal information from the metadata; The spatiotemporal consistency verification module is used to compare the actual collected spatiotemporal information with the claimed spatiotemporal information, and to perform physical authenticity verification operations based on the comparison results and a preset spatiotemporal verification strategy. The first type of decision-making module is used to generate a first verification conclusion and execute a first type of processing operation if the physical authenticity verification operation fails. The environmental visual analysis module is used to perform visual feature analysis on the on-site credential media file to obtain environmental feature information if the physical authenticity verification operation passes. The historical environment data acquisition module is used to acquire corresponding historical environment reference data based on the claimed spatiotemporal information. The multimodal logical reasoning module is used to input the event description text information, the environmental feature information, and the historical environmental reference data into the multimodal reasoning model to obtain the logical consistency analysis results. The second type of decision-making module is used to generate a second verification conclusion based on the logical consistency analysis results, and to perform a second type of processing operation according to the second verification conclusion.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and an event authenticity verification program stored in the memory and executable on the processor, wherein the event authenticity verification program, when executed by the processor, implements the steps of the event authenticity verification method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores an event authenticity verification program, which, when executed by a processor, implements the steps of the event authenticity verification method as described in any one of claims 1-7.

Images