Chain storage-based intelligent tamper-proofing method and system for medical data

Abstract

The application provides a chain storage-based intelligent tamper-proofing method and system for medical data, relates to the technical field of medical data security, and comprises the following steps: adaptively segmenting medical data according to semantic boundaries, extracting feature abstracts, establishing bidirectional chain identifiers, and embedding a storage structure, so that forward and backward tracing and checking are realized; when it is detected that the integrity checking is inconsistent, the tamper-proofing path and range are located based on the bidirectional chain identifiers, isolation marking and semantic boundary identification are adjusted, and a redundant chain association path is added, so that efficient tamper-proofing protection of medical data is realized.

Description

Technical Field

[0001] This invention relates to the field of medical data security technology, and in particular to an intelligent anti-tampering method and system for medical data based on chained storage. Background Technology

[0002] With the deepening application of information technology in the medical field, the secure storage and integrity protection of medical data have become particularly important. As a crucial component of patient privacy, the authenticity and integrity of medical data directly affect the reliability of medical diagnoses, treatment plans, and doctor-patient relationships. Traditional medical data storage often employs centralized database management, relying on access control and digital signatures to ensure data security. In recent years, distributed storage technologies such as blockchain have been introduced into the field of medical data management, enhancing the tamper-proof capabilities of medical data through decentralized consensus mechanisms and cryptographic techniques. Simultaneously, the development of medical big data and artificial intelligence technologies has also provided new ideas and methods for the intelligent management of medical data.

[0003] Existing technologies often employ a one-time summary calculation of the entire medical data for tamper-proofing. This approach struggles to pinpoint the specific areas of tampering, requiring the entire dataset to be marked as untrustworthy once tampered, thus failing to provide fine-grained protection for medical data. Traditional unidirectional linked storage structures only offer one-way traceability. When data undergoes complex tampering, it's difficult to construct a complete tampering timeline and accurately determine the scope of impact, reducing the effectiveness and reliability of the tamper-proofing system. Furthermore, existing technologies lack full utilization of the semantic features of medical data, failing to adaptively segment data based on semantic boundaries. This results in a mismatch between the tamper-proofing mechanism and the actual semantic structure of medical data, making it difficult to provide effective protection against precise tampering targeting specific medical semantic units. Summary of the Invention

[0004] This invention provides an intelligent anti-tampering method and system for medical data based on chained storage, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an intelligent anti-tampering method for medical data based on chained storage, comprising:

[0006] The medical data objects to be stored are adaptively segmented according to the data semantic boundaries to obtain multiple data segments, and feature summaries are extracted from each segment.

[0007] The feature summary of the current data segment is bidirectionally associated with the feature summaries of the preceding and subsequent data segments to generate a bidirectional chain identifier containing the dependencies between the preceding and subsequent segments, and the bidirectional chain identifier is embedded in the storage structure of the current data segment.

[0008] The feature summaries of each data segment are re-extracted and compared bidirectionally with the bidirectional chain identifier. Based on the bidirectional chain identifier, the medical data object is subjected to forward and backward traceability verification to generate an integrity verification result containing directional difference information.

[0009] When the integrity verification results are inconsistent, the temporal path and scope of the tampering operation are reconstructed based on the spatial distribution relationship between the forward and backward break positions of the bidirectional chain identifier, and the data segments within the scope of the impact are isolated and marked.

[0010] The tampered region is identified based on the isolation marker and the time sequence path, and the tampered region is used as feedback information to adjust the semantic boundary recognition, reduce the segmentation granularity of the tampered region, and add redundant chain association paths to the tampered region to realize intelligent anti-tampering of the medical data object.

[0011] The medical data objects to be stored are adaptively segmented according to data semantic boundaries to obtain multiple data segments, and feature summaries are extracted from each segment, including:

[0012] The format of the medical data object is identified to determine its data organization structure. Based on the data organization structure, the fields of the medical data object are parsed to obtain the field identifier and field content of each data field.

[0013] The data field type of each data field is determined based on the field identifier. By analyzing the reference and nesting relationships in the field content, the association relationship between the data fields is obtained, and the candidate position of the data semantic boundary is determined based on the association relationship.

[0014] Extract the semantic feature vectors of the preceding field set and the subsequent field set of the candidate position, calculate the vector distance between the semantic feature vectors as the semantic relevance of the data content of the candidate position, determine the candidate position with the semantic relevance lower than the preset semantic distinction threshold as the data semantic boundary, and segment the medical data object according to the data semantic boundary to obtain multiple data segments;

[0015] The data content within each data segment is traversed and categorized according to the data field type to obtain the type features corresponding to the data field type. The type features of different data field types are combined and encoded to generate the feature summary containing multi-dimensional features.

[0016] The feature summary of the current data segment is bidirectionally associated with the feature summaries of the preceding and subsequent data segments to generate a bidirectional chain identifier containing dependencies, including:

[0017] Calculate the positional difference distribution between the feature summary of the current data segment and the feature summary of the preceding data segment, and generate positional interleaving rules based on the positional difference distribution; interleave and mix the feature summaries according to the positional interleaving rules to obtain a forward mixed feature sequence, and perform nonlinear displacement transformation on each position element in the forward mixed feature sequence to generate a forward chain identifier;

[0018] The feature summary of the current data segment is XORed with the feature summary of the subsequent data segment to obtain backward difference features. A backward difference feature matrix is ​​constructed based on the backward difference features. Row and column permutation transformations and dimensionality reduction mappings are performed on the backward difference feature matrix to generate backward chain identifiers.

[0019] Based on a preset energy threshold, high-frequency components in the forward chain identifier and low-frequency components in the backward chain identifier are extracted; frequency domain fusion and inverse transform are performed on the high-frequency components and the low-frequency components to obtain a spatial domain fused identifier; the forward chain identifier, the backward chain identifier, and the spatial domain fused identifier are concatenated according to the splicing order determined by the parity of the position indices of the multiple data segments to generate the bidirectional chain identifier containing the preceding and following dependencies.

[0020] Re-extracting feature summaries from each data segment and performing a bidirectional comparison with the bidirectional chain identifier includes:

[0021] Multi-scale block sampling is performed on the initial data of the current data segment, the preceding data segment, and the subsequent data segment respectively. Local statistical characteristics of each sampling block are extracted, and the local statistical characteristics are aggregated through adaptive weights to obtain the current verification feature summary, the preceding verification feature summary, and the subsequent verification feature summary.

[0022] The current verification feature summary is concatenated with the previous verification feature summary. The concatenated feature sequence is then subjected to convolutional kernel filtering to extract the peak position and peak amplitude of the filter response and encode them as verification forward feature codes. The forward chain identifier is decoded to obtain the forward feature codes. The Hamming distance between the verification forward feature codes and the forward feature codes is calculated to obtain the forward dissimilarity.

[0023] The current verification feature digest and the subsequent verification feature digest are constructed into a two-dimensional matrix and singular value decomposition is performed. The principal singular values ​​and the corresponding singular vectors are serialized into the verification backward feature code. The backward chain identifier is decoded to obtain the backward feature code. The Euclidean distance between the verification backward feature code and the backward feature code is calculated to obtain the backward dissimilarity.

[0024] Based on the bidirectional chain identifier, forward and backward traceability verifications are performed on the medical data object to generate an integrity verification result containing directional difference information, including:

[0025] The spatial fusion identifier is decomposed into wavelet multi-level decomposition. Approximation coefficients and detail coefficients are extracted at each decomposition level. The energy ratio of the approximation coefficients and the sparsity of the detail coefficients at each level are calculated to construct a multi-level frequency domain feature matrix.

[0026] The forward frequency domain representation is obtained by performing a Fourier transform on the forward feature code of the verification, and the backward frequency domain representation is obtained by performing a discrete cosine transform on the backward feature code of the verification. The dominant mode components are extracted by performing a tensor product operation on the forward frequency domain representation and the backward frequency domain representation in the frequency domain space and then performing principal component projection.

[0027] The dominant mode component is matched layer by layer with the features of the corresponding level of the multi-level frequency domain feature matrix, the feature response intensity value of each level is calculated, and the cosine similarity between the feature response intensity value and the preset standard response mode is calculated to obtain the tampering confidence.

[0028] The position index is combined with the numerical values ​​of the forward difference and the backward difference to encode the directional difference information. The directional difference information is then encapsulated with the tamper confidence through field mapping and hierarchical nesting to generate the integrity verification result.

[0029] Based on the spatial distribution relationship between the forward and backward break positions of the bidirectional chain identifier, the temporal path and scope of influence of the tampering operation are reconstructed, including:

[0030] The forward chain identifier and the current feature summary of the adjacent preceding data segment are converted into frequency domain representation and the frequency domain amplitude spectrum is extracted. When the spectral correlation coefficient between the frequency domain amplitude spectra is lower than the preset forward correlation threshold, the interval between the current position index and the position index of the adjacent preceding data segment is marked as the forward break position.

[0031] The backward chain identifier and the current feature summary of the adjacent subsequent data segment are respectively constructed into multi-dimensional feature vectors. In the feature space, when the Mahalanobis distance of the multi-dimensional feature vector exceeds the preset backward distance threshold, the interval between the current position index and the position index of the adjacent subsequent data segment is marked as the backward break position.

[0032] Project all the forward fracture locations and the backward fracture locations into the location index space, and calculate the position offset between the distribution center of the forward fracture location and the distribution center of the backward fracture location.

[0033] The scope of influence of the tampering operation is determined based on the coverage of the forward and backward fracture locations in the location index space; the propagation direction of the tampering operation is determined based on the direction of the location offset; and according to the propagation direction, the data segments within the scope of influence are sorted in chronological order according to the location index to construct a chronological path.

[0034] The tampered region is identified based on the isolation marker and the temporal path, and the tampered region is used as feedback information to adjust the semantic boundary identification, reducing the segmentation granularity of the tampered region. Simultaneously, redundant chain-like association paths are added to the tampered region, including:

[0035] Based on the isolation marker and the temporal path, a location-temporal fusion feature space is constructed. In the location-temporal fusion feature space, each data segment is clustered, and the continuous region formed after clustering is marked as the tampered region.

[0036] Extract the semantic inflection points and semantic density distribution within the tampered region; perform secondary segmentation on the tampered region based on the semantic inflection points and semantic density distribution to generate enhanced data segments; and extract enhanced feature summaries for each enhanced data segment.

[0037] Based on the enhanced feature summary, cross-segment forward chain identifiers and cross-segment backward chain identifiers are constructed respectively between the current enhanced data segment and the non-adjacent preceding enhanced data segment and the non-adjacent subsequent enhanced data segment through skip-type association coding.

[0038] The cross-segment forward chain identifiers and cross-segment backward chain identifiers are constructed into an association matrix according to their position index order. The transitive closure between each cross-segment chain identifier in the association matrix is ​​calculated. Multiple backup verification paths are generated based on the reachable paths in the transitive closures. The backup verification paths are topologically fused with directly adjacent association paths to construct the redundant chain association paths of the tampered area.

[0039] A second aspect of this invention provides an intelligent anti-tampering system for medical data based on chained storage, comprising:

[0040] The first unit is used to adaptively segment the medical data objects to be stored according to the data semantic boundaries, resulting in multiple data segments and extracting feature summaries for each segment.

[0041] The second unit is used to perform bidirectional association weaving of the feature summary of the current data segment with the feature summaries of the preceding data segment and the subsequent data segment, respectively, to generate a bidirectional chain identifier containing the preceding and following dependencies, and to embed the bidirectional chain identifier into the storage structure of the current data segment.

[0042] The third unit is used to re-extract the feature summaries of each data segment and compare them bidirectionally with the bidirectional chain identifier. Based on the bidirectional chain identifier, the medical data object is subjected to forward and backward traceability verification to generate an integrity verification result containing directional difference information.

[0043] The fourth unit is used to reconstruct the temporal path and scope of the tampering operation based on the spatial distribution relationship between the forward and backward break positions of the bidirectional chain identifier when there is an inconsistency in the integrity verification results, and to isolate and mark the data segments within the scope of the influence.

[0044] The fifth unit is used to identify the tampered region based on the isolation marker and the time sequence path, and use the tampered region as feedback information to adjust the semantic boundary recognition, reduce the segmentation granularity of the tampered region, and add redundant chain association paths to the tampered region to realize intelligent anti-tampering of the medical data object.

[0045] A third aspect of the present invention,

[0046] An electronic device is provided, comprising:

[0047] processor;

[0048] Memory used to store processor-executable instructions;

[0049] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0050] Fourth aspect of the embodiments of the present invention,

[0051] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0052] The beneficial effects of this application are as follows:

[0053] By adaptively segmenting medical data according to semantic boundaries, flexible control over the granularity of data segmentation is achieved. Compared with traditional fixed-block methods, this better suits the structural characteristics of medical data and improves anti-tampering efficiency. An innovative bidirectional associative weaving mechanism for preceding and subsequent data segments is constructed, forming a bidirectional dependent chain structure. Compared with traditional unidirectional hash chains, this allows for data verification from two directions, significantly enhancing the reliability of tamper detection. Through a dual verification mechanism of forward and backward traceability, integrity verification results containing directional difference information can be generated, accurately locating the tampering site and solving the problem that traditional methods can only detect the existence of tampering but cannot pinpoint the location.

[0054] Through the above-mentioned technological innovations, this invention achieves intelligent and precise anti-tampering protection for medical data, significantly improving the credibility and integrity of medical data. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the intelligent anti-tampering method for medical data based on chained storage, as described in an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of the bidirectional chain identifier generation process. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0059] Figure 1 This is a flowchart illustrating an embodiment of the intelligent anti-tampering method for medical data based on chained storage according to the present invention. Figure 1 As shown, the method includes:

[0060] The medical data objects to be stored are adaptively segmented according to the data semantic boundaries to obtain multiple data segments, and feature summaries are extracted from each segment.

[0061] The feature summary of the current data segment is bidirectionally associated with the feature summaries of the preceding and subsequent data segments to generate a bidirectional chain identifier containing the dependencies between the preceding and subsequent segments, and the bidirectional chain identifier is embedded in the storage structure of the current data segment.

[0062] The feature summaries of each data segment are re-extracted and compared bidirectionally with the bidirectional chain identifier. Based on the bidirectional chain identifier, the medical data object is subjected to forward and backward traceability verification to generate an integrity verification result containing directional difference information.

[0063] When the integrity verification results are inconsistent, the temporal path and scope of the tampering operation are reconstructed based on the spatial distribution relationship between the forward and backward break positions of the bidirectional chain identifier, and the data segments within the scope of the impact are isolated and marked.

[0064] The tampered region is identified based on the isolation marker and the time sequence path, and the tampered region is used as feedback information to adjust the semantic boundary recognition, reduce the segmentation granularity of the tampered region, and add redundant chain association paths to the tampered region to realize intelligent anti-tampering of the medical data object.

[0065] In one optional implementation, the medical data object to be stored is adaptively segmented according to data semantic boundaries to obtain multiple data segments, and feature summaries are extracted from each segment, including:

[0066] The format of the medical data object is identified to determine its data organization structure. Based on the data organization structure, the fields of the medical data object are parsed to obtain the field identifier and field content of each data field.

[0067] The data field type of each data field is determined based on the field identifier. By analyzing the reference and nesting relationships in the field content, the association relationship between the data fields is obtained, and the candidate position of the data semantic boundary is determined based on the association relationship.

[0068] Extract the semantic feature vectors of the preceding field set and the subsequent field set of the candidate position, calculate the vector distance between the semantic feature vectors as the semantic relevance of the data content of the candidate position, determine the candidate position with the semantic relevance lower than the preset semantic distinction threshold as the data semantic boundary, and segment the medical data object according to the data semantic boundary to obtain multiple data segments;

[0069] The data content within each data segment is traversed and categorized according to the data field type to obtain the type features corresponding to the data field type. The type features of different data field types are combined and encoded to generate the feature summary containing multi-dimensional features.

[0070] Medical data can exist in JSON, XML, HL7, or custom formats. The process involves format identification of medical data objects to determine their data organization structure. This is achieved by examining file header information, feature tags, or content structure pattern matching. For example, for XML data, the root element and namespace are identified; for JSON data, its key-value pair structure is analyzed; and for HL7 data, its delimiter definition and message structure are parsed. After format identification, a data parsing rule mapping table is established for subsequent field parsing.

[0071] The data structure is traversed using a parser adapted to this format, deconstructing each level of elements into field identifier-field content pairs. For example, parsing the "patientInfo" node in a patient's electronic medical record yields field identifiers such as "patientID", "name", and "gender" and their corresponding contents "12345678", "Zhang Mou", and "Male". For nested fields, a recursive parsing method is used to preserve the hierarchical relationship between fields and construct field identifier paths, such as "labResults.bloodTest.whiteBloodCell" pointing to a specific test value.

[0072] Establish a classification model for medical data field types, including basic types (such as text, numeric, and date) and medical-specific types (such as diagnostic codes, drug codes, and physiological indicators). Field types are determined by field naming characteristics, content patterns, and contextual relationships. For example, fields containing "date" or "time" and conforming to a date format are classified as date types; physiological indicators containing numerical ranges are classified as indicator types; and fields containing ICD codes are classified as diagnostic codes.

[0073] The system detects reference markers in field content, such as "see the result of item x," and establishes a reference relationship graph. It analyzes changes in the nested structure hierarchy, marking significant changes in nesting levels as candidate boundaries. Typical semantic boundaries in medical data include those between different medical records, different examination items, and medical history versus present illness. By constructing a field association graph, the system calculates the reference strength between fields and marks points of abrupt changes in association strength as candidate semantic boundary locations.

[0074] A medical text vectorization model is adopted to convert field content into semantic vectors. For structured fields, appropriate feature extraction methods are selected according to field type: statistical features are calculated for numerical fields; topic features are extracted for textual fields; and time-series features are extracted for time-related fields. The feature vectors of each field are merged according to weights to form the overall semantic representation of the field set.

[0075] The semantic relevance can be evaluated by using metrics such as cosine similarity or Euclidean distance to calculate the vector distance between the semantic feature vectors of the preceding and subsequent field sets. A larger vector distance indicates a more significant semantic difference. An appropriate semantic discrimination threshold, such as 0.7 (cosine similarity), can be set. When the semantic relevance corresponding to the vector distance is lower than this threshold, the candidate position is confirmed as a valid data semantic boundary.

[0076] Medical data objects are segmented based on the confirmed semantic boundaries of the data, preserving the integrity of the original data structure during segmentation and ensuring that each segment is an independent and semantically coherent data unit. For example, a complete electronic medical record can be segmented into multiple semantically independent segments such as basic information segment, medical history description segment, examination results segment, and treatment plan segment.

[0077] Specialized feature extraction algorithms are used for different types of fields: for diagnostic code fields, disease category and severity features are extracted; for drug code fields, pharmacological classification and dosage features are extracted; and for physiological indicator fields, normal / abnormal state and trend features are extracted. For example, abnormal markers and degrees of abnormality for all test indicators are extracted from the segments of test results to form an abnormal feature set.

[0078] Feature vector concatenation or weighted fusion is used to integrate various feature types. Dimensionality control is applied to the summary to retain key feature information. The final generated feature summary contains the core semantic features, important medical information features, and structural features of the data segment, forming the feature fingerprint of that segment. This feature summary reflects both the semantic information of the data content and the data structure features, providing efficient support for subsequent indexing, retrieval, and similarity matching of medical data.

[0079] By using the aforementioned adaptive segmentation and feature summary extraction methods for medical data, efficient storage and retrieval of medical big data can be achieved while preserving the semantic integrity of the data, thereby improving the response efficiency and data utilization value of medical information systems.

[0080] In one optional implementation, the feature summary of the current data segment is bidirectionally associated with the feature summaries of the preceding and subsequent data segments to generate a bidirectional chain identifier containing dependencies, including:

[0081] Calculate the positional difference distribution between the feature summary of the current data segment and the feature summary of the preceding data segment, and generate positional interleaving rules based on the positional difference distribution; interleave and mix the feature summaries according to the positional interleaving rules to obtain a forward mixed feature sequence, and perform nonlinear displacement transformation on each position element in the forward mixed feature sequence to generate a forward chain identifier;

[0082] The feature summary of the current data segment is XORed with the feature summary of the subsequent data segment to obtain backward difference features. A backward difference feature matrix is ​​constructed based on the backward difference features. Row and column permutation transformations and dimensionality reduction mappings are performed on the backward difference feature matrix to generate backward chain identifiers.

[0083] Based on a preset energy threshold, high-frequency components in the forward chain identifier and low-frequency components in the backward chain identifier are extracted; frequency domain fusion and inverse transform are performed on the high-frequency components and the low-frequency components to obtain a spatial domain fused identifier; the forward chain identifier, the backward chain identifier, and the spatial domain fused identifier are concatenated according to the splicing order determined by the parity of the position indices of the multiple data segments to generate the bidirectional chain identifier containing the preceding and following dependencies.

[0084] like Figure 2 As shown, the method includes:

[0085] Obtain feature digests for the current data segment, the preceding data segment, and the subsequent data segment. These feature digests can be generated using hash functions, such as SHA-256 or MD5 algorithms, converting the original data into fixed-length feature values. For example, for binary data blocks, a rolling hash algorithm can be used to calculate the feature digest, ensuring that the digest reflects the unique characteristics of the original data.

[0086] During the forward chain identifier generation process, the positional difference distribution of the feature summaries between the current data segment and the previous data segment is calculated. Specifically, the two feature summaries are compared position by position, and the value at each position is recorded to form a difference distribution map. Based on this difference distribution, positional interleaving rules are generated. The positional interleaving rules can use the index sequence of the difference positions as the permutation order. For example, if the difference positions are the 2nd, 5th, and 9th positions, the interleaving rule is to prioritize the elements at these positions.

[0087] The feature summaries of the current data segment are interleaved and mixed according to the positional interleaving rules. The interleaving and mixing process is a rearrangement and combination of feature summaries under the guidance of the positional interleaving rules. For example, if the feature summaries are 16-bit binary sequences, the elements at different positions and the elements at the same positions are alternately arranged according to the interleaving rules to form a forward mixed feature sequence.

[0088] A bit-cycle shift operation is used to perform a nonlinear displacement transformation on each element in the forward mixed feature sequence, generating a forward chain identifier. The shift distance is determined based on the feature length modulo the product of the element's position index and a prime number. This operation enhances the complexity and anti-interference capability of the forward chain identifier.

[0089] Performing an XOR operation between the feature summary of the current data segment and the feature summary of the subsequent data segment yields backward difference features. The XOR operation can effectively capture the difference information between the two feature summaries, providing a basis for establishing subsequent connections.

[0090] Assuming the backward difference feature length is 64 bits, it can be reassembled into an 8×8 feature matrix. This matrix is ​​then subjected to row and column permutations, such as cyclic right shift of rows and cyclic down shift of columns, to shuffle the original order of the difference features. The permuted matrix is ​​then subjected to dimensionality reduction mapping, such as flattening the two-dimensional matrix into a one-dimensional vector, or using principal component analysis to extract the main features, thereby generating the backward chain identifier.

[0091] Based on a preset energy threshold, high-frequency components in the forward chain identifier and low-frequency components in the backward chain identifier are extracted. The chain identifier can be converted to the frequency domain by fast Fourier transform. The energy threshold is set to 70% of the total energy. Components with frequencies higher than the threshold in the forward chain identifier and components with frequencies lower than the threshold in the backward chain identifier are extracted.

[0092] The extracted high-frequency and low-frequency components are fused in the frequency domain and restored by inverse transformation using a weighted average method to obtain a spatial domain fusion identifier. Weights are assigned according to the importance of the two components, and then the fusion result is converted back to the spatial domain by inverse fast Fourier transform.

[0093] If the current data segment's position index is odd, it is concatenated in the order of forward chain identifier, spatial fusion identifier, and backward chain identifier; if it is even, it is concatenated in the order of backward chain identifier, spatial fusion identifier, and forward chain identifier. The forward chain identifier, backward chain identifier, and spatial fusion identifier are then concatenated in the determined concatenation order to generate a bidirectional chain identifier containing the preceding and following dependencies.

[0094] This bidirectional chain identifier not only contains the characteristic information of the current data segment but also integrates the association information with preceding and subsequent data segments, forming a tightly linked data chain. This identifier method effectively prevents data tampering or insertion because any modification to a single data segment will change its association with preceding and following segments, thus being detected. In practical applications, such as data storage systems or blockchain technology, this bidirectional chain identifier can be used to verify the integrity and authenticity of data, improving system security.

[0095] In one optional implementation, re-extracting the feature summaries of each data segment and performing a bidirectional comparison with the doubly linked identifier includes:

[0096] Multi-scale block sampling is performed on the initial data of the current data segment, the preceding data segment, and the subsequent data segment respectively. Local statistical characteristics of each sampling block are extracted, and the local statistical characteristics are aggregated through adaptive weights to obtain the current verification feature summary, the preceding verification feature summary, and the subsequent verification feature summary.

[0097] The current verification feature summary is concatenated with the previous verification feature summary. The concatenated feature sequence is then subjected to convolutional kernel filtering to extract the peak position and peak amplitude of the filter response and encode them as verification forward feature codes. The forward chain identifier is decoded to obtain the forward feature codes. The Hamming distance between the verification forward feature codes and the forward feature codes is calculated to obtain the forward dissimilarity.

[0098] The current verification feature digest and the subsequent verification feature digest are constructed into a two-dimensional matrix and singular value decomposition is performed. The principal singular values ​​and the corresponding singular vectors are serialized into the verification backward feature code. The backward chain identifier is decoded to obtain the backward feature code. The Euclidean distance between the verification backward feature code and the backward feature code is calculated to obtain the backward dissimilarity.

[0099] For the current data segment, the preceding data segment, and the subsequent data segment, an adaptive window size is used to divide the data into blocks. The window size can be set to different scales such as 8, 16, and 32 bytes. At each scale, the data is divided into several non-overlapping sampling blocks. For example, for a 1KB data segment, using a 16-byte window size can result in 64 sampling blocks.

[0100] Local statistical characteristics are extracted for each sampling block, including mean, variance, entropy, and frequency domain features. The mean of the sampling block is calculated by summing all data elements in the block and then dividing by the number of elements. The variance is obtained by summing the squares of the differences between each element and the mean and then dividing by the number of elements. The entropy is obtained by calculating the information entropy of the data. The frequency domain features are extracted by performing a fast Fourier transform on the sampling block data.

[0101] The extracted local statistical features are aggregated using adaptive weights, which are dynamically adjusted based on the information content and importance of the data blocks. Information entropy is used as the weighting factor, with data blocks containing more information receiving higher weights. In this way, the current verification feature summary, the preceding verification feature summary, and the subsequent verification feature summary are obtained. Each feature summary can be represented as a 128-dimensional vector.

[0102] During forward alignment, the current verification feature summary is concatenated with the previous verification feature summary in sequence to form a 256-dimensional feature sequence. This sequence is then processed by convolutional kernel filtering, using a set of preset convolutional kernels (such as Gabor filters) to extract feature patterns. The convolutional kernel size can be set to 16, using a total of 8 filters in different directions.

[0103] Peak positions and amplitudes are extracted from the filtered response. The top 10 peaks with the strongest response are selected, and their positions and amplitudes are encoded into a forward verification feature code. The encoding method uses binary encoding, with position information occupying 20 bits and amplitude information occupying 12 bits, forming a total of 320-bit binary sequence.

[0104] The forward chain identifier is decoded to obtain the forward feature code, and the ciphertext in the chain identifier is converted into the plaintext feature code. The Hamming distance between the two forward feature codes is calculated, which is the number of bits that are different at corresponding positions in the two binary sequences, to obtain the forward dissimilarity.

[0105] During backward alignment, the current verification feature summary and the subsequent verification feature summary are constructed into a two-dimensional matrix. If both feature summaries are 128-dimensional vectors, they can be constructed into a 128×128 matrix, or they can be reorganized into a 16×16×8 three-dimensional tensor and then flattened into a 256×64 matrix.

[0106] The constructed two-dimensional matrix is ​​subjected to singular value decomposition. The top 5 largest singular values ​​and their corresponding singular vectors are selected, and these values ​​are serialized into a backtested feature code. The serialization process converts the singular values ​​and singular vectors into a unified floating-point sequence with a length of 128 elements.

[0107] The backward chain identifier is decoded to obtain the backward feature code. The decoding method is similar to that of the forward chain identifier decoding, but a different key and decoding algorithm are used. The Euclidean distance between the two backward feature codes is calculated, which is the square root of the sum of the squares of the differences between corresponding elements of the two sequences, to obtain the backward dissimilarity.

[0108] By calculating the forward and backward discrepancies, we can comprehensively judge the consistency of the association between the current data segment and its preceding and subsequent data segments, thereby verifying the integrity and continuity of the data. When the discrepancy is lower than the preset threshold (e.g., forward discrepancy less than 32, backward discrepancy less than 0.5), the data segments are considered to be consistent; otherwise, the data is judged to have been tampered with or damaged.

[0109] In practical applications, sampling parameters, feature extraction methods, and difference thresholds can be adjusted according to data type and security requirements to achieve optimal verification results. For example, texture feature extraction can be added for image data, and semantic feature analysis can be added for text data, thereby improving the accuracy and robustness of verification.

[0110] In one optional implementation, the medical data object is subjected to forward and backward traceability verification based on the bidirectional chain identifier to generate an integrity verification result containing directional difference information, including:

[0111] The spatial fusion identifier is decomposed into wavelet multi-level decomposition. Approximation coefficients and detail coefficients are extracted at each decomposition level. The energy ratio of the approximation coefficients and the sparsity of the detail coefficients at each level are calculated to construct a multi-level frequency domain feature matrix.

[0112] The forward frequency domain representation is obtained by performing a Fourier transform on the forward feature code of the verification, and the backward frequency domain representation is obtained by performing a discrete cosine transform on the backward feature code of the verification. The dominant mode components are extracted by performing a tensor product operation on the forward frequency domain representation and the backward frequency domain representation in the frequency domain space and then performing principal component projection.

[0113] The dominant mode component is matched layer by layer with the features of the corresponding level of the multi-level frequency domain feature matrix, the feature response intensity value of each level is calculated, and the cosine similarity between the feature response intensity value and the preset standard response mode is calculated to obtain the tampering confidence.

[0114] The position index is combined with the numerical values ​​of the forward difference and the backward difference to encode the directional difference information. The directional difference information is then encapsulated with the tamper confidence through field mapping and hierarchical nesting to generate the integrity verification result.

[0115] A Haar wavelet basis function is used to perform wavelet multi-level decomposition on the spatial domain fusion identifier, resulting in a four-level decomposition forming a pyramid structure. At each decomposition level, approximation coefficients and detail coefficients are extracted. Approximation coefficients represent the low-frequency components of the image, while detail coefficients represent the horizontal, vertical, and diagonal high-frequency components. For the i-th decomposition level, the energy ratio Ei of the approximation coefficients is calculated, which is the ratio of the energy of the approximation coefficients at that level to the energy of the original identifier; simultaneously, the sparsity Si of the detail coefficients is calculated, represented by the proportion of non-zero coefficients. These features from each level are organized into a multi-level frequency domain feature matrix M, where each row contains the energy ratio and sparsity information for that level.

[0116] A Fast Fourier Transform (FFT) is applied to the forward feature code of the verification to obtain the forward frequency domain representation Ff. This representation includes the amplitude and phase spectra, reflecting the distribution characteristics of the data in the frequency domain. A Discrete Cosine Transform (DCT) is performed on the backward feature code of the verification to obtain the backward frequency domain representation Fb. The DCT has the characteristic of energy concentration, concentrating the main information of the signal in the low-frequency region. The forward frequency domain representation Ff and the backward frequency domain representation Fb are combined in the frequency domain through tensor product operations to form a feature tensor T. Principal component analysis (PCA) is applied to this tensor for dimensionality reduction, retaining principal components that explain more than 85% of the variance, thereby extracting the dominant mode component D.

[0117] The dominant modal component D is matched and compared with the features of each level of the multi-level frequency domain feature matrix M. At each level l, the feature response intensity value Rl is calculated, which is the cross-correlation coefficient between the dominant modal component and the feature of that level. Simultaneously, based on statistical analysis of a large number of untampered medical data objects, a standard response pattern S is pre-established. The cosine similarity between the feature response intensity value Rl and the pre-established standard response pattern S is calculated to obtain the tampering confidence level C. The cosine similarity value ranges from [-1, 1]. The closer the value is to 1, the more consistent it is with normal data characteristics, and the lower the probability of tampering; the closer the value is to -1, the greater the difference from the normal pattern, and the higher the probability of tampering.

[0118] The position index is combined with the numerical values of the forward difference degree and the backward difference degree and encoded into directional difference information. Specifically, for each detection position p in the medical data object, the forward difference degree df is compared with the backward difference degree db. If df > db, it indicates that there is tampering propagating forward; if df < db, it indicates that there is tampering propagating backward; if the two are close, it indicates that this position is the source point of tampering or has not been tampered with. These directional difference information is field-mapped with the tampering confidence C and encapsulated in a hierarchical nested manner in JSON format to generate the final integrity verification result.

[0119] In practical applications, taking the CT image data of patients in a certain hospital as an example, when it is detected that a certain sequence of images has been tampered with, verification is carried out through this method. After the spatial domain fusion identifier is decomposed by wavelet, the energy ratio of the third layer deviates significantly from the standard mode, and the cosine similarity between the calculated feature response intensity value and the preset mode is -0.78, indicating that there is probably tampering. The directional difference information shows that the forward difference degree in the area from the 57th frame to the 62nd frame of the data is significantly higher than the backward difference degree, indicating that the tampering behavior propagates from front to back. The generated integrity verification result clearly identifies the tampering position, tampering direction and confidence, providing an accurate basis for subsequent data recovery and security processing.

[0120] Through the above implementation manners, it is possible to perform a comprehensive integrity verification on the medical data object, not only to determine whether there is tampering, but also to identify the propagation direction and specific position of the tampering, providing a strong guarantee for the security of medical data.

[0121] In an optional implementation manner, based on the spatial distribution relationship between the forward break position and the backward break position of the bidirectional chain identifier, reconstructing the timing path and influence range of the tampering operation includes:

[0122] Converting the forward chain identifier and the current feature digest of the adjacent previous data segment into a frequency domain representation and extracting the frequency domain amplitude spectrum. When the spectral correlation coefficient between the frequency domain amplitude spectra is lower than the preset forward correlation threshold, the interval between the current position index and the position index of the adjacent previous data segment is marked as the forward break position;

[0123] Constructing the backward chain identifier and the current feature digest of the adjacent subsequent data segment into a multi-dimensional feature vector respectively. In the feature space, when the Mahalanobis distance of the multi-dimensional feature vector exceeds the preset backward distance threshold, the interval between the current position index and the position index of the adjacent subsequent data segment is marked as the backward break position;

[0124] Projecting all the forward break positions and the backward break positions in the position index space, and calculating the position offset between the distribution center of the forward break positions and the distribution center of the backward break positions;

[0125] The scope of influence of the tampering operation is determined based on the coverage of the forward and backward fracture locations in the location index space; the propagation direction of the tampering operation is determined based on the direction of the location offset; and according to the propagation direction, the data segments within the scope of influence are sorted in chronological order according to the location index to construct a chronological path.

[0126] Frequency domain analysis is performed on the forward chain identifiers, converting the current feature summaries of the forward chain identifiers and adjacent preceding data segments into frequency domain representations. Specifically, a Fast Fourier Transform (FFT) is used to map the time domain information to the frequency domain, and the frequency domain amplitude spectrum is extracted. When the spectral correlation coefficient between the calculated frequency domain amplitude spectra is lower than a preset forward correlation threshold, the interval between the current position index and the position index of the adjacent preceding data segment is marked as a forward break location. For example, for audio and video data, the spectral correlation coefficient can be calculated by dividing the inner product of two spectral vectors by the product of their respective norms. When the correlation coefficient is lower than the forward correlation threshold of 0.85, it is determined to be a forward break.

[0127] Simultaneously, backward breakpoint detection is performed. The backward chain identifier and the current feature summary of the adjacent subsequent data segment are each constructed into a multi-dimensional feature vector. These feature vectors can contain information such as content hash values, color distribution, and texture features, typically with dimensions between 64 and 256. In the feature space, the Mahalanobis distance between these multi-dimensional feature vectors is calculated. When the distance exceeds a preset backward distance threshold, the interval between the current position index and the position index of the adjacent subsequent data segment is marked as the backward breakpoint. Mahalanobis distance considers the correlation between the various dimensions of the features and reflects the true differences in the data within the feature space more accurately than Euclidean distance.

[0128] After obtaining all forward and backward fracture locations, they are projected into a location index space, a one-dimensional linear space where each index represents a specific location in the original data. The location offset between the distribution centers of the forward and backward fracture locations is obtained by calculating the distribution centers of these two centers. The distribution centers can be obtained by calculating the weighted average of all fracture point indices, with the weights set based on the fracture strength.

[0129] Based on the coverage of the forward and backward break locations in the location index space, the scope of the tampering operation is determined. The coverage is typically defined as the smallest continuous interval that includes all break locations and appropriate boundary extensions. For example, for a video data segment, if the forward break locations are distributed between frame indices 120 and 150, and the backward break locations are distributed between frame indices 145 and 175, then the scope of the tampering operation is determined to be the interval between frame indices 120 and 175.

[0130] If the forward fracture center is smaller than the backward fracture center, it indicates that the tampering propagated from front to back; otherwise, it indicates that it propagated from back to front. After obtaining the propagation direction, the data segments within the affected area are sorted according to the temporal order of their position indices to construct a complete tampering operation temporal path.

[0131] Throughout the reconstruction process, forward and backward analysis complemented each other, providing a dual verification mechanism for tracing the source of tampering. By combining the temporal path and the scope of impact, not only can the boundaries of tampering be determined, but the evolution process of tampering can also be revealed, providing strong support for verifying the authenticity of digital content.

[0132] In one optional implementation, the tampered region is identified based on the isolation marker and the temporal path, and the tampered region is used as feedback information to adjust the semantic boundary identification, thereby reducing the segmentation granularity of the tampered region. Simultaneously, adding redundant chain-like association paths to the tampered region includes:

[0133] Based on the isolation marker and the temporal path, a location-temporal fusion feature space is constructed. In the location-temporal fusion feature space, each data segment is clustered, and the continuous region formed after clustering is marked as the tampered region.

[0134] Extract the semantic inflection points and semantic density distribution within the tampered region; perform secondary segmentation on the tampered region based on the semantic inflection points and semantic density distribution to generate enhanced data segments; and extract enhanced feature summaries for each enhanced data segment.

[0135] Based on the enhanced feature summary, cross-segment forward chain identifiers and cross-segment backward chain identifiers are constructed respectively between the current enhanced data segment and the non-adjacent preceding enhanced data segment and the non-adjacent subsequent enhanced data segment through skip-type association coding.

[0136] The cross-segment forward chain identifiers and cross-segment backward chain identifiers are constructed into an association matrix according to their position index order. The transitive closure between each cross-segment chain identifier in the association matrix is ​​calculated. Multiple backup verification paths are generated based on the reachable paths in the transitive closures. The backup verification paths are topologically fused with directly adjacent association paths to construct the redundant chain association paths of the tampered area.

[0137] Constructing a location-time fusion feature space based on isolation markers and temporal paths involves fusing the data representation across both spatial location and time series dimensions. Specifically, isolation markers can be represented as feature vectors Fi and temporal paths as temporal feature vectors Ti. A location-time fusion feature vector PTi is generated using the feature fusion function F(Fi, Ti). The fusion function can be implemented using weighted summation, concatenation, or attention mechanisms, for example, PTi = α·Fi + β·Ti, where α and β are weighting coefficients. In the generated location-time fusion feature space, clustering algorithms, such as K-means or DBSCAN, are applied to each data segment to group data points with similar features into one class. After clustering, the resulting continuous regions are marked as tampered regions. Here, continuity refers to a set of data segments that are close together and form clusters in the location-time feature space.

[0138] The semantic inflection points and semantic density distribution within the tampered region are extracted. Semantic inflection points are locations where the content theme or semantics undergoes a significant change, identified by calculating the rate of change in semantic similarity between adjacent data segments. When the rate of change exceeds a preset threshold, the location is marked as a semantic inflection point. Semantic density distribution is measured by calculating the semantic information entropy or keyword frequency at each location within the region. Based on the extracted semantic inflection points and semantic density distribution, the tampered region is further segmented to generate enhanced data segments. Specifically, segmentation can be performed at semantic inflection points, and low-density regions can be further subdivided based on semantic density to obtain semantically relatively uniform enhanced data segments. For each enhanced data segment, an enhanced feature summary is extracted. This summary contains the segment's key semantic information, structural features, and content features, and deep learning models such as BERT or TransformerXL can be used for feature extraction.

[0139] Based on enhanced feature summarization, a chain-like identifier is constructed between the current enhanced data segment and non-adjacent data segments through skip-associative coding. Skip-associative coding considers not only the relationship between directly adjacent segments but also establishes connections between long-distance segments. Specifically, for the current enhanced data segment Si, the feature similarity or semantic association degree between it and the non-adjacent preceding enhanced data segments {S(ik), ..., S(i-2)} and the non-adjacent subsequent enhanced data segments {S(i+2), ..., S(i+k)} is calculated, where k is a preset association window size. Based on the calculation results, cross-segment forward chain identifier FLi and cross-segment backward chain identifier BLi are generated. These identifiers can be represented as similarity vectors or binary relation matrices.

[0140] The cross-segment forward and backward chain identifiers are constructed into an association matrix RM according to their position index order. The element RM[i,j] of the association matrix represents the direct association strength from the strengthened data segment Si to Sj. The transitive closure TC between each cross-segment chain identifier in the association matrix is ​​calculated. The transitive closure calculation can use the Warshall algorithm or matrix exponentiation. The transitive closure represents all possible indirect association paths, where TC[i,j] indicates whether there is one or more paths from Si to Sj. Based on the reachable paths in the transitive closure, multiple backup verification paths are generated. In specific implementations, shortest path algorithms such as Dijkstra's algorithm can be used to find a set of highly reliable paths in the transitive closure. The backup verification paths are topologically merged with directly adjacent association paths to construct redundant chain association paths for the tampered region. During the topology fusion process, direct paths and backup paths can be weighted and merged according to reliability to form an association network structure with redundancy protection.

[0141] In practical applications, such as medical video tampering detection, the above method can effectively identify edited or replaced video segments. When a segment in a medical video is replaced, the temporal continuity between that segment and the original video is disrupted. By constructing a position-temporal fusion feature space, these discontinuous regions can be quickly located. Subsequently, semantic analysis and secondary segmentation of the tampered region can accurately identify the tampering boundary and reduce false positives. The construction of redundant chain-like association paths provides a multiple verification mechanism for the detection results. Even if the tamperer attempts to imitate some features of the original content, it is difficult to simultaneously satisfy the consistency requirements of multiple verification paths, thereby improving the robustness of the detection system.

[0142] Through the above technical solutions, the accuracy of identifying tampered areas is improved, while the segmentation granularity is more refined. The introduction of redundant chain-like association paths also enhances the reliability of the detection results, providing an effective guarantee for the verification of the authenticity of digital media.

[0143] This invention relates to an intelligent anti-tampering system for medical data based on chained storage, the system comprising:

[0144] The first unit is used to adaptively segment the medical data objects to be stored according to the data semantic boundaries, resulting in multiple data segments and extracting feature summaries for each segment.

[0145] The second unit is used to perform bidirectional association weaving of the feature summary of the current data segment with the feature summaries of the preceding data segment and the subsequent data segment, respectively, to generate a bidirectional chain identifier containing the preceding and following dependencies, and to embed the bidirectional chain identifier into the storage structure of the current data segment.

[0146] The third unit is used to re-extract the feature summaries of each data segment and compare them bidirectionally with the bidirectional chain identifier. Based on the bidirectional chain identifier, the medical data object is subjected to forward and backward traceability verification to generate an integrity verification result containing directional difference information.

[0147] The fourth unit is used to reconstruct the temporal path and scope of the tampering operation based on the spatial distribution relationship between the forward and backward break positions of the bidirectional chain identifier when there is an inconsistency in the integrity verification results, and to isolate and mark the data segments within the scope of the influence.

[0148] The fifth unit is used to identify the tampered region based on the isolation marker and the time sequence path, and use the tampered region as feedback information to adjust the semantic boundary recognition, reduce the segmentation granularity of the tampered region, and add redundant chain association paths to the tampered region to realize intelligent anti-tampering of the medical data object.

[0149] A third aspect of the present invention provides an electronic device, comprising:

[0150] processor;

[0151] Memory used to store processor-executable instructions;

[0152] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0153] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0154] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0155] Finally, it should be noted that the above 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent anti-tampering of medical data based on chained storage, characterized in that, include: The medical data objects to be stored are adaptively segmented according to the data semantic boundaries to obtain multiple data segments, and feature summaries are extracted from each segment. The feature summary of the current data segment is bidirectionally associated and interwoven with the feature summaries of the preceding and subsequent data segments to generate a bidirectional chain identifier containing the dependencies between the preceding and subsequent data segments. This includes: calculating the positional difference distribution between the feature summary of the current data segment and the feature summary of the preceding data segment; generating a positional interleaving rule based on the positional difference distribution; interleaving and mixing the feature summaries according to the positional interleaving rule to obtain a forward mixed feature sequence; and performing a nonlinear shift transformation on each position element in the forward mixed feature sequence to generate a forward chain identifier. The feature summary of the current data segment is XORed with the feature summary of the subsequent data segment to obtain backward difference features. A backward difference feature matrix is ​​constructed based on the backward difference features. Row and column permutation transformations and dimensionality reduction mappings are performed on the backward difference feature matrix to generate backward chain identifiers. Based on a preset energy threshold, high-frequency components in the forward chain identifier and low-frequency components in the backward chain identifier are extracted; frequency domain fusion and inverse transformation are performed on the high-frequency components and the low-frequency components to obtain a spatial domain fusion identifier; the forward chain identifier, the backward chain identifier, and the spatial domain fusion identifier are concatenated according to the splicing order determined by the parity of the position indices of the multiple data segments to generate the bidirectional chain identifier containing the preceding and following dependencies, and the bidirectional chain identifier is embedded in the storage structure of the current data segment; The feature summaries of each data segment are re-extracted and compared bidirectionally with the bidirectional chain identifier. Based on the bidirectional chain identifier, the medical data object is subjected to forward and backward traceability verification to generate an integrity verification result containing directional difference information. When the integrity verification results are inconsistent, the temporal path and scope of the tampering operation are reconstructed based on the spatial distribution relationship between the forward and backward break positions of the bidirectional chain identifier, and the data segments within the scope of the impact are isolated and marked. The tampered region is identified based on the isolation marker and the time sequence path, and the tampered region is used as feedback information to adjust the semantic boundary recognition, reduce the segmentation granularity of the tampered region, and add redundant chain association paths to the tampered region to realize intelligent anti-tampering of the medical data object.

2. The method of claim 1, wherein, The medical data objects to be stored are adaptively segmented according to data semantic boundaries to obtain multiple data segments, and feature summaries are extracted from each segment, including: The format of the medical data object is identified to determine its data organization structure. Based on the data organization structure, the fields of the medical data object are parsed to obtain the field identifier and field content of each data field. The data field type of each data field is determined based on the field identifier. By analyzing the reference and nesting relationships in the field content, the association relationship between the data fields is obtained, and the candidate position of the data semantic boundary is determined based on the association relationship. Extract the semantic feature vectors of the preceding field set and the subsequent field set of the candidate position, calculate the vector distance between the semantic feature vectors as the semantic relevance of the data content of the candidate position, determine the candidate position with the semantic relevance lower than the preset semantic distinction threshold as the data semantic boundary, and segment the medical data object according to the data semantic boundary to obtain multiple data segments; The data content within each data segment is traversed and categorized according to the data field type to obtain the type features corresponding to the data field type. The type features of different data field types are combined and encoded to generate the feature summary containing multi-dimensional features.

3. The method of claim 1, wherein, Re-extracting feature summaries from each data segment and performing a bidirectional comparison with the bidirectional chain identifier includes: Multi-scale block sampling is performed on the initial data of the current data segment, the preceding data segment, and the subsequent data segment respectively. Local statistical characteristics of each sampling block are extracted, and the local statistical characteristics are aggregated through adaptive weights to obtain the current verification feature summary, the preceding verification feature summary, and the subsequent verification feature summary. The current verification feature summary is concatenated with the previous verification feature summary. The concatenated feature sequence is then subjected to convolutional kernel filtering to extract the peak position and peak amplitude of the filter response and encode them as verification forward feature codes. The forward chain identifier is decoded to obtain the forward feature codes. The Hamming distance between the verification forward feature codes and the forward feature codes is calculated to obtain the forward dissimilarity. The current verification feature digest and the subsequent verification feature digest are constructed into a two-dimensional matrix and singular value decomposition is performed. The principal singular values ​​and the corresponding singular vectors are serialized into the verification backward feature code. The backward chain identifier is decoded to obtain the backward feature code. The Euclidean distance between the verification backward feature code and the backward feature code is calculated to obtain the backward dissimilarity.

4. The method of claim 3, wherein, Based on the bidirectional chain identifier, forward and backward traceability verifications are performed on the medical data object to generate an integrity verification result containing directional difference information, including: The spatial fusion identifier is decomposed into wavelet multi-level decomposition. Approximation coefficients and detail coefficients are extracted at each decomposition level. The energy ratio of the approximation coefficients and the sparsity of the detail coefficients at each level are calculated to construct a multi-level frequency domain feature matrix. The forward frequency domain representation is obtained by performing a Fourier transform on the forward feature code of the verification, and the backward frequency domain representation is obtained by performing a discrete cosine transform on the backward feature code of the verification. The dominant mode components are extracted by performing a tensor product operation on the forward frequency domain representation and the backward frequency domain representation in the frequency domain space and then performing principal component projection. The dominant mode component is matched layer by layer with the features of the corresponding level of the multi-level frequency domain feature matrix, the feature response intensity value of each level is calculated, and the cosine similarity between the feature response intensity value and the preset standard response mode is calculated to obtain the tampering confidence. The position index is combined with the numerical values ​​of the forward difference and the backward difference to encode the directional difference information. The directional difference information is then encapsulated with the tamper confidence through field mapping and hierarchical nesting to generate the integrity verification result.

5. The method of claim 1, wherein, Based on the spatial distribution relationship between the forward and backward break positions of the bidirectional chain identifier, the temporal path and scope of influence of the tampering operation are reconstructed, including: The forward chain identifier and the current feature summary of the adjacent preceding data segment are converted into frequency domain representation and the frequency domain amplitude spectrum is extracted. When the spectral correlation coefficient between the frequency domain amplitude spectra is lower than the preset forward correlation threshold, the interval between the current position index and the position index of the adjacent preceding data segment is marked as the forward break position. The backward chain identifier and the current feature summary of the adjacent subsequent data segment are respectively constructed into multi-dimensional feature vectors. In the feature space, when the Mahalanobis distance of the multi-dimensional feature vector exceeds the preset backward distance threshold, the interval between the current position index and the position index of the adjacent subsequent data segment is marked as the backward break position. Project all the forward fracture locations and the backward fracture locations into the location index space, and calculate the position offset between the distribution center of the forward fracture location and the distribution center of the backward fracture location. The scope of influence of the tampering operation is determined based on the coverage of the forward and backward fracture locations in the location index space; the propagation direction of the tampering operation is determined based on the direction of the location offset; and according to the propagation direction, the data segments within the scope of influence are sorted in chronological order according to the location index to construct a chronological path.

6. The method of claim 1, wherein, The tampered region is identified based on the isolation marker and the temporal path, and the tampered region is used as feedback information to adjust the semantic boundary identification, reducing the segmentation granularity of the tampered region. Simultaneously, redundant chain-like association paths are added to the tampered region, including: Based on the isolation marker and the temporal path, a location-temporal fusion feature space is constructed. In the location-temporal fusion feature space, each data segment is clustered, and the continuous region formed after clustering is marked as the tampered region. Extract the semantic inflection points and semantic density distribution within the tampered region; perform secondary segmentation on the tampered region based on the semantic inflection points and semantic density distribution to generate enhanced data segments; and extract enhanced feature summaries for each enhanced data segment. Based on the enhanced feature summary, cross-segment forward chain identifiers and cross-segment backward chain identifiers are constructed respectively between the current enhanced data segment and the non-adjacent preceding enhanced data segment and the non-adjacent subsequent enhanced data segment through skip-type association coding. The cross-segment forward chain identifiers and cross-segment backward chain identifiers are constructed into an association matrix according to their position index order. The transitive closure between each cross-segment chain identifier in the association matrix is ​​calculated. Multiple backup verification paths are generated based on the reachable paths in the transitive closures. The backup verification paths are topologically fused with directly adjacent association paths to construct the redundant chain association paths of the tampered area.

7. A medical data intelligent anti-tampering system based on chained storage, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to adaptively segment the medical data objects to be stored according to the data semantic boundaries, resulting in multiple data segments and extracting feature summaries for each segment. The second unit is used to perform bidirectional association weaving of the feature summary of the current data segment with the feature summaries of the preceding data segment and the subsequent data segment, respectively, to generate a bidirectional chain identifier containing the preceding and following dependencies, and to embed the bidirectional chain identifier into the storage structure of the current data segment. The third unit is used to re-extract the feature summaries of each data segment and compare them bidirectionally with the bidirectional chain identifier. Based on the bidirectional chain identifier, the medical data object is subjected to forward and backward traceability verification to generate an integrity verification result containing directional difference information. The fourth unit is used to reconstruct the temporal path and scope of the tampering operation based on the spatial distribution relationship between the forward and backward break positions of the bidirectional chain identifier when there is an inconsistency in the integrity verification results, and to isolate and mark the data segments within the scope of the influence. The fifth unit is used to identify the tampered region based on the isolation marker and the time sequence path, and use the tampered region as feedback information to adjust the semantic boundary recognition, reduce the segmentation granularity of the tampered region, and add redundant chain association paths to the tampered region to realize intelligent anti-tampering of the medical data object.

8. An electronic device, comprising: include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images