Pathological section recognition neural network model weight encryption method

Abstract

The invention provides a pathological section recognition neural network model weight encryption method. The method comprises the following steps: acquiring a bit string as a pathological section recognition neural network model weight key; the bit string is written into the lower triangular matrix according to the fact that eight bits serve as a number, and a first data matrix is obtained; performing digital conversion on the first 8 bits in the bit string to obtain first element data; reading a weight matrix of the pathological section recognition neural network model into a blank matrix to obtain a weight vector matrix; combining the first element data matrix with the weight vector matrix to obtain a first data vector; according to the first data matrix and the first data vector, adopting a chaotic mapping algorithm to obtain an intermediate ciphertext vector; performing element data processing on the intermediate ciphertext vector to obtain a weight vector intermediate encryption matrix; and carrying out position disruption on the weight vector middle encryption matrix to obtain an encrypted weight vector matrix. According to the invention, the safety coefficient of the neural network weight can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of neural network models, and in particular to a weight encryption method for a neural network model for pathological slice recognition. Background Art

[0002] In recent years, emerging technologies such as deep neural networks (DNNs) have been widely developed and applied in medical pathology slice recognition with unprecedented performance. Deep neural networks simulate the neural connection structure of the human brain by building models and describe data features in layers through multiple transformation stages, which has brought breakthrough progress to large-scale data processing tasks such as images, videos and audio. Applying deep neural networks to medical case slice recognition, a pathology slice recognition neural network model is obtained, which provides auxiliary role for medical diagnosis. It can not only reduce the workload of diagnostic physicians and save time for analyzing medical case slices, but also assist diagnostic physicians and provide them with accurate analysis data information to improve the diagnostic accuracy of diagnostic physicians.

[0003] Pathological slice recognition neural network models usually need to be trained before they are put into formal use. During model training, it is necessary to capture a large amount of proprietary data from different scenarios in the target application, extensive computing resources, and adjust the network topology and correctly train the weights with the assistance of experts. Therefore, the weights of the pathological slice recognition neural network model have extremely high economic value and practical significance in the pathological slice recognition neural network model.

[0004] At present, the security protection level for weights in the processing of pathological section recognition neural network models is relatively poor. Therefore, the present invention proposes a pathological section recognition neural network model weight encryption method, which performs encryption processing on the pathological section recognition neural network model weights based on chaotic mapping, thereby ensuring that the pathological section recognition neural network model can run while improving the security of the neural network weights. Summary of the invention

[0005] The purpose of the present invention is to provide a pathological section recognition neural network model weight encryption method to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for encrypting weights of a neural network model for pathological slice recognition, comprising:

[0007] Obtain the bit string as the weight key of the neural network model for pathological slice recognition;

[0008] The weight key of the pathological section recognition neural network model is written into the lower triangular matrix according to 8 bits per number to obtain a first data matrix;

[0009] Perform digital conversion on the first 8 bits in the bit string to obtain the first element data;

[0010] The weight matrix of the pathological slice recognition neural network model is read into the blank matrix as a sub-element of the blank matrix to obtain a weight vector matrix;

[0011] Combining the first element data matrix with the weight vector matrix to obtain a first data vector;

[0012] Obtaining an intermediate ciphertext vector using a chaotic mapping algorithm according to the first data matrix and the first data vector;

[0013] Performing element data processing on the intermediate ciphertext vector to obtain an intermediate encryption matrix of the weight vector; selecting the first 8 elements in the intermediate ciphertext vector as the first part of the weight ciphertext in the weight vector matrix, using the last element to update the encryption of the second part of the weight in the weight vector matrix, obtaining the second part of the weight ciphertext of the weight vector matrix, and obtaining the intermediate encryption matrix of the weight vector according to the first part of the weight ciphertext and the second part of the weight ciphertext;

[0014] The position of the intermediate encrypted matrix of the weight vector is shuffled to obtain the encrypted weight vector matrix.

[0015] Furthermore, the lower triangular matrix is ​​a 9×9 integer matrix. When the weight key of the pathological section recognition neural network model is written into the lower triangular matrix as 8 bits per number, the elements on the diagonal are all 3, and the writing order is from 2 to 9 for row data and from 1 to row data minus 1 for column data.

[0016] Further, the intermediate ciphertext vector is obtained by using a chaotic mapping algorithm according to the first data matrix and the first data vector, including:

[0017] Performing transposition processing on the first data vector to obtain a first data vector processing vector;

[0018] A matrix operation is performed on the first data matrix and the first data vector processing vector to obtain an intermediate ciphertext vector.

[0019] Furthermore, after obtaining the encrypted weight vector matrix, the weight vector matrix is ​​updated using the encrypted weight vector matrix to obtain an updated weight vector matrix, and then the weight vector intermediate encrypted matrix is ​​re-obtained based on the updated weight vector matrix, wherein when the weight vector matrix is ​​updated using the encrypted weight vector matrix, a target number of cycles is determined, and the weight vector matrix is ​​updated and cycled multiple times according to the target number of cycles after the encrypted weight vector matrix is ​​obtained to obtain the final encrypted weight vector matrix.

[0020] Furthermore, after obtaining the encrypted weight vector matrix, the encrypted weight vector matrix is ​​stored and transmitted using encryption, the encrypted weight vector matrix, the first element data and the number of rows of the weight vector matrix are combined together to form ciphertext data, and then the ciphertext data is stored. When the pathological section recognition neural network model is applied, the ciphertext data is obtained by transmitting the ciphertext through a secure channel, and the ciphertext data is decrypted to obtain the plaintext data of the weight data, and then the plaintext data of the weight data is applied to the pathological section recognition neural network model for model processing.

[0021] Furthermore, after decrypting the ciphertext data, the plaintext data of the weight data is obtained, including:

[0022] After obtaining the weight key of the pathological slice recognition neural network model, write it into the lower triangular matrix according to 8 bits as a number, and obtain the inverse matrix of the lower triangular matrix of the written data to obtain the decryption transformation matrix;

[0023] Extracting the encrypted weight vector matrix from the ciphertext data, obtaining the data in the second part from the encrypted weight vector matrix, and combining the data in the second part with the first element data in the ciphertext data to obtain a second data vector;

[0024] An intermediate decryption vector is obtained by adopting a chaotic mapping algorithm according to the decryption transformation matrix and the second data vector;

[0025] Perform element data processing on the intermediate decrypted vector to obtain the intermediate plaintext matrix of the weight vector; use the first element in the intermediate decrypted vector to update the first element data, then use the first data to decrypt the next 8 elements and combine them with the updated first element data to obtain the intermediate plaintext matrix of the weight vector;

[0026] The weighted plaintext data of the pathological slice recognition neural network model is obtained by arranging the rows and columns of the intermediate plaintext matrix of the weight vector.

[0027] Furthermore, when decrypting the ciphertext data, the key source is shared through the cryptographic channel to obtain the pathological slice recognition neural network model weight key bit string in the encryption process, thereby obtaining the pathological slice recognition neural network model weight key.

[0028] Furthermore, when 8 bits are written into the lower triangular matrix as a number, data is added in the lower triangular matrix according to the cache elimination mechanism based on the index information combined with the two-layer hash table.

[0029] Further, when storing the ciphertext data, the ciphertext data is stored in a ciphertext data storage module, and the ciphertext data storage module stores the ciphertext data in a data block manner, including:

[0030] Obtain the number of data block nodes in the ciphertext data storage module;

[0031] Divide the ciphertext data into multiple sub-data according to the number of data block nodes by using a local mapping method;

[0032] The data block node is matched for the sub-data, and the sub-data is encrypted twice and stored in the data block node according to the matching result.

[0033] Furthermore, a ciphertext management system is constructed based on the ciphertext data storage module, the ciphertext management system comprising: a model application end, a ciphertext data storage module, a secure channel establishment module and a secure verification management module;

[0034] The model application end is used to request to obtain the ciphertext data and issue a request based on the identification information;

[0035] The security verification management module is used to verify the identification information according to the request requirements in combination with the data block node, obtain the verification result, issue a read instruction to the ciphertext data storage module according to the verification result; and perform changes and combinations according to the sub-data obtained by reading the ciphertext data storage module to obtain the ciphertext data;

[0036] The ciphertext data storage module is used to perform read instruction verification in the data block node, and when the read instruction verification passes, decrypt and read the information stored in the data block node, and provide the obtained sub-data to the security verification management module;

[0037] The security channel establishment module is used to establish a security channel between the security verification management module and the model application end according to the verification result, and transmit the ciphertext data to the model application end through the security channel after the security verification management module obtains the ciphertext data.

[0038] The present invention realizes efficient encryption processing of the weights of the neural network model for pathological slice recognition, improves the safety factor of the weights of the neural network model for pathological slice recognition, and can resist attacks during the encryption processing process, thereby ensuring the security of the weights of the neural network model for pathological slice recognition.

[0039] The present invention performs digital conversion and processing on the weight key of the neural network model for pathological section identification so as to obtain data convenient for performing the chaotic mapping algorithm based on the weight key of the neural network model for pathological section identification, thereby providing convenience for the implementation of the chaotic mapping algorithm. Moreover, the chaotic mapping algorithm is simple and convenient, and can quickly obtain an intermediate ciphertext vector in a short time, effectively reducing the time consumption of encryption processing. After obtaining the intermediate ciphertext vector, the element data processing and position scrambling are performed on the intermediate ciphertext vector, thereby further improving the security of the encrypted weight vector matrix and preventing the encrypted weight vector matrix from being easily cracked.

[0040] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0041] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0043] Figure 1 A schematic diagram of the steps of the method described in the present invention;

[0044] Figure 2 This is a schematic diagram of step six in the method described in the present invention;

[0045] Figure 3 A schematic diagram of the relationship between the encryption process and the decryption process in the method described in the present invention;

[0046] Figure 4 A schematic diagram of the steps of decryption processing in the method described in the present invention;

[0047] Figure 5 The schematic diagram of the ciphertext management system in the method described in the present invention is shown in FIG. DETAILED DESCRIPTION

[0048] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0049] like Figure 1 As shown, an embodiment of the present invention provides a method for encrypting weights of a neural network model for pathological slice recognition, comprising:

[0050] Step 1: Obtain a bit string as the weight key of the neural network model for pathological slice recognition;

[0051] Step 2: Write the weight key of the pathological section recognition neural network model into the lower triangular matrix according to 8 bits as a number to obtain a first data matrix;

[0052] Step 3: Digitally convert the first 8 bits in the bit string to obtain the first element data;

[0053] Step 4: read the weight matrix of the pathological section recognition neural network model into the blank matrix as a sub-element of the blank matrix to obtain a weight vector matrix;

[0054] Step 5: Combine the first element data matrix and the weight vector matrix to obtain a first data vector;

[0055] Step 6: Obtain an intermediate ciphertext vector using a chaotic mapping algorithm according to the first data matrix and the first data vector;

[0056] Step 7, perform element data processing on the intermediate ciphertext vector to obtain the intermediate encryption matrix of the weight vector; perform element data processing on the intermediate ciphertext vector to obtain the intermediate encryption matrix of the weight vector; select the first 8 elements in the intermediate ciphertext vector as the first part of the weight ciphertext in the weight vector matrix, use the last element to update the encryption of the second part of the weight in the weight vector matrix, obtain the second part of the weight ciphertext of the weight vector matrix, and obtain the intermediate encryption matrix of the weight vector according to the first part of the weight ciphertext and the second part of the weight ciphertext;

[0057] Step 8: shuffle the position of the weight vector intermediate encryption matrix to obtain the encrypted weight vector matrix.

[0058] In the above technical solution, the weight key of the pathological section recognition neural network model is a randomly generated bit string of length 8×n, where n<28.

[0059] In the above technical solution, the lower triangular matrix is ​​a 9×9 integer matrix.

[0060] In the above technical solution, when the first element data matrix is ​​combined with the weight vector matrix, the first element data is used as the first element in the first data vector, followed by the weight vector matrix. For example, the first element vector is u0, and the weight vector matrix is ​​IME 1×mn ,IME 1×mn Specifically expressed as: IME 1×mn =[IME 1×mn (1,1),IME 1×mn (1,2),IME 1×mn (1,8)], the first data vector is P = [u0,IME 1×mn (1,1),IME 1×mn (1,2),…,IME 1×mn (1,8)].

[0061] In the above technical solution, the first part of the weights in the weight vector matrix is ​​the first 8 elements, and the second part is the remaining elements.

[0062] The above technical solution realizes efficient encryption processing of the weights of the neural network model for pathological slice identification, improves the security factor of the weights of the neural network model for pathological slice identification, and can resist attacks during the encryption process to ensure the security of the weights of the neural network model for pathological slice identification. By digitally converting and processing the weight keys of the neural network model for pathological slice identification, data that is convenient for the chaotic mapping algorithm can be obtained based on the weight keys of the neural network model for pathological slice identification, which provides convenience for the implementation of the chaotic mapping algorithm. The chaotic mapping algorithm is simple and convenient, and can quickly obtain an intermediate ciphertext vector in a short time, effectively reducing the time consumption of encryption processing, and after obtaining the intermediate ciphertext vector, the element data processing and position scrambling are performed on the intermediate ciphertext vector, thereby further improving the security of the encrypted weight vector matrix and preventing the encrypted weight vector matrix from being easily cracked.

[0063] In one embodiment provided by the present invention, the lower triangular matrix is ​​a 9×9 integer matrix. When the weight key of the pathological slice recognition neural network model is written into the lower triangular matrix as 8 bits per number, the elements on the diagonal are all 3, and the writing order is from 2 to 9 for row data and from 1 to row data minus 1 for column data.

[0064] In the above technical solution, the writing order is from row data 2 to 9, and the column data is from 1 to row data minus 1. For example, if row data i=2, 3, 4, 5, 6, 7, 8, 9, then j=1, 2, ..., i-1.

[0065] The above technical solution makes the determinant of the lower triangular matrix always an odd number by taking 3 for all the elements on the diagonal, thereby ensuring that the determinant of the triangular matrix is ​​coprime with the modulus 256, thereby ensuring the existence of the cycle, and writing the weight key of the pathological slice recognition neural network model into the lower triangular matrix in sequence as a number of 8 bits to avoid the occurrence of disordered written data affecting the attribute characteristics of the lower triangular matrix.

[0066] like Figure 2 As shown, in one embodiment provided by the present invention, the intermediate ciphertext vector is obtained by using a chaotic mapping algorithm according to the first data matrix and the first data vector, including:

[0067] A1. Perform transposition processing on the first data vector to obtain a first data vector processing vector;

[0068] A2. Perform a matrix operation on the first data matrix and the first data vector processing vector to obtain an intermediate ciphertext vector.

[0069] In the above technical solution, when the transposition processing is performed on the first data vector, the first data vector is converted from the form of a row vector to the form of a column vector.

[0070] In the above technical solution, the number of rows of the first data vector processing vector is equal to the number of columns of the first data vector.

[0071] The above technical solution ensures the execution of the chaotic mapping algorithm by performing a transposition process on the first data vector so that the first data matrix and the first data vector can be operated, so that area-preserving mapping can be achieved, and the weights in the weight vector matrix are hidden in the mapping results, thereby improving the safety factor of the weights, so that the pathological section recognition neural network model can only be trained and processed by decrypting the weights under safe conditions, thereby preventing anyone from arbitrarily obtaining the weights and applying them to the pathological section recognition neural network model.

[0072] In one embodiment provided by the present invention, after obtaining the encrypted weight vector matrix, the weight vector matrix is ​​updated using the encrypted weight vector matrix to obtain an updated weight vector matrix, and then the weight vector intermediate encrypted matrix is ​​re-obtained based on the updated weight vector matrix, wherein when the weight vector matrix is ​​updated using the encrypted weight vector matrix, a target number of cycles is determined, and the weight vector matrix is ​​updated and cycled multiple times according to the target number of cycles after the encrypted weight vector matrix is ​​obtained to obtain the final encrypted weight vector matrix.

[0073] In the above technical solution, the target number of cycles can be set by oneself.

[0074] In the above technical solution, using the encrypted weight vector matrix to update the weight vector matrix means that the number of loops is 1. When the number of loops is 3, it means that the weight vector matrix is ​​updated three times, and the encrypted weight vector matrix obtained by performing steps five, six, seven and eight based on the updated weight vector matrix obtained for the third time is the final encrypted weight vector matrix.

[0075] The above technical solution uses the encrypted weight vector matrix to update the weight vector matrix so that the encrypted weight vector matrix can be obtained in multiple cycles, thereby realizing multiple encryption of the weight vector matrix and improving the security of the encrypted weight vector matrix.

[0076] In one embodiment provided by the present invention, after obtaining the encrypted weight vector matrix, the encrypted weight vector matrix is ​​stored and transmitted using encryption, the encrypted weight vector matrix, the first element data and the number of rows of the weight vector matrix are combined together to form ciphertext data, and then the ciphertext data is stored, such as Figure 3 As shown, when the pathological section recognition neural network model is applied, the ciphertext data is obtained by transmitting the ciphertext through a secure channel, and the ciphertext data is decrypted to obtain the plaintext data of the weight data, and then the plaintext data of the weight data is applied to the pathological section recognition neural network model for model processing.

[0077] In the above technical solution, the ciphertext data exists in the form of data pairs. For example, the encrypted weight vector matrix is ​​CIE 1×mn , the first element vector is u0, the number of rows of the weight vector matrix is ​​n, then the ciphertext data is (CIE 1×mn ,u0,n).

[0078] In the above technical solution, the secure channel is a transmission path with a higher transmission safety factor.

[0079] The above technical solution stores ciphertext data so that the weights of the pathological section recognition neural network model can be retained and used for a long time, thereby ensuring the effectiveness of the pathological section recognition neural network model. Moreover, when applying the pathological section recognition neural network model, the ciphertext data is obtained by transmitting the ciphertext through a secure channel, which can not only ensure the success rate of the ciphertext transmission, but also ensure the security of the ciphertext, avoiding illegal acquisition of the ciphertext data or damage to the ciphertext data during the ciphertext transmission process.

[0080] like Figure 4 As shown, in one embodiment provided by the present invention, the plaintext data of the weight data is obtained after decryption processing on the ciphertext data, including:

[0081] S1. After obtaining the weight key of the neural network model for pathological slice recognition, the key is written into the lower triangular matrix according to 8 bits as a number, and the inverse matrix of the lower triangular matrix of the written data is obtained to obtain the decryption transformation matrix;

[0082] S2, extracting the encrypted weight vector matrix from the ciphertext data, obtaining the data in the second part from the encrypted weight vector matrix, and combining the data in the second part with the first element data in the ciphertext data to obtain a second data vector;

[0083] S3, obtaining an intermediate decryption vector by using a chaotic mapping algorithm according to the decryption transformation matrix and the second data vector;

[0084] S4, performing element data processing on the intermediate decryption vector to obtain an intermediate plaintext matrix of the weight vector; using the first element in the intermediate decryption vector to update the first element data, and then using the first data to decrypt the last 8 elements and combine them with the updated first element data to obtain the intermediate plaintext matrix of the weight vector;

[0085] S5. Obtain the weighted plaintext data of the pathological slice recognition neural network model according to the row and column arrangement of the intermediate plaintext matrix of the weight vector.

[0086] In the above technical solution, the decryption and encryption processes for ciphertext data are inverse to each other.

[0087] In the above technical solution, when the data in the second part is combined with the first element data in the ciphertext data, the first element data is used as the last element in the second data vector. For example, the ciphertext data is CIE 1×mn The data in the second part of the encrypted weight vector matrix is: CIE(1,mn-7),CIE(1,mn-6),CIE(1,mn-5),…,CIE(1,mn), the first element vector is u0, and the second data vector C obtained at this time is C=[CIE(1,mn-7),CIE(1,mn-6),CIE(1,mn-5),…,CIE(1,mn),u0].

[0088] In the above technical solution, when the chaotic mapping algorithm is used to obtain the intermediate decryption vector according to the decryption transformation matrix and the second data vector, the decryption transformation matrix and the transposed vector of the second data vector are used to perform an operation to obtain the intermediate decryption vector, that is: D = IE 9×9 ·C T , where D is the intermediate decryption vector, IE 9×9 To decrypt the transformation matrix, C is the second data vector and T is the transposed symbol.

[0089] In the above technical solution, when the first element in the intermediate decryption vector is used to update the first element data, u0=D(1,1), where u0 is the first element vector and D(1,1) is the first element in the intermediate decryption vector.

[0090] The above technical solution realizes the decryption of ciphertext data, so that the weights of the pathological section recognition neural network model can be obtained based on the ciphertext data, and then the pathological section recognition neural network model can be trained and applied based on the existing weights, which provides convenience for diagnostic physicians, allowing diagnostic physicians to directly use the pathological section recognition neural network model without wasting time on training and determining the weights of the pathological section recognition neural network model. At the same time, it can also avoid arbitrarily decrypting ciphertext data to obtain the plaintext data of the weights, thereby improving the security of the pathological section recognition neural network model weights and ensuring the value of the pathological section recognition neural network model weights.

[0091] In one embodiment provided by the present invention, when decrypting ciphertext data, a key source is shared through a cryptographic channel to obtain a pathological slice recognition neural network model weight key bit string during the encryption process, thereby obtaining a pathological slice recognition neural network model weight key.

[0092] In the above technical solution, the pathological slice recognition neural network model weight key bit string obtained during the decryption process is consistent with the pathological slice recognition neural network model weight key bit string randomly generated during the encryption process.

[0093] The above technical solution realizes the sharing of the weight keys of the neural network model for pathological slice recognition during the encryption process and the decryption process by sharing the key source through the cryptographic channel. Moreover, the sharing of the key source can only be achieved through the cryptographic channel, which has a high security factor and is not easy to be tampered with or destroyed.

[0094] In an embodiment provided by the present invention, when a number is written into the lower triangular matrix as 8 bits, data is added in the lower triangular matrix according to the cache elimination mechanism based on the index information combined with the two-layer hash table.

[0095] In the above technical solution, the cache elimination mechanism is implemented by combining the LRU (Least recently used) algorithm and the timeout period.

[0096] The above technical solution uses a cache elimination mechanism to make data that has been recently accessed have a higher probability of being accessed in the future. Data can be eliminated based on historical access records to keep access data active, facilitating digital writing in the lower triangular matrix.

[0097] In one embodiment provided by the present invention, when storing ciphertext data, the ciphertext data is stored in a ciphertext data storage module, and the ciphertext data storage module stores the ciphertext data in a data block manner, including:

[0098] Obtain the number of data block nodes in the ciphertext data storage module;

[0099] Divide the ciphertext data into multiple sub-data according to the number of data block nodes by using a local mapping method;

[0100] The data block node is matched for the sub-data, and the sub-data is encrypted twice and stored in the data block node according to the matching result.

[0101] In the above technical solution, when the sub-data is double-encrypted and stored in the data block node according to the matching result, different data block nodes can use the same double-encryption storage method or different double-encryption storage methods, but at least two double-encryption storage methods are used in all nodes of the data block, and they are mixed in the data block nodes.

[0102] In the above technical solution, when the ciphertext data is locally mapped, the local mapping includes multiple branches. In each branch, only part of the ciphertext data can be output according to the ciphertext data, thereby splitting the ciphertext data and obtaining multiple sub-data.

[0103] The above technical solution realizes the storage of ciphertext data through the ciphertext data storage module, so that when the pathological section recognition neural network model is trained or applied, the plaintext data of the weight can be obtained through the ciphertext data, thereby improving the confidentiality of the pathological section recognition neural network model weight, ensuring the security of the pathological section recognition neural network model weight, and will not cause the ciphertext data to be damaged or lost. In addition, the ciphertext data storage module adopts a data block method to store the ciphertext data, which effectively improves the security performance of the ciphertext storage module, so that only authorized people can obtain the ciphertext data completely. In addition, the sub-data is encrypted twice and stored in the data block node according to the matching result, which increases the security factor of the sub-data and prevents the data stored in the data block node from being easily cracked abnormally.

[0104] like Figure 5 As shown, in one embodiment provided by the present invention, a ciphertext management system is constructed based on the ciphertext data storage module, and the ciphertext management system includes: a model application end, a ciphertext data storage module, a secure channel establishment module and a security verification management module;

[0105] The model application end is used to request to obtain the ciphertext data and issue a request based on the identification information;

[0106] The security verification management module is used to verify the identification information according to the request requirements in combination with the data block node, obtain the verification result, issue a read instruction to the ciphertext data storage module according to the verification result; and perform changes and combinations according to the sub-data obtained by reading the ciphertext data storage module to obtain the ciphertext data;

[0107] The ciphertext data storage module is used to perform read instruction verification in the data block node, and when the read instruction verification passes, decrypt and read the information stored in the data block node, and provide the obtained sub-data to the security verification management module;

[0108] The security channel establishment module is used to establish a security channel between the security verification management module and the model application end according to the verification result, and transmit the ciphertext data to the model application end through the security channel after the security verification management module obtains the ciphertext data.

[0109] In the above technical scheme, after the security verification management module obtains the ciphertext data and transmits the ciphertext data to the model application end through the security channel, the security channel and the ciphertext data are monitored, including: obtaining the initial state information of the security channel and the initial information of the ciphertext data in the security channel; when the ciphertext data enters the security channel, real-time monitoring is performed on the transmission of the ciphertext data through the security channel, and the monitoring data of the ciphertext data in the security channel is obtained to obtain the current monitoring data; the current monitoring data is compared with the monitoring data of the previous time to obtain the current monitoring difference information; according to the initial state information of the security channel and the initial information of the ciphertext data in the security channel, a rationality analysis is performed on the current monitoring difference information to determine whether the current monitoring difference information is due to the difference phenomenon caused by the transmission of the ciphertext data in the security channel, and an analysis and judgment result is obtained; according to the analysis and judgment result, an early warning is issued when the previous monitoring difference information is not due to the difference phenomenon caused by the transmission of the ciphertext data in the security channel, and at the same time, the ciphertext data is locked and the transmission is stopped in the security channel.

[0110] In the above technical solution, the number of model application terminals is not unique.

[0111] In the above technical solution, the model application end issues a request based on the instructions of the diagnosing doctor.

[0112] In the above technical solution, the model application end is connected to the security verification management module, and the security verification management module is connected to the ciphertext data storage module and the security channel establishment module.

[0113] In the above technical solution, when a read instruction is issued to the ciphertext data storage module according to the verification result, the read instruction is generated based on the identification information of the model application end, and each read instruction acts on a data block node.

[0114] The above technical scheme realizes the management of ciphertext data through the ciphertext management system, realizes the connection between the encryption process and the decryption processing, so that the diagnostic physician can safely obtain the ciphertext data and perform decryption processing to obtain the plaintext data of the pathological section recognition neural network model weights when the pathological section recognition neural network model weights are needed. In addition, the security verification management module verifies the identification information and the ciphertext data storage module verifies the reading instructions, realizing double verification, avoiding the model application end that does not meet the requirements from stealing the ciphertext data in the ciphertext data storage module by abnormal means, ensuring the value of the pathological section recognition neural network model weights, and ensuring the security of the ciphertext data. In addition, the security channel establishment module establishes a secure channel between the security verification management module and the model application end only after the security verification management module passes the identification information verification. This can not only ensure that the security verification management module responds to the request requirements of the model application end and securely transmits the ciphertext data to the model application end, but also effectively avoid the leakage of ciphertext data caused by the disorder of ciphertext data transmission. In addition, after the security verification management module obtains the ciphertext data and transmits the ciphertext data to the model application end through a secure channel, it monitors the secure channel and ciphertext data to promptly detect any tampering with the secure channel or ciphertext data, thereby ensuring the security of the ciphertext data. Moreover, by comparing the difference between the current monitoring data and the monitoring data of the previous time, real-time analysis of the real-time monitoring data can be achieved, so that the theft of the ciphertext data can be detected in time, and then it can be blocked in time, reducing the possibility of leakage of the ciphertext data.

[0115] Those skilled in the art should understand that the first and second in the present invention merely refer to different application stages.

[0116] Other embodiments of the present disclosure will be readily apparent to those skilled in the art after considering the specification and practicing the disclosure disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0117] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A weight encryption method for a neural network model for pathological slice recognition, characterized in that: The method comprises: Obtain the bit string as the weight key of the neural network model for pathological slice recognition; The weight key of the pathological section recognition neural network model is written into the lower triangular matrix according to 8 bits per number to obtain a first data matrix; Perform digital conversion on the first 8 bits in the bit string to obtain the first element data; The weight matrix of the pathological slice recognition neural network model is read into the blank matrix as a sub-element of the blank matrix to obtain a weight vector matrix; Combining the first element data matrix with the weight vector matrix to obtain a first data vector; Obtaining an intermediate ciphertext vector using a chaotic mapping algorithm according to the first data matrix and the first data vector; Performing element data processing on the intermediate ciphertext vector to obtain an intermediate encryption matrix of the weight vector; selecting the first 8 elements in the intermediate ciphertext vector as the first part of the weight ciphertext in the weight vector matrix, using the last element to update the encryption of the second part of the weight in the weight vector matrix, obtaining the second part of the weight ciphertext of the weight vector matrix, and obtaining the intermediate encryption matrix of the weight vector according to the first part of the weight ciphertext and the second part of the weight ciphertext; The position of the intermediate encrypted matrix of the weight vector is shuffled to obtain the encrypted weight vector matrix.

2. The method according to claim 1, characterized in that The lower triangular matrix is ​​a 9×9 integer matrix. When the weight key of the pathological section recognition neural network model is written into the lower triangular matrix as 8 bits per number, the elements on the diagonal are all 3, and the writing order is from 2 to 9 for row data and from 1 to row data minus 1 for column data.

3. The method according to claim 1, characterized in that The intermediate ciphertext vector is obtained by using a chaotic mapping algorithm according to the first data matrix and the first data vector, including: Performing transposition processing on the first data vector to obtain a first data vector processing vector; A matrix operation is performed on the first data matrix and the first data vector processing vector to obtain an intermediate ciphertext vector.

4. The method according to claim 1, characterized in that: After obtaining the encrypted weight vector matrix, the weight vector matrix is ​​updated using the encrypted weight vector matrix to obtain an updated weight vector matrix, and then the weight vector intermediate encrypted matrix is ​​re-obtained based on the updated weight vector matrix, wherein when the weight vector matrix is ​​updated using the encrypted weight vector matrix, a target number of cycles is determined, and the weight vector matrix is ​​updated and cycled multiple times according to the target number of cycles to obtain the encrypted weight vector matrix, to obtain the final encrypted weight vector matrix.

5. The method according to claim 1, characterized in that: After obtaining the encrypted weight vector matrix, the encrypted weight vector matrix is ​​stored and transmitted using encryption, the encrypted weight vector matrix, the first element data and the number of rows of the weight vector matrix are combined together to form ciphertext data, and then the ciphertext data is stored. When the pathological slice recognition neural network model is applied, the ciphertext data is obtained by transmitting the ciphertext through a secure channel, and the ciphertext data is decrypted to obtain the plaintext data of the weight data, and then the plaintext data of the weight data is applied to the pathological slice recognition neural network model for model processing.

6. The method according to claim 5, characterized in that After decrypting the ciphertext data, the plaintext data of the weight data is obtained, including: After obtaining the weight key of the pathological slice recognition neural network model, write it into the lower triangular matrix according to 8 bits as a number, and obtain the inverse matrix of the lower triangular matrix of the written data to obtain the decryption transformation matrix; Extracting the encrypted weight vector matrix from the ciphertext data, obtaining the data in the second part from the encrypted weight vector matrix, and combining the data in the second part with the first element data in the ciphertext data to obtain a second data vector; An intermediate decryption vector is obtained by adopting a chaotic mapping algorithm according to the decryption transformation matrix and the second data vector; Perform element data processing on the intermediate decrypted vector to obtain the intermediate plaintext matrix of the weight vector; use the first element in the intermediate decrypted vector to update the first element data, then use the first data to decrypt the next 8 elements and combine them with the updated first element data to obtain the intermediate plaintext matrix of the weight vector; The weighted plaintext data of the pathological slice recognition neural network model is obtained by arranging the rows and columns of the intermediate plaintext matrix of the weight vector.

7. The method according to claim 6, characterized in that When decrypting ciphertext data, the key source is shared through the cryptographic channel to obtain the pathological slice recognition neural network model weight key bit string in the encryption process, and obtain the pathological slice recognition neural network model weight key.

8. The method according to claim 6, characterized in that When writing 8 bits as a number into the lower triangular matrix, data is added in the lower triangular matrix according to the cache elimination mechanism based on the index information combined with the two-layer hash table.

9. The method according to claim 5, characterized in that When storing the ciphertext data, the ciphertext data is stored in the ciphertext data storage module, and the ciphertext data storage module uses a data block method to store the ciphertext data, including: Obtain the number of data block nodes in the ciphertext data storage module; Divide the ciphertext data into multiple sub-data according to the number of data block nodes by using a local mapping method; The data block node is matched for the sub-data, and the sub-data is encrypted twice and stored in the data block node according to the matching result.

10. The method according to claim 9, characterized in that Constructing a ciphertext management system based on the ciphertext data storage module, the ciphertext management system includes: a model application end, a ciphertext data storage module, a secure channel establishment module and a security verification management module; The model application end is used to request to obtain the ciphertext data and issue a request based on the identification information; The security verification management module is used to verify the identification information according to the request requirements in combination with the data block node, obtain the verification result, issue a read instruction to the ciphertext data storage module according to the verification result; and perform changes and combinations according to the sub-data obtained by reading the ciphertext data storage module to obtain the ciphertext data; The ciphertext data storage module is used to perform read instruction verification in the data block node, and when the read instruction verification passes, decrypt and read the information stored in the data block node, and provide the obtained sub-data to the security verification management module; The security channel establishment module is used to establish a security channel between the security verification management module and the model application end according to the verification result, and transmit the ciphertext data to the model application end through the security channel after the security verification management module obtains the ciphertext data.

Images