Data analysis method and device based on multi-party security computing

Abstract

The invention provides a data analysis method and device based on multi-party security computing. The method comprises the steps that different data holding units share and store image data and metadata of the units; determining an image content identification or analysis task calculation model; according to the calculation model, using a multi-party security calculation protocol to carry out joint deep learning on the image data, and training a content identification model; performing content identification prediction calculation on new image data by using the trained model and a multi-party security calculation protocol; performing related calculation analysis on feature data extracted from the image data according to analysis task requirements; each participating entity carries out calculation according to the distributed multi-party security calculation subtasks; after calculation is completed, the participating entities compare calculation results in a public verification mode; and outputting an image analysis calculation result after privacy processing. Through the scheme of the invention, the data analysis efficiency can be improved, and the credibility and transparency of the calculation result are realized.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data analysis method and device based on multi-party secure computing. Background Art

[0002] Multi-party computation (MPC) is a computing method that allows two or more close friends to collaborate to calculate the value of a function while ensuring that the input and calculation results of each party are kept confidential. The working principle of the common MPC protocol is to decompose the computing task into multiple simple arithmetic and logical operations through randomization, encryption and masking technology, and then the participating parties perform the calculations alternately and share the results, so as to realize the function calculation without leaking the input information.

[0003] Current data analysis methods (especially those for image data) are not safe enough, are not intelligent enough, and are inefficient. Summary of the invention

[0004] Based on the above problems, the present invention proposes a data analysis method and device based on multi-party secure computing. Through the scheme of the present invention, the idea of ​​multi-party secure computing is utilized to realize in-depth development of multi-party image data, while taking into account data privacy protection, solving practical problems and achieving good technical results.

[0005] In view of this, one aspect of the present invention proposes a data analysis method based on multi-party secure computing, comprising: Different data holders use blockchain technology to share and store their own privacy-processed image data and metadata; Determine the computational model for image content recognition or analysis tasks; According to the computing model, the multi-party secure computing protocol is used to conduct joint deep learning on the image data stored in the blockchain in an encrypted domain to train a content recognition model; Use the trained model and multi-party secure computing protocol to perform content recognition prediction calculations on new image data; Perform relevant calculation and analysis on the feature data extracted from the image data according to the analysis task requirements; Each participating entity performs calculations according to the assigned multi-party secure computation subtask; After the calculation is completed, the participating entities compare the calculation results through public verification; Output the privacy-protected image analysis calculation results.

[0006] Optionally, the step of using blockchain technology to share and store the privacy-processed image data and metadata of different data holders includes: Each data holder performs privacy encryption on the image data and metadata it holds; The privacy-processed image data and metadata are packaged into blocks and digitally signed using the public-private key pair of the data holder; The data holder broadcasts the block to the blockchain network, and other participating nodes verify the integrity and source of the data block through decryption and signature verification; Successfully verified blocks will be added to the blockchain for storage, establishing a database in the form of a distributed ledger.

[0007] Optionally, the step of determining the image content recognition or analysis task computing model includes: The data holder selects the appropriate deep learning model architecture based on task requirements; Decompose the model structure into basic operation units that can be calculated in parallel; Use homomorphic encryption or MPC technology to encode the privacy of parameters in the model; Use MPC protocol to implement basic unit calculation in the encryption domain; Each participating unit conducts distributed training based on the model part it holds, iteratively updates parameters and reaches consensus; The network structure and parameters are obtained through public verification and serve as the calculation basis for image analysis tasks.

[0008] Optionally, the step of performing joint deep learning on the image data stored in the blockchain in an encrypted domain using a multi-party secure computing protocol according to the computing model to train a content recognition model includes: Each data holding unit takes out a portion of its own encrypted image data as a training subset; Use the MPC protocol to encrypt and distribute the weight parameters and biases in the neural network model; Each participating unit calculates the forward propagation results and gradient information corresponding to its own data; Using the secure aggregation algorithm, the loss function gradients of each participating unit are aggregated in the MPC environment; Based on the aggregation results, each participating unit updates the locally held model segment through the MPC protocol; Repeat the forward and backward propagation calculations, and update the parameters in each round of gradient descent; After multiple rounds of iterative training, the final encrypted model of joint learning is obtained.

[0009] Optionally, the step of using the trained model and the multi-party secure computing protocol to perform content recognition prediction calculation on new image data includes: The data holder encrypts the new image to be predicted to generate a ciphertext image; According to the MPC protocol, each participating unit uses the model parameters stored in the local distributed state to perform forward propagation calculations; Each participating unit will encrypt and report its normalized data, and other participating units will securely aggregate it; Obtain the encrypted expression of the feature maps of each layer after aggregation, and perform forward calculations on subsequent layers; After executing the entire neural network, the encrypted probability distribution of the predicted class can be obtained; Each participating unit confirms the final prediction results through a public verification algorithm.

[0010] Optionally, the step of performing relevant computational analysis on the feature data extracted from the image data according to the analysis task requirements includes: Convert the image feature vector into a homomorphic encryption domain for representation; Perform encrypted statistical analysis on features; Use secure comparison protocol to achieve the sorting of feature encryption values; Calculate the correlation analysis results of the feature set using the secure aggregation protocol; Supports statistical analysis and calculation of image metadata through the MPC protocol.

[0011] Optionally, the step of each participating entity performing calculation according to the assigned multi-party secure computation subtask includes: Decompose the global computing model into multiple basic computing units; Perform task division and scheduling optimization based on the computing capabilities of each participating entity; According to the MPC protocol, each basic computing unit is encoded as a subtask and distributed to the corresponding participating entities; Each participating entity completes the corresponding security calculation program according to the type of computing unit to realize the basic operations within the unit; Report the processing results to the task coordinator in encrypted form; Repeat until all unit tasks are completed.

[0012] Optionally, after the calculation is completed, the step of comparing the calculation results between the participating entities in a public verification manner includes: Each participating entity uses its own plaintext test data to independently calculate the test indicators; Encrypt the calculation results and publish them together with a random digital signature; After the other participating entities verify the signature, they compare the encryption indicators through the MPC protocol to determine whether they are consistent; For indicators that can be encrypted linearly, verification is done directly in the encryption domain; For nonlinear indicators, prove the correctness of the calculation process and results.

[0013] Optionally, the step of outputting the privacy-processed image analysis calculation result includes: Encoding the calculation result with a security mechanism; Privacy techniques such as applying random noise perturbations to the encoded results and combining them into ciphertext; Take partial display to show statistical distribution; When transmitting the encoded results over the network, use transmission encryption to provide anti-tampering protection; Configure validity period control for the encoded results, and automatically delete them after expiration to reduce the risk of leakage; Provide users with an interface to filter the result granularity to balance the effect and privacy requirements.

[0014] Another aspect of the present invention provides a data analysis device based on multi-party secure computing, comprising: a control module, a determination module, a training module, a prediction module and a processing module; The control module is used to control different data holding units to share and store the privacy-processed image data and metadata of the units using blockchain technology; The determination module is used to determine the image content recognition or analysis task calculation model; The training module is used to perform joint deep learning on the image data stored in the blockchain in an encrypted domain using a multi-party secure computing protocol according to the computing model to train a content recognition model; The prediction module is used to perform content recognition prediction calculation on new image data using the trained model and the multi-party secure computing protocol; The processing module is used for: Perform relevant calculation and analysis on the feature data extracted from the image data according to the analysis task requirements; Each participating entity performs calculations according to the assigned multi-party secure computation subtask; After the calculation is completed, the participating entities compare the calculation results through public verification; Output the privacy-protected image analysis calculation results.

[0015] The technical solution of the present invention is adopted, and the data analysis method based on multi-party secure computing includes: different data holding units use blockchain technology to share and store the image data and metadata that have been privacy-processed in the unit; determine the image content recognition or analysis task calculation model; according to the calculation model, use the multi-party secure computing protocol to perform joint deep learning on the image data stored in the blockchain under the encryption domain to train the content recognition model; use the trained model and the multi-party secure computing protocol to perform content recognition prediction calculation on the new image data; perform relevant calculation analysis on the feature data extracted from the image data according to the analysis task requirements; each participating entity calculates according to the assigned multi-party secure computing subtask; after the calculation is completed, the participating entities compare the calculation results through public verification; and output the privacy-processed image analysis calculation results. Through the scheme of the present invention, deep learning and calculation tasks of multi-party image data from different sources can be realized under the premise of protecting the data privacy of all participating entities; distributed computing resources can be fully utilized to improve data analysis efficiency; various multi-party image calculation tasks under privacy rules can be realized; and the credibility and transparency of the calculation results can be achieved through public verification. In summary, this method utilizes the idea of ​​multi-party secure computing to achieve in-depth development of multi-party image data, while taking into account data privacy protection, solving practical problems, and achieving good technical results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of a data analysis method based on multi-party secure computing provided by an embodiment of the present invention; Figure 2 It is a schematic block diagram of a data analysis device based on multi-party secure computing provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0018] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0019] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0020] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0021] Refer to the following Figure 1 to Figure 2 To describe the data analysis method and device based on multi-party secure computing provided according to some embodiments of the present invention.

[0022] like Figure 1 As shown, an embodiment of the present invention provides a data analysis method based on multi-party secure computing, including: Different data holders use blockchain technology to share and store their own privacy-processed image data and metadata; In this step, image data includes but is not limited to: pixel array (recording the luminosity or color value of each pixel), pixel format (such as RGB, GRAYSCALE, etc.), image size (number of pixel rows and columns), compression format (such as JPG, PNG, etc.), etc. Image metadata includes but is not limited to: file name (uniquely identifies the image), file type (such as picture, video, etc.), shooting date / time (recording the image acquisition time), GPS location information (providing the image geographic location coordinates), camera parameters (such as aperture, shutter speed, etc.), hue information (such as white balance, saturation and other adjustment parameters), label information (person / object name, scene classification, etc.), annotation (text description of the image), author / copyright information (recording the image creator), application information (in what system the image is used), metadata standard (structural specification of describing metadata), etc. By sharing these core image data and descriptive metadata, in-depth analysis tasks based on visual content can be carried out. Before understanding and applying the content of image data, it is necessary to extract and process the metadata of the image, including: automatically extracting metadata from the image file itself, such as device parameters and time information stored in the EXIF ​​standard; identifying objects and scenes from the image content to generate new tagged metadata; manually inputting descriptive metadata, such as location, event information, etc.; identifying text information from the image through OCR technology to generate auxiliary metadata; extracting formatted metadata such as the location frame coordinates of the image subject through object detection technology; distinguishing between core metadata and auxiliary metadata with higher importance according to different purposes; using different storage formats for different metadata types, such as using structured databases to store digital metadata; using standard formats such as EXIF / IPTC to uniformly describe image metadata from different sources; using metadata cleaning and completion technology to handle missing and redundant information; and providing open interfaces to support dynamic upgrades and expansions of metadata. Through a systematic metadata extraction and processing process, the shared application of image data can be effectively supported.

[0023] Determine the computational model for image content recognition or analysis tasks; According to the computing model, the multi-party secure computing protocol is used to conduct joint deep learning on the image data stored in the blockchain in an encrypted domain to train a content recognition model; Use the trained content recognition model and multi-party secure computing protocol to perform content recognition prediction calculations on new image data; Perform relevant computational analysis (including statistics, sorting, etc.) on the feature data extracted from the image data according to the analysis task requirements; Each participating entity performs calculations according to the assigned multi-party secure computation subtask; After the calculation is completed, the participating entities compare the calculation results through public verification; Output the privacy-protected image analysis calculation results.

[0024] By adopting the technical solution of the embodiment of the present invention, it is possible to realize deep learning and computing tasks of multi-party image data from different sources under the premise of protecting the data privacy of all participating entities; it is possible to make full use of distributed computing resources to improve data analysis efficiency; it is possible to realize various multi-party image computing tasks under privacy rules; and it is possible to realize the credibility and transparency of computing results through public verification. In short, this method can realize the deep development of multi-party image data by using the idea of ​​multi-party secure computing, while taking into account data privacy protection, solving practical problems, and achieving good technical results.

[0025] In some possible implementations of the present invention, the steps of different data holders using blockchain technology to share and store the privacy-processed image data and metadata of the units include: Each data holder performs privacy encryption on the image data and metadata it holds (homomorphic encryption). In this step, each data holder, such as a hospital, research institute, etc., has certain image data and related metadata. In order to protect the privacy of this data, each data holder needs to encrypt its own data. The privacy encryption processing method mentioned here can use homomorphic encryption, that is, encrypt the original data without affecting subsequent calculations. Specifically, each unit can perform specific homomorphic encryption operations on image pixel arrays, metadata attributes, etc.: encrypt the image pixel array using an encryption algorithm based on additive homomorphism, use a homomorphic encryption scheme that supports multiplication for numerical and identification metadata fields, and consider random hash encryption for text and unstructured fields. The encrypted image data and metadata retain computational compatibility with the original data, but it is difficult to decrypt the encrypted data back to the original plaintext, so as to achieve privacy protection of the data during the calculation process.

[0026] The privacy-processed image data and metadata are packaged into blocks and digitally signed using the public-private key pair of the data holder; In this step, after each data holding unit completes the homomorphic encryption processing of the image data and metadata, it is packaged into a block. Here, a block can contain one or a small group of image data records and corresponding metadata descriptions; in order to achieve data integrity and immutability, each block needs to be digitally signed; the digital signature uses public-private key pair technology, and each data holding unit has a pair of exclusive public-private keys; the private key held by the unit is used to digitally sign the block content to generate signature information; other units can use the public key of this unit to verify whether this block is issued by this unit and the content has not been modified, so that each block carries both the data content and the identity of the unit that generated it. Digital signatures can achieve non-rejection and identity recognition of blocks, providing technical support for subsequent tasks that require verification of data source and integrity.

[0027] The data holder broadcasts the block to the blockchain network, and other participating nodes verify the integrity and source of the data block through decryption and signature verification; In this step, the blocks formed after each data holder completes processing need to be published to the blockchain network, where the blockchain network refers to a shared storage data chain built specifically for multi-party computing services; the data holder uses a broadcast mechanism to send the encrypted data block signed by itself to other nodes in the network; after receiving the broadcast, other nodes participating in the calculation will use the signature to verify: use the public key of the data unit to verify the validity of the digital signature and verify whether the content of the signed block has been modified; at the same time, the participating nodes cannot decrypt the encrypted data block to obtain the plaintext content, but can verify the source and integrity of the data, and only store the data block data on the local blockchain after verification. Through this verification and publishing mechanism, the tamper-proof transmission of data can be achieved.

[0028] Successfully verified blocks will be added to the blockchain for storage, establishing a database in the form of a distributed ledger.

[0029] In this embodiment, data holders can query and track blocks initiated by themselves, but cannot directly view the original image data or metadata of other units; the distributed ledger jointly constructed by all participating units contains the image IDs, descriptive metadata, etc. contributed by each unit, promoting cross-institutional resource sharing.

[0030] Through the solution of this embodiment, secure shared storage of image data and metadata between different data sources, decentralized verification management of data integrity and ownership, and real-time monitoring and tracking of each data holder's own resources are achieved, laying a data foundation for subsequent privacy-based joint learning and analysis, thereby achieving deep sharing and utilization of multi-party resources while protecting data privacy.

[0031] In some possible implementations of the present invention, the following methods may be used to solve the trust problem between data holders: MPC technology is used to re-encrypt data so that no party can obtain the data content of other parties alone. This can establish a mutual distrust mechanism at the computing level. Establish a data usage tracking mechanism, such as recording the reading and sharing of data on the blockchain. Any use of data must be agreed upon by other parties to avoid private use of data. Introducing a third-party regulator to supervise operations between data holders and impose immediate penalties if violations are discovered, thereby increasing the cost of compliance; Use MPC protocols that can pre-verify the correctness of results, such as zero-knowledge proofs, which allow parties to verify each other before and after the calculation, but the process and data are still hidden; The data holder can choose to provide only anonymous ID or feature data for calculation instead of raw pixel data, which can alleviate the trust issue to a certain extent; Establish a credit rating mechanism for data holders, record the historical behavior of the units, regard integrity as an important community value, and influence subsequent data acquisition and sharing permissions; A pledge mechanism is introduced, and violations will result in partial confiscation of the pledge as appropriate, thereby increasing the cost of violations.

[0032] In summary, the combination of the above technologies and mechanisms can effectively alleviate the problem of mutual distrust between data holders.

[0033] In some possible implementations of the present invention, the step of determining the image content recognition or analysis task computing model includes: The data holder selects the appropriate deep learning model architecture (such as CNN, RNN and other classic image classification models) according to task requirements; In this step, the data holder needs to complete an image classification or recognition computing task; before starting the multi-party model training, it is necessary to decide what type of deep learning model architecture to use; the choice of model architecture depends on the task requirements and data characteristics, and different architectures perform differently for different tasks. For image classification tasks, commonly used models include: convolutional neural network (CNN), which is suitable for extracting high-level features from low-level attributes; recurrent neural network (RNN), which considers sequence relationships and is suitable for some ordered tasks. Data units need to choose an appropriate architecture based on their own data characteristics (such as static images or video data), or refer to experience to choose classic models with good classification effects, such as ResNet, VGG, etc. Inform other participating units of the model type information they have chosen. By clarifying the model type, technical guidance is provided for subsequent multi-party security learning, while taking into account both task effects and computational complexity.

[0034] Decompose the model structure into basic operation units that can be calculated in parallel (such as convolution, pooling layers, etc.); In this step, the classic deep learning model (such as CNN) is usually composed of a complete network structure composed of different types of stacked layers. These layers contain different computing units, such as the operations of each convolution kernel in the convolution layer. In order to perform multi-party distributed computing, the entire model task needs to be split into independent subtasks. The data holding unit splits the selected model according to the hierarchical structure and extracts the basic computing units in each layer or part of the layer. For example, the convolution layer can be split into the multiplication and addition computing units of each convolution kernel, and the pooling layer can be split into the maximization unit in each window. These basic computing units are defined as subtasks that can be calculated in parallel, with the purpose of further scheduling and allocating tasks according to the network structure and computing resources. Therefore, model decomposition is the process of breaking down the end-to-end model task into basic computing tasks that can be processed in a distributed manner.

[0035] Use homomorphic encryption or MPC technology to encode the parameters (such as filters) in the model in a privacy-preserving manner; In this step, a large number of parameters (such as convolutional filters in the convolutional layer, etc.) are required in the training process of the deep learning model. These parameters contain sensitive information, and direct sharing will leak privacy, so the parameters need to be privately encoded, mainly using two technologies: homomorphic encryption (encrypting parameters while maintaining computability, such as Paillier encryption) and MPC (multi-party computing protocols distribute parameters across nodes and obtain results interactively during calculations). Specific operations include: digitizing filter parameters into homomorphic domains or MPC secret sharing formats; encoding other model parameters such as biases accordingly; the encoded parameters can be directly used for distributed model training. However, the outside world cannot directly interpret the plaintext parameter values ​​from the encoded data. The purpose is to protect the security of parameters used in model training in distributed deep learning.

[0036] Use the MPC protocol to implement basic unit calculations in the encrypted domain (such as convolution using secure multiplication, etc.); In this step, the model is decomposed into multiple basic operation unit tasks, such as each convolution kernel operation in the convolution layer. These basic units involve specific operations, such as the product and addition of convolution kernels. Direct calculation in the plaintext domain will leak intermediate data, so it is necessary to calculate in the encrypted domain. The MPC protocol can realize secure calculation under the condition that it is encrypted for all participants. For example, for the product of convolution kernels: use the encryption method supported by MPC for multiplication to encode the filter and feature map. According to the MPC protocol, each participant uses sharing to complete the multiplication operation, and the encrypted product result is obtained instead of plaintext; the sum of the results also uses the secure addition of MPC to calculate the result from the sharing. In this way, the basic unit operations are completed in the encrypted domain, protecting the data privacy of each participant. Therefore, this process is to use MPC to perform the calculation of basic operations of deep learning models under confidentiality conditions.

[0037] Each participating unit conducts distributed training based on the model part it holds, iteratively updates parameters and reaches consensus; In this step, the model is decomposed into several basic computing unit tasks. These unit tasks are distributed to each participating unit for parallel computing according to computing power. Each unit holds a part of the model, such as part of the convolution kernel operation of the convolution layer. Through multiple rounds of iterative training: each unit calculates the local gradient based on the part of the model and part of the samples it holds; uses the MPC protocol to exchange gradient information and aggregate and calculate the global gradient; calculates the new value of the updated encrypted parameter; parameter updates require all units to reach a consensus. Repeating this iterative update, the model parameters are continuously optimized and improved. Finally, each unit holds the encrypted model after distributed training and the updated sharing. Through a distributed and iterative approach, distributed joint training of deep learning models that protect privacy is achieved.

[0038] The network structure and parameters are obtained through public verification and serve as the calculation basis for image analysis tasks.

[0039] It should be noted that in the embodiments of the present invention, the participating unit refers to an entity organization that participates in data sharing and computing tasks, such as hospitals, research institutions, etc. Each participating unit holds a certain amount of image data and corresponding metadata. The participating unit needs to perform privacy processing on its own data and then publish it to the blockchain network for sharing. The participating node refers to the computing node that constitutes the blockchain technology infrastructure. After receiving the encrypted data broadcast by the participating unit, the participating node is responsible for storage and verification. The participating node is also the computing entity that performs the MPC multi-party secure computing task. Each participating unit may have one or more participating nodes deployed in the background to support computing. Simply put, the participating unit is the provider of data and tasks, and the participating node is the blockchain and distributed computing infrastructure, which is operated and managed by the participating unit. The participating unit issues tasks to the participating node, and the participating node supports the participating unit to complete privacy computing.

[0040] Through the solution of this embodiment, the computing framework of the deep learning model is determined while protecting data privacy, distributed secure computing is implemented to train the recognition model of the pattern between learning samples, multi-party security predictive analysis of the image content is performed based on the training results, and reliable computing power support is set to process a large number of image tasks. While improving the efficiency of sharing and utilizing image resources, the data sovereignty of all parties is guaranteed, thereby efficiently completing image-based sharing tasks.

[0041] In some possible implementations of the present invention, for an image content recognition model trained using the MPC protocol, the following method may be used to evaluate its performance in a specific task: Separate test data set: Each data holder pre-allocates a portion of the data for model testing, which does not participate in training.

[0042] Ciphertext testing: Encode the test image using technologies such as homomorphic encryption or MPC, and run the model in the encrypted domain for prediction.

[0043] Disclose some information: All parties involved can verify the prediction results based on some information such as the disclosed probability distribution.

[0044] Indicator calculation: Participants independently calculate common classification indicators such as accuracy and recall based on their own test data and prediction results.

[0045] Result comparison: Finally, all participants will cross-compare the classification indicator results to verify whether they are consistent, that is, the model effect.

[0046] Diagnostic testing: Diagnose the classification results given by the marginal example model and debug the parts that are easily misclassified by the model.

[0047] Performance tracking: Track the performance of models at different training stages on the same test set to understand the model performance curve.

[0048] Multiple executions: Run the same test set multiple times to average the results and reduce the impact of random errors on model evaluation.

[0049] The above process can effectively evaluate the effect of the MPC training model in actual image recognition tasks while protecting data privacy.

[0050] In some possible implementations of the present invention, the step of performing joint deep learning on the image data stored in the blockchain in an encrypted domain using a multi-party secure computing protocol according to the computing model to train a content recognition model includes: Each data holding unit takes out a portion of its own encrypted image data as a training subset; In this step, each data holding unit (such as a hospital) shares the image data to the blockchain network after privacy processing; in order to distribute the training model, it is necessary to extract sub-datasets from these shared data for training; because the data is completely confidential and the data of other units cannot be directly accessed, each unit only selects the data uploaded by itself as the local training data set. For example, if a unit has 1,000 pictures, it can randomly select 500 of them. These 500 encrypted pictures constitute the local sub-dataset of the unit for distributed training. The data of other units will not be accessed during the training process, but the global model can be updated. In this way, each unit only uses a part of the encrypted data it saves for training calculations. The purpose is to fully utilize the data resources stored in each unit for distributed deep learning while ensuring data privacy.

[0051] Use the MPC protocol to encrypt and distribute the weight parameters and biases in the neural network model; In this step, weight parameters and bias values ​​are very important hyperparameters in the deep learning model training process. Directly sharing these parameter values ​​will leak private data feature information, so these parameters need to be encrypted and protected in a distributed representation. The MPC protocol can achieve this goal: each parameter value is transformed into multiple secret shares using MPC, and these secret shares are randomly distributed to each participating node through the MPC protocol. Each node only holds a part of the secret share. For example, a weight value w is distributed after MPC processing as follows: Node 1 holds w1, Node 2 holds w2, and so on; the original parameter value can be restored through the collaboration of the secret shares held by each node, which realizes the encrypted protection representation of model parameters in a distributed scenario.

[0052] Each participating unit calculates the forward propagation results and gradient information corresponding to its own data; In this embodiment, each participating unit (such as a hospital) takes out a portion of its own data as a local training subset. In each iteration cycle of distributed training: The participating units first perform forward propagation calculations based on the global model (in MPC distributed representation): the local subset data is input into the network to obtain the prediction results; the results are expressed in encrypted form and cannot be cracked; Calculate the loss function and error rate of local data: compare with the true label and calculate the loss of this round; the loss value is also kept in encrypted form; The gradient information is calculated through the back-propagation algorithm: the gradient reflects the source of the error and the direction that needs to be optimized, and is also expressed in an encrypted distributed form; Each participating unit only calculates and holds its own partial gradient information; The goal is to use MPC to implement privacy-preserving forward and back-propagation calculations.

[0053] Using the secure aggregation algorithm, the loss function gradients of each participating unit are aggregated in the MPC environment; In this embodiment, each participating unit (such as a hospital) calculates its own gradient information through back propagation, but the gradient of a single participating unit is not enough to update the global model, and the gradient information of each participating unit needs to be aggregated to obtain the global gradient. Here, a secure aggregation algorithm is used to complete this process: through the MPC protocol, each participating unit shares its own gradient value with other nodes; each node uses the addition operation supported by MPC to add up the received gradient shares. In this way, the addition operation results are shared among each node, but no single node sees the plaintext gradient. This method realizes the aggregation operation of gradients, but maintains their privacy. What is finally obtained is the global gradient under MPC, which provides a basis for subsequent parameter updates. This process uses the characteristics of MPC to achieve secure aggregation of the gradient values ​​of participating units without leaking individual gradient details.

[0054] Based on the aggregation results, each participating unit updates the locally held model segment through the MPC protocol; In this embodiment, global gradient information has been obtained in the MPC environment through secure aggregation. This gradient will be used to update and optimize the model parameters, but each participating unit only holds a part of the model's secret share. Through the MPC protocol, the participating unit needs to update the part of the model held locally in combination with the global gradient: calculate the updated value of the parameter according to the gradient descent rule; apply the updated value to the locally held model share through operations supported by MPC (such as ciphertext domain subtraction); ensure that the shared participating units remain consistent after the update; after several rounds of iterations, the local model part is effectively optimized; continuous iteration of each unit will gradually converge the full model. By updating the ontology divisions in a decentralized and parallel manner by the units, distributed model training is achieved while protecting privacy.

[0055] Repeat the forward and backward propagation calculations, and update the parameters in each round of gradient descent; After multiple rounds of iterative training, the final encrypted model of joint learning is obtained.

[0056] Through the solution of this embodiment, deep learning tasks are carried out on the basis of protecting data privacy. By optimizing the model with better training effect globally, the data scale advantage from all parties is effectively utilized to provide safe and reliable multi-party joint learning services, make up for the shortcomings of a single data source, improve recognition capabilities, and thus use distributed thinking to enhance the model learning effect.

[0057] In some possible implementations of the present invention, the step of using the trained model and the multi-party secure computing protocol to perform content recognition prediction calculations on new image data includes: The data holder encrypts the new image to be predicted to generate a ciphertext image; In this step, the data holder (such as a hospital) collects new image data that needs to be predicted. Directly sharing these images may leak privacy information, so these new images need to be encrypted and encoded: the image pixel data is encrypted using the same homomorphic encryption mechanism as before; the image metadata is also encrypted like the previous training data; the generated new images are represented in an encrypted form, called ciphertext images; these ciphertext images retain the image structure features, but prevent direct inspection of the content; they can be directly used for subsequent distributed classification prediction tasks. The purpose is to use the established encryption system to provide privacy protection for new data to be predicted, and the generated ciphertext images can support downstream tasks while protecting privacy.

[0058] According to the MPC protocol, each participating unit uses the model parameters stored in the local distributed state to perform forward propagation calculations; In this step, through the previous distributed training, each participating unit (such as a hospital) locally saves some of the updated model parameters. These parameters are encrypted and distributed using the MPC algorithm and saved in multiple copies in each participating unit. When there is a new prediction task, the model needs to be used for inference prediction: According to the MPC algorithm, each participating unit uses its own model parameter sharing to start forward propagation calculation tasks, such as multiplication, addition, etc. These calculations are all performed in the encrypted domain, and the participating units will not retrieve the plaintext parameters. However, through MPC interaction, forward propagation with distributed parameter calculations can be achieved, and the prediction results obtained are also represented in an encrypted distributed form. In this way, while protecting data privacy, the distributed trained model is used to perform prediction tasks, realizing an end-to-end protection calculation process.

[0059] Each participating unit will encrypt and report its normalized data, and other participating units will securely aggregate it; In this step, each participating unit holds a part of the original data set. For the training effect, the data set needs to be standardized / normalized. Since the data is distributed among the participating units, it is necessary to adopt a distributed method for standardization: each unit calculates the mean and standard deviation based on its own data for local normalization. The normalized mean and standard deviation are encrypted and reported to other participating units. Other participating units use the MPC security convergence algorithm to summarize the received encrypted normalization parameters. After convergence, the mean and standard deviation of the global data set are obtained, but the encrypted state is kept to prevent the value from being exposed. This global standardized parameter is used in training and prediction in the future. The purpose is to perform unified standardized preprocessing of distributed data sets while ensuring data privacy.

[0060] Obtain the encrypted expression of the feature maps of each layer after aggregation, and perform forward calculations on subsequent layers; In this step, a convolutional neural network is used for image classification prediction tasks. The CNN network contains multiple convolution-pooling layers, and each layer calculates a feature map as input to the lower layer. Since the model and data are distributed in each participating unit: each unit uses the MPC protocol to perform forward propagation to calculate the feature map of this layer; encodes and packages its own feature map of this layer and sends it to other participating units; other units use the MPC protocol to securely aggregate and summarize the received feature maps; the aggregated feature maps are represented in an encrypted manner and propagated as the input of the lower layer. In this way, each layer repeatedly executes the secure aggregation-lower layer calculation sequence to calculate the entire network, and finally the output layer obtains the classification prediction result in encrypted form. This process ensures the confidential transmission of feature information during distributed computing and supports end-to-end model forward execution.

[0061] After executing the entire neural network, the encrypted probability distribution of the predicted class can be obtained; Each participating unit confirms the final prediction results through a public verification algorithm.

[0062] The solution of this embodiment performs deep learning reasoning tasks while protecting the privacy of images, uses distributed resources to accelerate predictions and reduce online response time, has the ability to identify the content of new images, and different entities can perform joint prediction services, providing a secure and reliable multi-party deep learning prediction interface, thereby efficiently carrying out image-based security content understanding services.

[0063] In some possible implementations of the present invention, in performing content recognition prediction calculations on new image data, the specific implementation process of each participating unit encrypting and reporting its own normalized data is as follows: Each participating unit calculates the feature map of the new image in its own area, such as channel value, response map, etc. Adding random noise to the feature map data to generate encrypted feature map data; Use homomorphic encryption algorithms (such as Paillier encryption) to encrypt encrypted feature graph data; After generating the encrypted feature map ciphertext, the participating units digitally sign it together with the random number encryption key; Publish the encrypted feature graph ciphertext and digital signature to the MPC network; Other participating units can obtain the feature map after verifying the digital signature, but cannot obtain the plaintext feature map data; The above process is repeated between units to report the encrypted feature map of their respective areas; Use MPC protocols for secure aggregation, such as homomorphic addition of feature maps; Get the encrypted expression of the global converged feature map and input it to the subsequent layer calculation; The privacy of intermediate data of each participating unit is guaranteed throughout the entire process.

[0064] Through the solution of this embodiment, the privacy security of the image data held by the unit can be protected to the maximum extent while performing joint reasoning.

[0065] In some possible implementations of the present invention, the step of performing relevant computational analysis on the feature data extracted from the image data according to the analysis task requirements includes: Convert the image feature vector into a homomorphic encryption domain for representation; In this step, when performing deep learning image classification tasks, the feature vector representation of the image will be extracted after calculations at each layer; these feature vectors contain sensitive information and need to be privacy protected; homomorphic encryption can achieve calculations in the encrypted domain to protect the data itself. The specific operation is: digitize the original floating-point feature vector into an integer vector of appropriate equation length, use a homomorphic encryption system (such as Paillier) to encrypt each integer element, and obtain a series of encrypted integers as the feature vector expression of the image in the encrypted domain. The operations of the subsequent classification model layers are all performed in this encrypted domain, and the final output is also the predicted category in encrypted form. This method converts features from the plaintext domain to a representation in the homomorphic encryption domain, supporting model calculations while protecting feature privacy.

[0066] Perform encrypted statistical analysis on features (such as calculating expected value, variance and other indicators); In this step, during the training of the deep learning image classification model, it is necessary to perform statistical analysis on the feature data. These statistical analyses include calculating feature averages (expected values), variances and other indicators. Directly using plaintext features for statistics may leak privacy information, so statistical calculations need to be performed in the encrypted domain: the features extracted by each participating unit (such as a hospital) are homomorphically encrypted; encrypted features are used to calculate encrypted expected values, variances and other statistics through homomorphic operations supported by homomorphic encryption. For example, the calculation of expected values ​​uses encrypted summation and division operations; the final statistical indicators are also expressed in encrypted form, and plaintext information is not disclosed to the outside. However, these encrypted statistics can support statistical learning algorithms, such as feature selection. Through homomorphic encryption, the statistical analysis and calculation of features required for model training can be achieved while protecting privacy.

[0067] Use secure comparison protocol to achieve the sorting of feature encryption values; In this step, during the training of the deep learning model, it is necessary to sort the feature data, such as sorting features according to their importance, but the feature values ​​have been converted into encrypted values ​​through homomorphic encryption, and cannot be directly compared in the plaintext domain. Here, the secure comparison protocol supported by MPC is used: each participating unit holds a part of the encrypted feature value; through the MPC protocol, the corresponding two encrypted feature values ​​are compared to determine the size relationship; the comparison will not leak the feature plaintext, and an encrypted comparison result is obtained; all feature values ​​are compared pairwise to obtain a complete encrypted sorting result. This can be provided to subsequent algorithms, such as feature selection and sorting importance; but the external system cannot know the plaintext sorting, which protects the privacy of the features. Through secure sorting, the feature sorting requirements are realized in the encrypted domain, supporting the model statistical learning function.

[0068] Calculate the correlation analysis results of the feature set using the secure aggregation protocol; In this step, during the feature selection process, it is necessary to evaluate the correlation between different features; however, the eigenvalues ​​have been homomorphically encrypted, and the plaintext correlation coefficient cannot be directly calculated; here, the secure aggregation protocol supported by MPC is used for calculation: each participating unit holds some encrypted pairs of eigenvalues; after MPC summation, integration and other operations, the encrypted covariance matrix is ​​aggregated and calculated; then the correlation coefficient is calculated based on the covariance, and the encryption state is maintained. For example, the correlation coefficient is calculated using the ratio formula of covariance and standard deviation; finally, the complete encrypted feature correlation analysis results are obtained. The external system cannot know the plaintext, but it can support the subsequent feature selection algorithm decision. Through secure aggregation, the statistical analysis task is achieved without cracking the feature plaintext.

[0069] Supports statistical analysis and calculation of image metadata through the MPC protocol.

[0070] In this step, the deep learning model not only uses image pixel features, but also metadata features, such as age and gender, for auxiliary classification. However, these metadata also need to protect personal privacy and cannot be shared in plain text. Here, the MPC protocol is used to implement metadata statistical analysis and calculation: each participating unit ciphertext encodes its own metadata; through the MPC protocol, each participating unit shares the encrypted metadata with other nodes; other nodes use operations supported by MPC, such as summation and averaging, to perform statistical calculations, such as calculating the average age, gender ratio distribution, etc. The results are still given in encrypted form; but these encrypted statistical results can provide support for algorithms such as decision trees or logistic regression. That is, the metadata statistical needs are met while protecting the personal privacy data itself. Through MPC collaborative computing, it supports the use of metadata for model training without cracking the privacy content itself.

[0071] In this embodiment, deep data analysis is performed while protecting feature privacy, the statistical distribution law of image features is obtained, feature vectors are sorted for clustering and other processing, feature correlation is calculated to support data mining, image attribute relationships are analyzed to support decision-making, multi-party data analysis services are provided to protect personal privacy, and distributed digital display analysis solutions are provided for medical and other fields, improving data resource utilization while taking into account security rules. The method of this embodiment can efficiently carry out secure joint analysis and give full play to the data advantages of all parties.

[0072] In some possible implementations of the present invention, when performing image data analysis, the following factors should be considered when selecting a suitable feature extraction method: Analyze task types: tasks such as classification, detection, and segmentation require different levels and types of features; Feature interpretability: Low-level features such as color and texture are easy to understand, while high-level features are difficult to directly interpret; Number of features: Different tasks have restrictions on feature dimensions, and important features should be selected for extraction; Computational complexity: In the MPC environment, choose a feature extraction method with high computational efficiency; Homomorphic operability: Choose features that can be directly computed in the encrypted domain, such as SUM statistics; Based on this, the following feature extraction methods can generally be considered: Color histogram: suitable for classifying low-level features; Power spectrum: emphasizes morphological structure and classification and recognition; Texture features: Texture descriptors that support homomorphic operations; Convolutional neural features: The feature dimensions of subsequent layers are small but the recognition ability is strong; Based on geolocation / time metadata: support semantic image understanding; Shape-based moment or contour features: applicable to medical imaging, etc.

[0073] In summary, combining factors such as task type and computational cost, flexibly selecting coarse to fine feature combinations is beneficial to image data mining under multi-party computing.

[0074] In some possible implementations of the present invention, the step of each participating entity performing calculations according to the assigned multi-party secure computing subtask includes: Decompose the global computing model into multiple basic computing units (such as convolutional layers, pooling layers, etc.); In this step, when the deep learning model performs computing tasks, it needs to perform forward and backward propagation according to the hierarchical structure. However, considering that the data and model are distributed in each participating unit and cannot be calculated as a whole at one time, the model computing task needs to be decomposed into each basic computing unit: for example, the convolutional neural network can be decomposed into basic modules such as convolution layer, pooling layer, regular layer, and fully connected layer; each unit completes independent operations, such as convolution and pooling; the operations of each unit are performed in sequence to form an overall computing pipeline; each unit is completed in parallel in multiple units through MPC; and the results of the entire network are calculated step by step. This decomposition helps to perform deep learning model training and prediction tasks step by step in a distributed environment, and exchange information after each unit completes the operation.

[0075] Perform task division and scheduling optimization based on the computing capabilities of each participating entity; In this step, in the distributed deep learning system, the computing resources of each participating entity (such as different hospitals, schools, etc.) may be different. If a large computing task is directly evenly distributed, it may cause some entities to have high loads and tight resources. Therefore, it is necessary to intelligently divide the tasks according to the CPU / GPU / storage configuration of each entity: first conduct a capability assessment to understand the computing capabilities of each entity; refine the original task (such as convolutional layer calculation) into multiple subtask units; prioritize subtasks with large computing volume to the configured entity according to the capability conditions; continue to distribute the remaining tasks according to the load balancing principle; if the real-time load changes, the task redistribution will be adjusted dynamically. This method flexibly schedules tasks according to the actual environment of the entity, fully utilizes the resources of each entity, and improves the overall training efficiency.

[0076] According to the MPC protocol, each basic computing unit is encoded as a subtask and distributed to the corresponding participating entities; In this step, the deep learning model task is decomposed into multiple basic computing units. For example, a convolutional layer can be counted as a unit task. These unit tasks need to be assigned to each participating entity for calculation as needed. The specific steps are as follows: Encode each basic unit to indicate the input, output and other information required for its calculation; Determine the unit task assignment entity based on the previous task scheduling decision; According to the MPC protocol, the unit tasks are packaged and distributed to the corresponding entities in ciphertext form; Each entity is decoded and the assigned unit is calculated locally through the MPC; The calculation results are also returned in ciphertext for the next unit calculation; The whole model calculation is completed collaboratively. Using the MPC protocol for encrypted task distribution can safely and efficiently execute distributed task scheduling.

[0077] Each participating entity completes the corresponding secure computing program according to the type of computing unit and implements the basic operations within the unit (such as multiplication and addition operations, function estimation, etc.); In this step, each participating entity obtains the basic operation unit to be executed through task distribution. Each operation unit represents a basic computing module of the model, such as a convolution layer or a pooling layer. Each participating entity implements the corresponding security calculation program in advance according to different unit types: for convolution units, implement the convolution operation algorithm under homomorphic encryption; for pooling units, implement the maximum / average operation under pooling; for nonlinear units such as ReLU, provide the calculation method of the activation function in the encrypted domain; and so on. Various basic arithmetic and function calculations are supported through security algorithms. When a unit task is obtained, the participating entity directly executes the corresponding security program: completes the basic operation calculations such as multiplication and addition, function estimation, etc. within this unit. In this way, each participating entity can fundamentally unify the interface and efficiently execute the different unit tasks assigned to it.

[0078] Report the processing results to the task coordinator in encrypted form; In this step, each participating entity will obtain a processing result after completing the assigned unit task calculation. Directly reporting the plaintext result may leak intermediate information and destroy data privacy, so the processing result needs to be encrypted and reported: the participating entity encrypts the result according to the homomorphic encryption algorithm used; the ciphertext result is reported to the task coordination center through secure communication; in this way, the task coordination center obtains the encrypted processing results of all parties involved; subsequent work can be performed based on these encrypted results for the next step of calculation or distribution; but no single entity or coordination center can obtain the plaintext intermediate results. Through encrypted reporting, participating entities can safely provide calculation results, and the coordination center can also perform subsequent operations in the encrypted domain, thereby protecting the data privacy in the overall workflow.

[0079] Repeat until all unit tasks are completed.

[0080] In this embodiment, efficient task allocation is performed based on entity capabilities, and the encryption module calculates and implements basic arithmetic and logical operations to ensure that entity resources are fully utilized to achieve the target efficiency. Each entity does not have the ability to access other parties' plaintext or secret data, and distributed multi-party collaborative computing tasks are implemented, thereby using MPC technology to complete end-to-end operations in an orderly and uniform manner.

[0081] In some possible implementations of the present invention, after the calculation is completed, the step of comparing the calculation results between the participating entities in a public verification manner includes: Each participating entity uses its own plaintext test data to independently calculate test indicators (such as accuracy); In this step, after the distributed deep learning has trained the model, the test effect of the model needs to be evaluated. However, the model and data are distributed among different participating entities, and the entities do not send plaintext test data to each other. Therefore, each participating entity uses part of the plaintext test data it holds: input the entity test data into the trained model for prediction; calculate the degree of consistency with the true label, such as accuracy, recall and other indicators; but the test indicator results calculated at this time remain within the scope of the entity and are not leaked; after all entities calculate independently, these indicators can be aggregated through a security protocol to obtain the overall test effect; but any single entity cannot obtain the test data details of other entities. In this way, the overall effect of the model on each entity test data set can be evaluated without leaking the test data.

[0082] Encrypt the calculation results and publish them together with a random digital signature; After the other participating entities verify the signature, they compare the encryption indicators through the MPC protocol to determine whether they are consistent; For indicators that can be linearly encrypted (such as sum and average), verification is done directly in the encryption domain; For nonlinear indicators, (using zk-SNARKs) prove the correctness of the calculation process and results (no intermediate results or plaintext character information are disclosed during the proof process, and finally reach a consensus that the calculation results are correct).

[0083] The solution of this embodiment verifies the correctness of the calculation without leaking secrets, implements an asymmetric verification mechanism for the results, improves the credibility of the results for each entity, meets the integrity requirements of secure multi-party computing, protects the individual data privacy of all parties involved, and effectively verifies and applies distributed computing results.

[0084] In some possible implementations of the present invention, the step of outputting the privacy-processed image analysis calculation results includes: Encoding the calculation results (such as classification labels, detection boxes, etc.) with a security mechanism (such as differential privacy processing); In this step, various calculation results, such as classification labels or object detection boxes, are generated during the prediction of the distributed deep learning model. Directly providing these results may leak certain privacy information, so the results need to be securely encoded: use privacy protection technologies such as differential privacy to achieve encoding; add random perturbations to the plaintext results to obtain a slightly biased but privacy-protected version of the results; such as adding tiny random perturbations to the classification labels, or blurring the detection box coordinates; the encoded results retain the main prediction information and eliminate the risk of direct privacy leakage; and finally provide the requester with the encoded results in the form of privacy-protected results. Through the implementation of differential privacy, the prediction needs of deep learning applications and privacy security considerations are taken into account.

[0085] Privacy techniques such as applying random noise perturbations to the encoded results and combining them into ciphertext; In this step, differential privacy technology is used to perform primary encoding on the prediction results. However, the results of this step may leak some information, so additional privacy technologies are needed to process the results again: apply random noise perturbation (add tiny random noise to the encoded results to generate a new version of the results; this step discretizes the results and further reduces the risk of information leakage), use secure computing technology to compose the results into ciphertext (use obfuscation encryption and other technologies to synthesize the results processed in the first two steps into ciphertext; in this way, the result content cannot be directly obtained by the outside world, maximizing privacy), and retain the usability of the results (although the results are in ciphertext form, they can still support certain subsequent calculations). By combining multiple privacy enhancement technologies, the prediction results are deeply processed, greatly improving the privacy protection capabilities.

[0086] Take partial display to show statistical distribution; When transmitting the encoded results over the network, use transmission encryption to provide anti-tampering protection; Configure validity period control for the encoded results, and automatically delete them after expiration to reduce the risk of leakage; Provide users with an interface to filter the result granularity to balance the effect and privacy requirements.

[0087] In this embodiment, personal information is effectively protected at the result level, computing value is outputted taking into account both computing efficiency and privacy requirements, direct leakage of sensitive features of data cohesion is avoided, a secure and controllable multi-party result sharing service is provided, and decision-making is made using analysis results while protecting privacy, and protective output is performed, taking into account both analysis needs and privacy rules.

[0088] See also Figure 2 ,Another embodiment of the present invention provides a data analysis device based on multi-party secure computing, including: a control module, a determination module, a training module, a prediction module and a processing module; The control module is used to control different data holding units to share and store the privacy-processed image data and metadata of the units using blockchain technology; The determination module is used to determine the image content recognition or analysis task calculation model; The training module is used to perform joint deep learning on the image data stored in the blockchain in an encrypted domain using a multi-party secure computing protocol according to the computing model to train a content recognition model; The prediction module is used to perform content recognition prediction calculation on new image data using the trained model and the multi-party secure computing protocol; The processing module is used for: Perform relevant calculation and analysis on the feature data extracted from the image data according to the analysis task requirements; Each participating entity performs calculations according to the assigned multi-party secure computation subtask; After the calculation is completed, the participating entities compare the calculation results through public verification; Output the privacy-protected image analysis calculation results.

[0089] See also Figure 2 Another embodiment of the present invention provides a data analysis device based on multi-party secure computing, including:

[0090] It should be known that Figure 2 The block diagram of the data analysis device based on multi-party secure computing is for illustration only, and the number of modules shown does not limit the scope of protection of the present invention. The data analysis device based on multi-party secure computing provided in this embodiment can be used to execute the corresponding implementation schemes of the data analysis method based on multi-party secure computing. For the specific implementation process, please refer to the description of the implementation schemes of each method, which will not be repeated here.

[0091] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0092] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0093] In the several embodiments provided in the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of the above-mentioned units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0094] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0095] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0096] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a memory, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the above-mentioned methods of each embodiment of the present application. The aforementioned memory includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or CD-ROM and other media that can store program codes.

[0097] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable memory, and the memory can include: a flash drive, a read-only memory (English: Read-Only Memory, abbreviated as: ROM), a random access memory (English: Random Access Memory, abbreviated as: RAM), a magnetic disk or an optical disk, etc.

[0098] The embodiments of the present application are introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

[0099] Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions without departing from the spirit and scope of the present invention, and can make various changes and modifications, including the combination of the above-mentioned different functions and implementation steps, including software and hardware implementation methods, all of which are within the scope of protection of the present invention.

Claims

1. A data analysis method based on multi-party secure computing, characterized in that: include: Different data holders use blockchain technology to share and store their own privacy-processed image data and metadata; Determine the computational model for image content recognition or analysis tasks; According to the computing model, the multi-party secure computing protocol is used to conduct joint deep learning on the image data stored in the blockchain in an encrypted domain to train a content recognition model; Use the trained model and multi-party secure computing protocol to perform content recognition prediction calculations on new image data; Perform relevant calculation and analysis on the feature data extracted from the image data according to the analysis task requirements; Each participating entity performs calculations according to the assigned multi-party secure computation subtask; After the calculation is completed, the participating entities compare the calculation results through public verification; Output the privacy-protected image analysis calculation results.

2. The data analysis method based on multi-party secure computing according to claim 1 is characterized in that: The steps of different data holding units using blockchain technology to share and store the privacy-processed image data and metadata of the units include: Each data holder performs privacy encryption on the image data and metadata it holds; The privacy-processed image data and metadata are packaged into blocks and digitally signed using the public-private key pair of the data holder; The data holder broadcasts the block to the blockchain network, and other participating nodes verify the integrity and source of the data block through decryption and signature verification; Successfully verified blocks will be added to the blockchain for storage, establishing a database in the form of a distributed ledger.

3. The data analysis method based on multi-party secure computing according to claim 2 is characterized in that: The step of determining the image content recognition or analysis task computing model includes: The data holder selects the appropriate deep learning model architecture based on task requirements; Decompose the model structure into basic operation units that can be calculated in parallel; Use homomorphic encryption or MPC technology to encode the parameters in the model for privacy processing; Use MPC protocol to implement basic unit calculation in the encryption domain; Each participating unit conducts distributed training based on the model part it holds, iteratively updates parameters and reaches consensus; The network structure and parameters are obtained through public verification and serve as the calculation basis for image analysis tasks.

4. The data analysis method based on multi-party secure computing according to claim 3 is characterized in that: The step of performing joint deep learning on the image data stored in the blockchain in an encrypted domain using a multi-party secure computing protocol according to the computing model to train a content recognition model includes: Each data holding unit takes out a portion of its own encrypted image data as a training subset; Use the MPC protocol to encrypt and distribute the weight parameters and biases in the neural network model; Each participating unit calculates the forward propagation results and gradient information corresponding to its own data; Using the secure aggregation algorithm, the loss function gradients of each participating unit are aggregated in the MPC environment; Based on the aggregation results, each participating unit updates the locally held model segment through the MPC protocol; Repeat the forward and backward propagation calculations, and update the parameters in each round of gradient descent; After multiple rounds of iterative training, the final encrypted model of joint learning is obtained.

5. The data analysis method based on multi-party secure computing according to claim 4 is characterized in that: The step of using the trained model and the multi-party secure computing protocol to perform content recognition prediction calculation on new image data includes: The data holder encrypts the new image to be predicted to generate a ciphertext image; According to the MPC protocol, each participating unit uses the model parameters stored in the local distributed state to perform forward propagation calculations; Each participating unit will encrypt and report its normalized data, and other participating units will securely aggregate it; Obtain the encrypted expression of the feature maps of each layer after aggregation, and perform forward calculations on subsequent layers; After executing the entire neural network, the encrypted probability distribution of the predicted class can be obtained; Each participating unit confirms the final prediction results through a public verification algorithm.

6. The data analysis method based on multi-party secure computing according to claim 5 is characterized in that: The step of performing relevant calculation and analysis on the feature data extracted from the image data according to the analysis task requirements includes: Convert the image feature vector into a homomorphic encryption domain for representation; Perform encrypted statistical analysis on features; Use secure comparison protocol to achieve the sorting of feature encryption values; Calculate the correlation analysis results of the feature set using the secure aggregation protocol; Supports statistical analysis and calculation of image metadata through the MPC protocol.

7. The data analysis method based on multi-party secure computing according to claim 6 is characterized in that: The step of each participating entity performing calculation according to the assigned multi-party secure computing subtask includes: Decompose the global computing model into multiple basic computing units; Perform task division and scheduling optimization based on the computing capabilities of each participating entity; According to the MPC protocol, each basic computing unit is encoded as a subtask and distributed to the corresponding participating entities; Each participating entity completes the corresponding security calculation program according to the type of computing unit to realize the basic operations within the unit; Report the processing results to the task coordinator in encrypted form; Repeat until all unit tasks are completed.

8. The data analysis method based on multi-party secure computing according to claim 7 is characterized in that: After the calculation is completed, the step of comparing the calculation results between the participating entities through a public verification method includes: Each participating entity uses its own plaintext test data to independently calculate the test indicators; Encrypt the calculation results and publish them together with a random digital signature; After the other participating entities verify the signature, they compare the encryption indicators through the MPC protocol to determine whether they are consistent; For indicators that can be encrypted linearly, verification is done directly in the encryption domain; For nonlinear indicators, prove the correctness of the calculation process and results.

9. The data analysis method based on multi-party secure computing according to claim 8 is characterized in that: The step of outputting the privacy-processed image analysis calculation result comprises: Encoding the calculation result with a security mechanism; Privacy techniques such as applying random noise perturbations to the encoded results and combining them into ciphertext; Take partial display to show statistical distribution; When transmitting the encoded results over the network, use transmission encryption to provide anti-tampering protection; Configure validity period control for the encoded results, and automatically delete them after expiration to reduce the risk of leakage; Provide users with an interface to filter the result granularity to balance the effect and privacy requirements.

10. A data analysis device based on multi-party secure computing, characterized in that: include: Control module, determination module, training module, prediction module and processing module; The control module is used to control different data holding units to share and store the privacy-processed image data and metadata of the units using blockchain technology; The determination module is used to determine the image content recognition or analysis task calculation model; The training module is used to perform joint deep learning on the image data stored in the blockchain in an encrypted domain using a multi-party secure computing protocol according to the computing model to train a content recognition model; The prediction module is used to perform content recognition prediction calculation on new image data using the trained model and the multi-party secure computing protocol; The processing module is used for: Perform relevant calculation and analysis on the feature data extracted from the image data according to the analysis task requirements; Each participating entity performs calculations according to the assigned multi-party secure computation subtask; After the calculation is completed, the participating entities compare the calculation results through public verification; Output the privacy-protected image analysis calculation results.

Images