CN115238288A - Safety processing method for industrial internet data - Google Patents

Abstract

The invention relates to a security processing method for industrial internet data, belonging to the field of industrial internet data processing. The method includes: the factory and the collaborating party generate a pair of secret keys using the ElGamal encryption algorithm; the contracting party initializes the model parameters and uploads them and the public key on the blockchain; the factory downloads the initial model and the collaborating party's public key from the blockchain , train the model and use the differential privacy algorithm to extract the model parameters; the factory encrypts the model parameters and stores them in IPFS to obtain the hash value; encrypts the hash value and adds it and the factory public key to the blockchain; Retrieve the model parameters in IPFS to train the global model, use SK to encrypt the model parameters and store them in IPFS, encrypt the IPFS hash value, encrypt the SK with the factory public key, and add the result to the blockchain; the factory receives the current global model parameters are updated. The present invention solves the privacy and trust problems of machine learning in the industrial Internet system.

Description

A security processing method for industrial Internet data

technical field

The invention belongs to the technical field of industrial internet data security processing, and relates to a security processing method for industrial internet data.

Background technique

In recent years, with the large-scale deployment and application of intelligent sensing devices, massive amounts of data have been generated. These data contain information in various fields of production and life, and have become an important production factor. As a product of the deep integration of a new generation of information technology and manufacturing, the Industrial Internet, through the comprehensive interconnection of people, machines, and things, builds a new industrial manufacturing and service system that is fully connected with all elements, the entire industrial chain, and the entire value chain. The way to realize digital transformation is the key force to realize the transformation of old and new kinetic energy. The Industrial Internet is in a new stage of exploding data scale, sharply increasing computing power, and improving algorithm performance. Intelligent technologies such as deep learning, reinforcement learning, and federated learning have become important supports for improving the comprehensive service performance of the Industrial Internet.

The confidentiality of industrial production data is extremely high. At present, the commonly used centralized model training method has the problem of data privacy leakage. It is difficult to take into account the dual needs of data sharing and privacy protection. Not only does a new privacy-preserving computing paradigm need to complete multi-party joint modeling, but also Flexible and intelligent adaptation methods are required to ensure the efficient operation of new computing paradigms. And security is the guarantee of the Industrial Internet. In the future, the construction of the emerging infrastructure of the Industrial Internet will continue to advance to a wider, deeper and higher level. In the face of new technologies and new situations of security protection, the industrial Internet security ecosystem will continue to face new challenges. Therefore, it is urgent to deal with industrial Internet data security methods to improve the ability to deal with security risks and promote the prosperity and development of the Industrial Internet.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide a security processing method for industrial Internet data, aiming at the problem that it is difficult to take into account the dual requirements of industrial data sharing and privacy protection, by combining federated learning, differential privacy, ElGamal encryption, etc. The industrial Internet can better empower the industrial Internet and promote industrial upgrading.

To achieve the above object, the present invention provides the following technical solutions:

A security processing method for industrial Internet data, which specifically includes the following steps:

S1: The factory and the collaborating party generate a pair of secret keys; then the collaborating party initializes the model parameter w ₀ and uploads the model parameter w ₀ and its public key to the blockchain;

S2: Each factory downloads the initialization model parameter w ₀ and the public key of the collaborator from the blockchain, and uses a specific number of local data samples to train a deep neural network model; after the training is completed, the model is processed with a differential privacy algorithm to generate Local differential privacy machine learning model, after which the factory extracts model parameters;

S3: The factory uses the collaborator's public key to encrypt the model parameters, then stores the encrypted model parameters in IPFS, and obtains a unique IPFS hash value; then also uses the collaborator's public key to encrypt the IPFS hash value, with the help of It is packaged with the factory public key and added to the blockchain in the smart contract; the factory notifies the collaborator that the current round of local model training is completed;

Store encrypted model parameters in IPFS, a peer-to-peer protocol for file storage that uses content-based addressing rather than location-based. Files in the IPFS network are assigned a hash value, which is similar to a "fingerprint", which is calculated from the content of the file.

S4: The collaborator decrypts with the private key, encrypts the IPFS hash value with its public key, retrieves the corresponding encryption model parameters in IPFS with the IPFS hash value, and then trains the global model and extracts the model parameters after decryption;

S5: Use SK (symmetric key produced by the collaborator) to encrypt the model parameters and store them in IPFS, and then will be given an IPFS hash value, use SK to encrypt the IPFS hash value, and use the public key of each factory Encrypt SK separately, add the encryption result to the blockchain (this multi-layer encryption method ensures the security of data), and notify the factory that this round of aggregation is over;

S6: After receiving the notification from the collaborator that the current federation cycle is over, the factory retrieves the encrypted global model parameters, uses its private key to decrypt to obtain SK, uses SK to decrypt the encrypted IPFS hash value, and uses the IPFS hash value to retrieve model parameters; Use the updated model parameters for local training, start the next round of joint training, and repeat all steps for a predefined number of joint rounds.

Further, in step S1, the factory and the collaborating party generate a pair of secret keys with the ElGamal encryption algorithm; the ElGamal encryption algorithm is an asymmetric encryption algorithm based on Diffie-Hellman key exchange, and the principle based on the EIGamal encryption algorithm is : It is difficult to solve the discrete logarithm, and its inverse operation can be effectively calculated by the method of square multiplication. In the corresponding group G, the exponential function is a one-way function.

的批量序号b，计算批梯度g_k ^(b)，本地更新模型参数：

其中η表示学习率，w_t表示当前的模型参数，D_k表示第k个协作方所拥有的数据集，M表示指定的客户更新是使用的mini-batch的大小；Further, step S2 specifically includes: the factory uses a specific number of local data samples to train a deep neural network; obtains the latest model parameters from the server, ranging from 1 to the number of batches

The batch sequence number b is calculated, the batch gradient g _k ^(b) is calculated, and the model parameters are updated locally:

where η represents the learning rate, _wt represents the current model parameters, D _k represents the data set owned by the k-th collaborator, and M represents the size of the mini-batch used for the specified customer update;

来训练模型参数；在SGD的每一步，计算梯度，对于采样子集，剪切每个梯度的范数l₂，计算平均值，为保护隐私添加噪声，向这个平均噪声梯度的相反方向迈出一步进行梯度下降反向传播完成训练最后输出模型。During the training process, differential privacy-based neural network training is implemented in the SGD calculation by minimizing the empirical loss function

to train the model parameters; at each step of SGD, compute the gradient, for a sampled subset, clip the norm l ₂ of each gradient, compute the average, add noise for privacy, and step in the opposite direction of this average noise gradient One step of gradient descent backpropagation completes the training and finally outputs the model.

Further, step S4 specifically includes: assuming that there are K participants (ie factories) in a federated learning system, D _k represents the data set owned by the kth participant, and P _k represents the index of the data point located in participant k Set; let n _k represent the cardinality of P _k , assuming that the kth participant has n _k data points, when there are a total of K participants, the collaborating party aggregates the received model parameters, that is, the received model Weighted average of parameters:

Update model parameters:

表示在给定的模型参数w上对样本(x_i,y_i)进行预测所得到的损失结果，x_i和y_i分别表示第i个训练数据点及其相关的标签。η表示学习率，n表示训练数据的数量；协作方检查损失函数是否收敛或者是否达到最大训练轮次；若是，则协作方给各参与方发信号，使其全部停止模型训练。of which, of which

represents the loss result obtained by predicting the sample ( _xi , y _i ) on the given model parameter w, where x _i and y _i represent the ith training data point and its associated label, respectively. η represents the learning rate, and n represents the number of training data; the collaborating party checks whether the loss function converges or whether the maximum training round is reached; if so, the collaborating party sends a signal to each participant to stop model training.

Further, in step S6, the updated model parameters of the factory are encrypted, stored and uploaded.

The beneficial effect of the present invention is that the present invention enhances the privacy and credibility of industrial Internet data by combining differential privacy, federated learning, block chain and smart contracts. It improves the ability to deal with security risks and promotes the prosperity and development of the Industrial Internet.

Other advantages, objects, and features of the present invention will be set forth in the description that follows, and will be apparent to those skilled in the art based on a study of the following, to the extent that is taught in the practice of the present invention. The objectives and other advantages of the present invention may be realized and attained by the following description.

Description of drawings

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be preferably described in detail below with reference to the accompanying drawings, wherein:

Fig. 1 is the framework diagram of the industrial Internet data security processing method involved in the present invention;

Fig. 2 is the factory data processing process flow chart of the industrial Internet data security processing method of the present invention;

FIG. 3 is a flow chart of the data processing process of the collaborator of the industrial Internet data security processing method of the present invention.

Detailed ways

The embodiments of the present invention are described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments are only used to illustrate the basic idea of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

Please refer to Figures 1 to 3. Figure 1 shows the architecture of an industrial Internet data security processing method, Figure 2 shows the factory data processing process of the industrial Internet data security processing method, and Figure 3 shows the industrial Internet data The collaborating party's data processing process for the secure processing method.

This method solves the privacy and trust issues of machine learning in industrial Internet systems, including factories, collaborators, IPFS, blockchain and other entities and data transmission between entities. Specifically includes the following steps:

Step 1: The factory and the collaborator first generate a pair of secret keys with the ElGamal encryption algorithm. The ElGamal encryption algorithm is an asymmetric encryption algorithm based on Diffie-Hellman key exchange. The principle of the EIGamal encryption algorithm is to solve the discrete pair Numbers are difficult, and their inverse operations can be efficiently calculated using the square multiplication method. In the corresponding group G, the exponential function is a one-way function.

Step 2: Each factory downloads the initialization model w ₀ and the public key of the collaborating party from the blockchain, and trains a deep neural network with a specific number of local data samples. After the training is completed, the differential privacy algorithm is used to process the model to generate a local differential privacy machine learning model.

的批量序号b，计算批梯度g_k ^(b)，本地更新模型参数：

η表示学习率。The factory trains a deep neural network using a specific number of local data samples. Get the latest model parameters from the server, from 1 to the number of batches

The batch sequence number b is calculated, the batch gradient g _k ^(b) is calculated, and the model parameters are updated locally:

η represents the learning rate.

来训练模型参数。在SGD的每一步，计算梯度，对于采样子集，剪切每个梯度的范数l₂，计算平均值，为保护隐私添加噪声，向这个平均噪声梯度的相反方向迈出一步进行梯度下降反向传播完成训练最后输出模型。During the training process, differential privacy-based neural network training is implemented in the SGD calculation by minimizing the empirical loss function

to train the model parameters. At each step of SGD, compute the gradient, for the sampled subset, clip the norm l ₂ of each gradient, compute the average, add noise for privacy, take a step in the opposite direction of this average noise gradient and perform the gradient descent inverse The training is completed to the final output model of the propagation.

Federated learning is able to train deep learning models with the help of a central server while keeping the training data distributed across clients. The client only needs to submit its local gradient to the cloud server during model training, which avoids the risk of user data privacy leakage to a certain extent. Differential privacy is an efficient privacy protection technology, which can quantify the degree of data privacy protection, set an appropriate privacy budget, and achieve a good balance between data availability and privacy protection. Differential privacy is often used to protect the privacy of training data of artificial intelligence models, and can provide privacy guarantees for federated learning.

On the premise that the user data is not local, a common model is established through parameter exchange and optimization under the encryption mechanism or the perturbation mechanism. The performance of this shared model is close to that of a model trained by aggregating data from all parties into one piece. The data joint modeling scheme does not reveal user privacy and conforms to the principle of data security protection.

Step 3: The factory encrypts the model parameters with the public key of the collaborating party, then stores the encrypted model parameters in IPFS, and obtains a unique IPFS hash value. Then, the IPFS hash value is also encrypted with the public key of the contracting party, and it is packaged with the factory public key and added to the blockchain with the help of a smart contract. The factory notifies the collaborator that the current round of local model training is complete.

Store encrypted model parameters in IPFS, a peer-to-peer protocol for file storage that uses content-based addressing rather than location-based. Files in the IPFS network are assigned a hash value, which is similar to a "fingerprint", which is calculated from the content of the file.

Step 4: The collaborating party uses private key decryption to encrypt the IPFS hash value through its public key, and retrieves the corresponding encryption model parameters in IPFS with the IPFS hash value. After decryption, the global model is trained and the model parameters are extracted.

The collaborating party first decrypts to obtain the IPFS hash value, retrieves the corresponding encryption model parameters with the IPFS hash value, and continues to use its private key to decrypt to obtain the parameters of the local model update.

Suppose there are K parties in a federated learning system, _Dk denotes the dataset owned by the Kth party, and _Pk denotes the index set of data points located at customer k. Let n _k denote the cardinality of P _k , assuming that the kth participant has n _k data points, and there are K participants in total

The collaborators aggregate the received model parameters, that is, perform a weighted average of the received model parameters:

where η represents the learning rate, and the collaborator checks whether the loss function converges or whether the maximum training epoch is reached. If so, the cooperating party sends a signal to all the participating parties to stop model training.

Step 5: Use SK (a symmetric key produced by the collaborator) to encrypt the model parameters and store them in IPFS, and then will be given an IPFS hash value, use SK to encrypt the IPFS hash value, and use the factory's public key. Key encryption SK, the encryption result is added to the blockchain, and the factory is notified that this round of aggregation is over.

Step 6: The factory uses the updated model parameters for local training, first retrieves the encrypted global model parameters, decrypts with the private key to obtain SK, decrypts and encrypts the IPFS hash value with the SK, retrieves the model parameters with the IPFS hash value, and then uses the IPFS hash value to retrieve the model parameters. Parameterize the rear wing in the same way as in step 2 and perform model training. After that, model parameters are encrypted, stored and uploaded.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent replacements, without departing from the spirit and scope of the technical solution, should all be included in the scope of the claims of the present invention.

Claims (5)

1. a security processing method for industrial internet data, it is characterised in that the method specifically comprises the following steps: S1: The factory and the collaborating party generate a pair of secret keys; then the collaborating party initializes the model parameter w ₀ and uploads the model parameter w ₀ and its public key to the blockchain; S2: Each factory downloads the initialization model parameter w ₀ and the public key of the collaborator from the blockchain, and uses a specific number of local data samples to train a deep neural network model; after the training is completed, the model is processed with a differential privacy algorithm to generate Local differential privacy machine learning model, after which the factory extracts model parameters; S3: The factory uses the collaborator's public key to encrypt the model parameters, then stores the encrypted model parameters in IPFS, and obtains a unique IPFS hash value; then also uses the collaborator's public key to encrypt the IPFS hash value, with the help of It is packaged with the factory public key and added to the blockchain in the smart contract; the factory notifies the collaborator that the current round of local model training is completed; S4: The collaborator decrypts with the private key, encrypts the IPFS hash value with its public key, retrieves the corresponding encryption model parameters in IPFS with the IPFS hash value, and then trains the global model and extracts the model parameters after decryption; S5: Use the symmetric key SK produced by the collaborator to encrypt the model parameters and store them in IPFS, and then will be given an IPFS hash value, use SK to encrypt the IPFS hash value, and use the public key of each factory to encrypt SK, add the encrypted result to the blockchain and notify the factory that this round of aggregation is over; S6: After receiving the notification from the collaborator that the current federation cycle is over, the factory retrieves the encrypted global model parameters, uses its private key to decrypt to obtain SK, uses SK to decrypt the encrypted IPFS hash value, and uses the IPFS hash value to retrieve model parameters; Use the updated model parameters for local training, start the next round of joint training, and repeat all steps for a predefined number of joint rounds. 2. the security processing method of industrial internet data according to claim 1, is characterized in that, in step S1, factory and collaborating party generate a pair of secret keys with ElGamal encryption algorithm; Described ElGamal encryption algorithm is a Diffie-based encryption algorithm. Asymmetric encryption algorithm for Hermann key exchange. 3. The method for safe processing of industrial internet data according to claim 1, wherein step S2 specifically comprises: the factory uses a specific number of local data samples to train a deep neural network; obtain the latest model parameters from the server, from 1 to batch quantity

The batch sequence number b is calculated, the batch gradient g _k ^(b) is calculated, and the model parameters are updated locally:

where η represents the learning rate, _wt represents the current model parameters, D _k represents the data set owned by the k-th collaborator, and M represents the size of the mini-batch used for the specified customer update; During the training process, differential privacy-based neural network training is implemented in the SGD calculation by minimizing the empirical loss function

to train the model parameters; at each step of SGD, the gradients are calculated, and for a sampled subset, the norm of each gradient is clipped

Calculate the average, take a step in the opposite direction of the average noise gradient and perform gradient descent backpropagation to complete the training and finally output the model. 4. The security processing method for industrial Internet data according to claim 3, wherein step S4 specifically comprises: assuming that there are K participants in a federated learning system, D _k represents the data owned by the kth participant. Data set, P _k represents the index set of data points located in participant k; let n _k represent the cardinality of P _k , assuming that the kth participant has n _k data points, when there are a total of K participants, the collaborating party Aggregate the received model parameters, that is, perform a weighted average of the received model parameters:

Update model parameters:

of which, of which

Represents the loss result obtained by predicting the sample ( _xi , y _i ) on the given model parameter w, x _i and y _i represent the i-th training data point and its associated label respectively; η represents the learning rate, n represents the number of training data; the collaborating party checks whether the loss function converges or whether the maximum training round is reached; if so, the collaborating party sends a signal to each participant to stop model training. 5 . The method for safely processing industrial Internet data according to claim 1 , wherein, in step S6 , the updated model parameters of the factory are encrypted, stored and uploaded. 6 .

Images