Internet of Things security method based on multidirectional projection feature enhancement and negative gain aggregation

Abstract

The invention relates to an Internet of Things security method based on multidirectional projection feature enhancement and negative gain aggregation, comprising the following steps: S1, a cloud server establishes a communication link with an edge client, and deploys a negative gain aggregation hyper-parameter for the edge client; s2, the edge client receives the current global model from the cloud server, carries out local model training on the current global model through feature enhancement, and uploads the local model after training is completed; s3, the edge client calculates a parameter adaptation degree and transmits the parameter adaptation degree to the cloud server; s4, the cloud server aggregates the local model according to the parameter adaptation degree and the negative gain aggregation hyper-parameter to obtain a new global model; and S5, receiving the new global model by the edge client, and returning to S2 to carry out next round of local training on the new global model through feature enhancement. According to the method, the overall efficiency of the system is improved while privacy is protected through feature enhancement and negative gain aggregation technologies.

Description

Technical Field

[0001] The present invention relates to the field of Internet of Things security technology, and in particular to an Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation. Background Art

[0002] IoT devices have greatly promoted the development of artificial intelligence, but they rely on users' sensitive data to perform tasks, which has aroused public concerns about the privacy and security of personal data. Limited by the architectural design of IoT devices themselves, it is difficult to deploy security firewalls with large computing loads. Therefore, federated learning (FL), a technology that focuses on user security and privacy protection, has begun to be adopted. Federated learning (FL) can improve the performance of AI models through a distributed training mechanism without leaking original data.

[0003] However, there are many types of IoT devices, and the data generated is heterogeneous. If we only rely on FL, the performance of AI models will often drop sharply. There is an urgent need to develop new security solutions without sacrificing the security of IoT devices and the privacy of user data. We need to find new methods to supplement federated learning (FL) to solve the heterogeneity problem of IoT. Summary of the invention

[0004] The purpose of the present invention is to provide an Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation. Through feature enhancement and negative gain aggregation technology, the performance of AI models in heterogeneous data environments is optimized, thereby improving the overall performance of the system while protecting privacy.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] The IoT security method based on multi-directional projection feature enhancement and negative gain aggregation includes:

[0007] S1. The cloud server establishes a communication link with the edge client and deploys negative gain aggregation hyperparameters for the edge client;

[0008] S2, the edge client receives the current global model from the cloud server, the edge client performs local model training on the current global model through feature enhancement, and uploads the local model after the training is completed;

[0009] S3, the edge client calculates the parameter adaptation and transmits it to the cloud server;

[0010] S4, the cloud server aggregates the local model according to the parameter fitness and the negative gain aggregation hyperparameter to obtain a new global model;

[0011] S5. The edge client receives the new global model and returns to S2 to perform the next round of local training on the new global model through feature enhancement.

[0012] Optionally, the edge client in S2 performs local model training on the current global model through feature enhancement, including:

[0013] The edge client divides the current global model into a main body and a head, and uses a multi-directional projection feature enhancement method to enhance the features of the main body after training, wherein the current global model is a CNN model, the main body is the first two layers of the CNN model, and the head is a fully connected layer and an output layer;

[0014] The head is trained according to the enhanced features, and the local model is uploaded after the training is completed.

[0015] Optionally, the multi-directional projection feature enhancement method is used to enhance the features after subject training, including:

[0016] The output features of the subject are obtained, and the output features of the subject are used as input features of a multi-directional projection feature enhancement method. The multi-directional projection feature enhancement method performs a convolution operation and outputs intermediate data, wherein the intermediate data is F q 、F k and F v ;

[0017] The intermediate data is enhanced by using a multi-directional projection window and an attention mechanism to obtain a first output feature, and a convolution operation is performed on the intermediate data by using a convolution to obtain a second output feature;

[0018] According to the first output feature and the second output feature, a feature after feature enhancement is obtained.

[0019] Optionally, the intermediate data is enhanced by using a multi-directional projection window and attention to obtain a first output feature, which includes:

[0020] Use multi-directional projection windows to project the intermediate data F k and F v Perform feature enhancement to obtain F k and F v Generated intermediate variables;

[0021] F k and F v The generated intermediate variables are similar to F q Through the attention mechanism, the first output feature is output.

[0022] Optionally, a multi-directional projection window is used to project the intermediate data F k and F v The feature enhancement is:

[0023] {F k_m ,F v_m}=R(σf 7×7 (pad(F k ,F v ))+b)

[0024] Where pad(·) represents a padding operation of size P, and f k×k (·) represents the projection window operation of size k×k, R(·) represents the Reshape operation, and F k_m and F v_m F k and F v The generated intermediate variable, σ is the activation function, and b is the bias term.

[0025] Optionally, the parameter adaptability in S3 is obtained through a global model according to the data resources of the edge client.

[0026] Optionally, the cloud server in S4 aggregates the local model according to the parameter fitness and the negative gain aggregation hyperparameter, including:

[0027] According to the parameter fitness and the negative gain aggregation hyperparameter, obtaining a weight redistribution coefficient of model aggregation;

[0028] The weight redistribution coefficient of the model aggregation is:

[0029]

[0030] Among them, α i represents the aggregation weight redistribution coefficient of client i, S i is the fitness of client i, β is the hyperparameter of the negative gain aggregation method, N represents the number of all clients participating in the training, S k is the fitness of client k;

[0031] The local models are aggregated according to the weight redistribution coefficients to obtain the new global model.

[0032] Optionally, the new global model is:

[0033]

[0034] in, represents the local model of the i-th client, w t represents the global model, represents the total number of data of edge client i, The total number of data representing all clients.

[0035] The beneficial effects of the present invention are as follows: the method of the present invention can maintain the security performance of the system and improve the performance of the IoT system and devices in a heterogeneous environment where data is not shared. This method mainly targets the heterogeneity problem of IoT devices and optimizes the performance of AI models in a heterogeneous data environment through feature enhancement and negative gain aggregation technology, thereby improving the overall performance of the system while protecting privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 It is a flow chart of an Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation according to an embodiment of the present invention;

[0038] Figure 2 A data processing diagram for multi-directional projection feature enhancement according to an embodiment of the present invention;

[0039] Figure 3 This is a comparison result of the accuracy of the method of the embodiment of the present invention and the baseline method on the cifar10 dataset;

[0040] Figure 4 This is a comparison chart of the accuracy of the method according to the embodiment of the present invention and the baseline method on the cifar100 dataset. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] This embodiment provides an IoT security method based on multi-directional projection feature enhancement and negative gain aggregation, including:

[0044] S1. The cloud server establishes a communication link with the edge client and deploys negative gain aggregation hyperparameters for the edge client;

[0045] S2, the edge client receives the current global model from the cloud server, and performs local model training on the current global model through feature enhancement. After the training is completed, the local model is uploaded;

[0046] S3, the edge client calculates the parameter adaptation and transmits it to the cloud server;

[0047] S4, the cloud server aggregates the local models according to the parameter fitness and negative gain aggregation hyperparameters to obtain a new global model;

[0048] S5. The edge client receives the new global model and returns to S2 to perform the next round of local training on the new global model through feature enhancement.

[0049] Furthermore, the edge client in S2 performs local model training on the current global model through feature enhancement, including:

[0050] The edge client divides the current global model into a main body and a head, and uses a multi-directional projection feature enhancement method to enhance the features of the main body after training. The current global model is a CNN model, the main body is the first two layers of the CNN model, and the head is the fully connected layer and the output layer;

[0051] The head is trained based on the enhanced features, and the local model is uploaded after training.

[0052] Specifically, the main body generally extracts data features, and the head is mainly responsible for high-level abstract feature judgment to complete subsequent classification or regression tasks. The multi-directional projection feature enhancement method can be regarded as a combination of two structures through parameters and , one of which is a combination of multi-directional projection and attention, and the other is a convolution scheme to avoid the poor effect of early multi-directional projection.

[0053] Furthermore, the multi-directional projection feature enhancement method is used to enhance the features after subject training, including:

[0054] The output features of the subject are obtained, and the output features of the subject are used as the input features of the multi-directional projection feature enhancement method. The multi-directional projection feature enhancement method outputs intermediate data, where the intermediate data is F q 、F k and F v ;

[0055] The intermediate data is enhanced by using a multi-directional projection window and attention to obtain a first output feature, and the intermediate data is convolved by using convolution to obtain a second output feature;

[0056] According to the first output feature and the second output feature, a feature after feature enhancement is obtained.

[0057] Specifically, the multi-directional projection first obtains three intermediate data F through the following formula q 、F k and F v :

[0058] {F k ,F q ,F v = R(Cov 1×1 (UF in +b))

[0059] Cov i×i Represents the convolution operation with kernel i, where i is 1. R(·) represents the Reshape operation, which is used to adjust the output size. b is the bias term and U is the weight matrix.

[0060] After that, the multi-directional projection uses the multi-directional projection window to enhance the intermediate data, and then uses attention to refocus the projection window. The output of this path is defined as F s That is the first output feature.

[0061] Design a new convolution structure to get the output F c That is the second output feature.

[0062] The output F of the multi-directional projection feature enhancement method out The expression is:

[0063] F out =γF c +δF s

[0064] Among them, γ and δ are the hyperparameters of multi-directional projection, which are used to dynamically adjust the proportion of dual-channel output. c is the output of the convolution, F s This is the output of one multi-directional projection.

[0065] Furthermore, the intermediate data is enhanced using multi-directional projection windows and attention, and the first output features are obtained including:

[0066] Use multi-directional projection windows to project the intermediate data F k and F v Perform feature enhancement to obtain F k and F v Generated intermediate variables;

[0067] F k and F v The generated intermediate variables are similar to F q Through attention, the first output feature is output.

[0068] Furthermore, the multi-directional projection window is used to project the intermediate data Fk and F v The feature enhancement is:

[0069] {F k_m ,F v_m}=R(σf 7×7 (pad(F k ,F v ))+b)

[0070] Where pad(·) represents a padding operation of size P, and f k×k (·) represents a projection window operation of size k×k. In this embodiment, the value of k is 7. R(·) represents a Reshape operation. k_m and F v_m F k and F v The generated intermediate variable, σ is the activation function, and b is the bias term.

[0071] Furthermore, the parameter adaptation in S3 is obtained through the current global model according to the data resources of each edge client.

[0072] Furthermore, the cloud server in S4 aggregates the local model according to the parameter fitness and negative gain aggregation hyperparameters, including:

[0073] According to the parameter fitness and negative gain aggregation hyperparameters, obtain the weight redistribution coefficient of model aggregation;

[0074] The weight redistribution coefficient of model aggregation is:

[0075]

[0076] Among them, α i represents the aggregation weight redistribution coefficient of client i, S i is the fitness of client i, β is the hyperparameter of the negative gain aggregation method, and N is the number of all clients participating in the training;

[0077] The local models are aggregated according to the weight redistribution coefficients to obtain a new global model.

[0078] Furthermore, the new global model is:

[0079]

[0080] in, represents the local model of the i-th client, w t represents the global model, represents the total number of data of edge client i, The total number of data representing all clients.

[0081] like Figure 1 As shown, the IoT security method based on multi-directional projection feature enhancement and negative gain aggregation is further described as follows:

[0082] Step 1: During the system initialization phase, the cloud server establishes a communication link with each edge client and deploys the negative gain aggregation method. At the same time, each edge client prepares the relevant parameters of the training model and its own data resources.

[0083] Among them, the hyperparameter β of the negative gain aggregation method is set to 1, and some parameters related to model training are learning rate η, sampling rate a, and training batch size B.

[0084] Step 2: Each edge client C i The initial model CNN received from the cloud server is divided into the head and the body, and the multi-directional projection feature enhancement method is used in the two and the initialization related parameters are given. The body is the first two layers of the CNN model, that is, the first convolution layer plus the first pooling layer is the first layer, and the second convolution layer plus the second pooling layer is the second layer. The head is the remaining part, that is, the fully connected layer and the output layer.

[0085] Among them, the main body generally extracts data features, and the head is mainly responsible for high-level abstract feature judgment to complete subsequent classification or regression tasks. The values ​​of the initialization parameters γ and δ related to the multi-directional projection feature enhancement method are both set to 0.5.

[0086] Step 3: Edge Client C i Perform local model training. The edge client receives the initial model CNN and related parameters from the cloud server, and uses the multi-directional projection feature enhancement method to enhance the features after the main body training, and then continues to train the head. After the training is completed, upload the local model

[0087] Specifically, the model CNN is divided into two parts and trained sequentially: first, the main body is started to obtain the abstract features of the image, and the features are stored in the last layer of the main body. Then, the multi-directional projection feature enhancement method is used to obtain more effective image features, and then the head training is carried out.

[0088] Among them, the multi-directional projection feature enhancement method can be regarded as a combination of two structures through parameters γ and δ, one of which is a combination of multi-directional projection and attention, and the other is a convolution scheme to avoid the poor effect of early multi-directional projection. Figure 2 shown.

[0089] The output after the subject training is used as the input of the multi-directional projection feature enhancement method, and the input is defined as F in , the output is defined as F outThe multi-directional projection first obtains three data F through the following formula q 、F k and F v .

[0090] {F k ,F q ,F v = R(Cov 1×1 (UF in +b))

[0091] Cov i×i Represents the convolution operation with kernel i, where i is 1. R(·) represents the Reshape operation, which is used to adjust the output size. b is the bias term and U is the weight matrix.

[0092] After that, the multi-directional projection uses the multi-directional projection window to project the intermediate data F k and F v After enhancement, the projection window is weighted and refocused using attention. The output of this path is defined as F s The specific mechanism is shown in the following formula:

[0093] {F k_m ,F v_m}=R(σf 7×7 (pad(F k ,F v ))+b)

[0094] Where pad(·) represents a padding operation of size P. k×k (·) represents a projection window operation of size k×k. R(·) represents a Reshape operation. k_m and F v_m F k and F v The generated intermediate variables, subsequent and F q Generate output F through attention together s , σ is the activation function (tanh or ReLU), and b is the bias term.

[0095] Design a new convolution structure to get the output F c , as a supplement to the multi-directional projection window. The new convolution structure is shown in the following formula:

[0096] F c =Cov 1×1 (W f [h t ,x t ]R(F k ,F q ,F v )+b f )

[0097] Cov i×i represents the convolution operation with kernel i, R(·) represents the Reshape operation, and F q 、F k and F v Yes Input F in The output obtained by the formula is W f and b f is the weight matrix and bias vector of the forget gate, h t is the current hidden state, x t is the time step input.

[0098] Finally, the output F of the designed multi-directional projection feature enhancement method is out The expression is:

[0099] F out =γF c +δF s

[0100] Among them, γ and δ are the hyperparameters of multi-directional projection, which are used to dynamically adjust the proportion of dual-channel output. c is the output of the convolution, F s This is the output of one multi-directional projection.

[0101] Step 4: The edge client calculates the parameter fitness S required for the designed negative gain aggregation i , and transmitted to the cloud server.

[0102] Among them, the fitness S i The initial global model received by the edge client and the edge client local data resources It is calculated that the local data resources are first input into the initial global model as a test set to obtain the prediction accuracy, and the obtained accuracy is used as the adaptation of the edge client, and S i ∈[0,1].

[0103] Step 5: The cloud server uses the designed negative gain aggregation method to aggregate model parameters. The cloud server is based on the fitness S of all edge clients. i and local models The global model w is obtained by the negative gain aggregation method t .

[0104] Among them, the local model As the generated global model w t Raw materials, suitability S i The weight redistribution coefficient α of the model aggregation will be generated through the newly designed formula i, and adjust the model aggregation accordingly. The cloud server aggregates data through the proposed negative gain aggregation formula, as follows:

[0105]

[0106]

[0107] in, represents the local model of the i-th client, w t represents the global model, represents the total number of data of edge client i, Represents the total number of data from all clients. S i is the fitness of client i, and β is the hyperparameter of the negative gain aggregation method to prevent the aggregation weight from exceeding the range. i represents the aggregate weight redistribution coefficient of client i, N represents the number of all clients participating in the training, S k is the fitness of client k.

[0108] Step 6: The edge client updates the local model. The edge client receives the global model w aggregated by the negative gain aggregation method in the previous round from the cloud server. t , a multi-directional projection window method is used for the next round of local training and global communication of the model.

[0109] After T rounds of communication between the cloud server and the edge client, an image classification detection model based on the deep learning model can be finally obtained. i In each communication round, local training is performed once. Each training first accepts the global model w of the previous round. t As the initial parameter setting for training, generate a local model Compute global models and data resources simultaneously The fitness S i After local training is completed, the local model and fitness S i For the cloud server, it receives the relevant models and parameters of all edge clients in each round of communication and obtains the global model w through the negative gain aggregation method. t , and then distributed to each edge client in the next round of communication for the next round of training.

[0110] Experimental proof examples:

[0111] In order to comprehensively evaluate the superior performance of the innovative method proposed by the present invention in practical applications, this embodiment selects special image datasets CIFAR-10 and CIFAR-100, and compares the accuracy of the proposed method with the traditional federated learning method. The experiment was carried out on the Ubuntu 18.04.3 LTS platform, using Intel CORE I5 12600KFCPU and NVIDIA GeForce RTX 4060TI GPU (32GB RAM). The federated learning model of this embodiment is run in an environment built using Python3.9.19 programming language and PyTorch 2.4.0 framework. In the specific implementation, for the two different datasets of CIFAR-10 and CIFAR-100, operations are performed separately, and each dataset is divided into a training set and a test set. Specifically, the training set of CIFAR-10 contains 50,000 images, and the test set contains 10,000 images; similarly, the training set of CIFAR-100 also contains 50,000 images, and the test set contains 10,000 images.

[0112] In a specific implementation, the method of the present invention is experimentally compared with the classic federated learning methods FedAvg and FedGen models. Figure 3 As shown in Figure 2, the accuracy of CIFAR-10 is improved by 3.18%. Figure 4 As shown in the figure, the accuracy of CIFAR-100 is improved by 6.64%, while achieving privacy protection for the Internet of Things.

[0113] In summary, compared with the classic federated learning method, the method proposed in the present invention can not only achieve IoT privacy security but also has higher accuracy in the actual production environment of the IoT network.

[0114] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. An IoT security method based on multi-directional projection feature enhancement and negative gain aggregation, characterized in that: include: S1. The cloud server establishes a communication link with the edge client and deploys negative gain aggregation hyperparameters for the edge client; S2, the edge client receives the current global model from the cloud server, the edge client performs local model training on the current global model through feature enhancement, and uploads the local model after the training is completed; S3, the edge client calculates the parameter adaptation and transmits it to the cloud server; S4, the cloud server aggregates the local model according to the parameter fitness and the negative gain aggregation hyperparameter to obtain a new global model; S5. The edge client receives the new global model and returns to S2 to perform the next round of local training on the new global model through feature enhancement.

2. The Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation according to claim 1 is characterized in that: The edge client in S2 performs local model training on the current global model through feature enhancement, including: The edge client divides the current global model into a main body and a head, and uses a multi-directional projection feature enhancement method to enhance the features of the main body after training, wherein the current global model is a CNN model, the main body is the first two layers of the CNN model, and the head is a fully connected layer and an output layer; The head is trained according to the enhanced features, and the local model is uploaded after the training is completed.

3. The Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation according to claim 2 is characterized in that: The multi-directional projection feature enhancement method is used to enhance the features after subject training, including: The output features of the subject are obtained, and the output features of the subject are used as input features of a multi-directional projection feature enhancement method. The multi-directional projection feature enhancement method performs a convolution operation on the input features and outputs intermediate data, wherein the intermediate data is F q 、F k and F v ; The intermediate data is enhanced by using a multi-directional projection window and an attention mechanism to obtain a first output feature, and a convolution operation is performed on the intermediate data by using a convolution to obtain a second output feature; According to the first output feature and the second output feature, a feature after feature enhancement is obtained.

4. The Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation according to claim 3 is characterized in that: The intermediate data is enhanced by using a multi-directional projection window and attention to obtain a first output feature, comprising: Use multi-directional projection windows to project the intermediate data F k and F v Perform feature enhancement to obtain F k and F v Generated intermediate variables; F k and F v The generated intermediate variables are similar to F q Through the attention mechanism, the first output feature is output.

5. The Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation according to claim 4 is characterized in that: Use multi-directional projection windows to project the intermediate data F k and F v The feature enhancement is: {F k_m ,F v_m }=R(σf 7×7 (pad(F k ,F v ))+b) Where pad(·) represents a padding operation of size P, and f k×k (·) represents the projection window operation of size k×k, R(·) represents the Reshape operation, and F k_m and F v_m F k and F v The generated intermediate variable, σ is the activation function, and b is the bias term.

6. The Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation according to claim 1 is characterized in that: The parameter adaptation degree in S3 is obtained according to the data resources of the edge client and the initial global model.

7. The Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation according to claim 1 is characterized in that: The cloud server in S4 aggregates the local model according to the parameter fitness and the negative gain aggregation hyperparameter, including: According to the parameter fitness and the negative gain aggregation hyperparameter, obtaining a weight redistribution coefficient of model aggregation; The weight redistribution coefficient of the model aggregation is: Among them, α i represents the aggregation weight redistribution coefficient of client i, S i is the fitness of client i, β is the hyperparameter of the negative gain aggregation method, N represents the number of all clients participating in the training, S k is the fitness of client k; The local models are aggregated according to the weight redistribution coefficients to obtain the new global model.

8. The Internet of Things security method based on multi-directional projection feature enhancement and negative gain aggregation according to claim 7 is characterized in that: The new global model is: in, represents the local model of the i-th client, w t represents the global model, represents the total number of data of edge client i, represents the total number of data of all clients, and N represents the number of all clients participating in the training.

Images