Internet of Things cooperative sensing method based on encryption protection

Abstract

The invention relates to an Internet of Things cooperative sensing method based on encryption protection, and belongs to the technical field of Internet of Things. According to the method, security protection of sensitive data in communication transmission and cloud computing processes is ensured by using a fully homomorphic encryption algorithm, and a convolutional neural network model framework is designed to perform data prediction and abnormal source identification in combination with multi-source factor data. The model fuses a space environment topological structure and content features, and is beneficial for improving the scientificity of collaborative perception decision-making in the scene of the Internet of Things. According to the invention, through combination of the fully homomorphic encryption technology and the graph convolutional neural network, data security and sharing problems in an Internet of Things data integration processing process are solved, security sharing application of cross-regional perception information is realized, and the depth decision level of Internet of Things monitoring management is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet of Things, and relates to an Internet of Things collaborative perception method based on encryption protection. Background Art

[0002] With the in-depth application of IoT technology in different perception scenarios and professional fields, various data collection methods have gradually achieved comprehensive and real-time coverage. Due to the characteristics of data diversity, massive and high-frequency updates, and often involving cross-departmental application needs, data sharing and protection face huge challenges. The main problems of current monitoring data include: insufficient protection of sensitive information, high difficulty in data utilization, and insufficient understanding of data value. These problems limit the in-depth utilization and value maximization of monitoring data.

[0003] At present, the sensing and monitoring networks at all levels have been basically established. For example, there are more than 3,000 water conservancy monitoring points in a certain city, and the amount of data generated every day reaches TB level. Such data contains the layout and parameter information of key infrastructure. Once leaked, it may have a serious impact on corporate interests. At the same time, since the analysis of monitoring data often involves multiple units and regions, it is difficult to share and use data due to technical conditions, which hinders the further improvement of management efficiency.

[0004] Fully homomorphic encryption technology can realize direct calculation of encrypted data while ensuring data privacy. In the process of continuous monitoring, joint analysis services can effectively solve the problem of data sharing. However, the current application of homomorphic encryption in the monitoring field still needs to overcome the difficulties of algorithm efficiency and customization of deep learning models. The characteristics of different types of variables vary in encryption representation and calculation mode, so further in-depth research is needed.

[0005] In the field of perception monitoring, the dynamic changes of multi-source data present multi-variable and nonlinear characteristics, and artificial neural networks have good predictive capabilities for nonlinear problems. Although the deployment of automatic monitoring stations provides a large amount of high-frequency data, which can better reflect the dynamic changes of variables than traditional sampling frequencies, the neural network monitoring data analysis model based on spatial association is still relatively lacking.

[0006] Therefore, a new IoT collaborative perception method is urgently needed to solve the above problems. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide an Internet of Things collaborative perception method based on encryption protection, which solves the data security and sharing problems in the process of Internet of Things data integration and processing by combining fully homomorphic encryption technology with graph convolutional neural network, realizes the safe sharing and application of cross-regional perception information, and improves the deep decision-making level of Internet of Things monitoring and management.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] A collaborative perception method for the Internet of Things based on encryption protection uses the fully homomorphic encryption (CKKS) algorithm to ensure the security of sensitive data during communication transmission and cloud computing. At the same time, it combines multi-source factor data and designs a graph convolutional neural network (GCNN) model framework for data prediction and abnormal source identification. The model integrates the spatial environment topology structure and content characteristics, which is conducive to improving the scientific nature of collaborative perception decision-making in the Internet of Things scenario.

[0010] The method specifically comprises the following steps:

[0011] S1: On the remote terminal control system device (RTU) of the IoT monitoring station, communicate with various types of sensors to collect monitoring data in the current scenario at a preset frequency. For example, monitoring data in a water conservancy environment includes water level, flow, water quality, and seepage pressure parameters, etc. The monitoring data of multiple sensors are synchronized and fused to generate standardized time point monitoring data.

[0012] S2: Convert and standardize the collected raw monitoring data, unify timestamps and fill missing values, aggregate monitoring data by time series, build a multi-dimensional monitoring data set, downsample and compress the data set to reduce the data volume, and then perform deterministic verification on the processed data for subsequent homomorphic encryption and computing;

[0013] S3: Time-sorting and dividing the pre-processed monitoring data set according to the preset storage strategy, and storing it in the local cache space of the RTU;

[0014] S4: Use the CKKS algorithm to generate a public-private key pair, and use the public key to homomorphically encrypt the cached data to generate encrypted ciphertext data. The public-private key pair needs to be updated regularly, and the private key is stored in the cloud server for decryption operations after subsequent calculations. At the same time, perform module length management and expansion on the ciphertext data to ensure that the encrypted data meets the numerical range requirements of homomorphic calculations. Among them, the CKKS algorithm represents a fully homomorphic encryption algorithm.

[0015] S5: Transmit the homomorphically encrypted ciphertext data to the cloud server via a wireless network, encapsulate and encode it according to the communication protocol, and transmit the ciphertext data in real time to ensure that real-time data is transmitted first. In case of abnormality, the reliability of transmission is ensured through a disconnection and reconnection mechanism.

[0016] S6: receiving the uploaded homomorphic encrypted ciphertext on the cloud server, and performing forward propagation and back propagation calculations on the monitoring data in the ciphertext domain through the GCNN model, completing model training or inference, and finally generating encrypted ciphertext of the prediction result; wherein GCNN represents graph convolutional neural network;

[0017] S7: The ciphertext result obtained by cloud computing is transmitted to the terminal through the network, and the ciphertext is decrypted using the private key. The decrypted data is restored to the plaintext result, and the decryption result report is displayed on the front-end page. At the same time, the decryption log is recorded and potential improper use is monitored.

[0018] Furthermore, in step S3, the cache space uses a ring buffer or a fixed cache structure to manage the data of each time slice, and regularly clears outdated data to ensure cache resource utilization of the embedded device.

[0019] Further, in step S4, the specific steps of using the CKKS algorithm for encryption and decryption are:

[0020] (1) Initialize basic parameters: Given the basic integer p and the initial modulus q0, initialize the security parameter λ, the power N of the polynomial modulus, and the noise distribution parameter X key ,X err ,x enc ; Analyze the data characteristics of different types of data sources, including statistics such as mean, variance, maximum and minimum values, so as to select the appropriate ciphertext modulus for each data type;

[0021] (2) Define the adaptive parameter γ i : Set adaptive parameters γ according to the data characteristics of different data types i In order to balance noise growth and data accuracy during encryption; γ can be defined according to the variance of different data types and data change rate. i , for example: γ i =α·σ i ;

[0022] Among them, α is a tuning parameter used to control the overall range of the ciphertext modulus size. σ i is the standard deviation of data source i, the larger the variance, γ i Larger values ​​provide higher modulus for data with faster noise growth.

[0023] (3) Define the adaptive modulus q l : Based on the adaptive parameter γ i and the number of encryption layers l i , set the corresponding modulus sequence for different data sources l i It is the current calculation depth of a data source, and the modulus is dynamically adjusted according to its data characteristics and noise accumulation. For example, position data position: where γ position Represents the adaptive parameter for position data; 1≤l i≤L, where L represents the global maximum number of encryption layers;

[0024] This adaptive modulus setting ensures that for data with large variations (such as position data), a larger modulus q is assigned l , to control the rapid growth of noise during the calculation process; for data with small changes, such as temperature, a smaller modulus q is assigned l , in order to reduce the encryption computation overhead.

[0025] (4) Generate public parameters PP: Adaptive modulus sequence based on each data source Construct adaptive public parameters PP, including adaptive moduli and basic encryption parameters; assuming the polynomial modulus N of the encryption ring and the ring size k = 2, the generated public parameters include: Where L represents the global maximum number of encryption layers; Represents the modulus sequence of different data sources, which is used for noise control and adaptive parameter selection in subsequent encryption and calculation processes;

[0026] (5) Execute key generation, encryption and decryption operations: Use the defined adaptive public parameter PP to generate keys, encrypt data and decrypt data respectively:

[0027] Key generation: Generate public key pk, private key sk and evaluation key evk through PP:

[0028] KeyGen(PP)→(pk,sk,evk)

[0029] Encryption: Use the public key pk to encrypt different data sources and generate adaptive modulus ciphertext ct i :

[0030] Encrypt(m i ,pk)→ct i

[0031] Among them, m i Represents different types of sensor data (such as temperature, humidity, location, device status, etc.), and each type of data selects the corresponding module according to its characteristics

[0032] Decryption: Use the private key sk to decrypt the ciphertext ct i Decrypt and get the original data m i :

[0033] Decrypt(ct i ,sk)→m i

[0034] (6) Adaptive modulus adjustment in addition and multiplication operations: When performing addition and multiplication operations between different data sources, the appropriate modulus q is selected based on the adaptive modulus of the data. l To control the noise growth. For example, for the addition operation of temperature and humidity data, select the modulus of humidity data (larger modulus) to suppress the noise growth:

[0035]

[0036] For multiplication operations, a data modulus with slower noise growth can be selected, or the modulus size can be dynamically adjusted to balance the calculation accuracy and noise accumulation issues.

[0037] An adaptive polynomial parameter selection is proposed to make the ciphertext modulus q l It can dynamically adapt to the characteristics of different data sources in IoT scenarios to improve encryption computing efficiency and data security.

[0038] Further, in step S4, in the CKKS algorithm, for the object to be encrypted, which is information data collected in different IoT scenarios, such as water level, water quality, rainfall and seepage pressure in water conservancy scenarios, it is necessary to functionalize the data before encryption, and map the numerical monitoring data to a finite-dimensional complex number field. Using the default polynomial representation of CKKS can more conveniently represent the changing laws of various monitoring data, and at the same time, it is more compact and easy to regenerate the original curve, which is conducive to decryption and restoration as well as multi-source data integration and analysis.

[0039] Further, in step S6, the GCNN model based on the topological relationship of multiple monitoring points is constructed in the homomorphic ciphertext space, which specifically includes the following steps:

[0040] (1) Distribute IoT monitoring points at different locations in the collection area, collect geographic information data at different monitoring points, and build a graph structure between monitoring points based on geographic information, where nodes represent monitoring points, edges represent the relationship between them, and edge weights can reflect the similarity or distance between nodes;

[0041] Given N monitoring points, the entire topology information can be represented by a graph G = (V, E), where V = (v0, v1, ..., v N-1 ) is the set of monitoring points, and E represents the set of edges between monitoring points;

[0042] (2) The multiple data uploaded by the monitoring point to the cloud server include time series data of important information indicators, such as water quality indicators in the water conservancy environment, including dissolved oxygen, ammonia nitrogen, pH value, etc. The node feature matrix X is composed of the information data indicator values ​​obtained from the most recent monitoring of each monitoring point, and the adjacency matrix is ​​used to represent the topological relationship; the feature matrix X is expressed as:

[0043]

[0044] Among them, x i The feature vector representing monitoring point i contains the data values ​​of different monitoring indicators;

[0045] (3) Since the IoT scenario is a real-time, dynamically changing system, according to the first law of geography, there are different degrees of influence relationships between monitoring points. Use the following formula to calculate the distance d between two monitoring points, and use the reciprocal as the edge weight value and store it in the adjacency matrix A;

[0046]

[0047] Among them, Δh=|h1-h2|, Δφ=φ2-φ1, Δλ=λ2-λ1; Δh represents the height difference between the two monitoring points, η is the vertical distance weight factor, which reflects the influence of vertical distance in the calculation; Δφ and Δλ represent the latitude difference and longitude difference respectively; w represents the weight factor, which is used to dynamically adjust the importance of distance in a specific IoT scenario and can be set according to the environmental characteristics of the monitoring point or the device priority; R is the radius of the earth;

[0048] (4) The GCNN model uses the geographical location information, topological structure G and feature matrix X of the monitoring points to train and learn a complex mapping function to extract spatial features;

[0049] In the GCNN model, node features are represented by aggregating the features of neighboring nodes and convolving them to generate high-level feature representations; in graph convolution, node v i The feature aggregation is expressed as:

[0050]

[0051] in, is the feature representation of node i at layer k; is the neighborhood of node i; d i and d j are the degrees of nodes i and j respectively; W (k) is the weight matrix of the kth layer, used to learn the feature map; σ is the activation function used to increase nonlinearity; w ij is the edge weight, which indicates the similarity relationship between nodes i and j. It is the inverse of the distance d, indicating that the smaller the distance, the greater the edge weight.

[0052] The graph convolution operation is expressed as the convolution calculation of the node feature matrix X through the adjacency matrix A as follows:

[0053]

[0054] in, That is, add self-loops to the adjacency matrix, and I is the identity matrix; yes The degree matrix of (k) is the feature matrix of the kth layer; H represents the number of multi-heads, that is, the number of multi-head aggregations; and denote the adjacency matrix and adaptive edge weight coefficient of the h-th head respectively; W (k，h) Represents the weight matrix of the hth head in the kth layer, and each head captures different feature representations through an independent weight matrix. This formula realizes the smooth propagation of node features in each layer and extracts spatial features. At the same time, a multi-head aggregation mechanism is used to learn and aggregate neighborhood information from multiple subspace perspectives, thereby improving the richness of feature expression and forming an enhanced feature representation.

[0055] Furthermore, in step S6, the GCNN model training is completed in the homomorphic encryption space, specifically including: calculating the loss function in the ciphertext space and maintaining the privacy of the loss value through homomorphic encryption; updating the model weight W through gradient descent in the ciphertext space:

[0056]

[0057] Among them, η is the learning rate and Enc(L) is the encrypted loss function.

[0058] The beneficial effects of the present invention are:

[0059] (1) The present invention, based on the research of fully homomorphic encryption, can perform calculations without exposing the original data, thereby effectively protecting the security of monitoring data. Storing monitoring data in the cloud and performing calculations in the cloud has many advantages, such as resource sharing and elastic expansion. However, there may be a risk of information leakage when transmitting sensitive monitoring data to the cloud. The fully homomorphic encryption of the present invention provides a security protection method for performing calculations in the cloud, which can perform useful calculations without exposing plaintext data.

[0060] (2) The topological relationship diagram of the monitoring points constructed by the present invention can more comprehensively capture the complexity and interrelationship of the perception system. By training the graph convolutional neural network in the ciphertext domain and utilizing its capabilities on graph data, the topological structure between different monitoring points can be comprehensively considered, which helps to more accurately understand the overall characteristics of the perception system. In collaborative management, data cooperation between different units is involved. The research based on fully homomorphic encryption of the present invention helps to realize cross-unit data calculation in an encrypted state, promote the sharing and cooperation of monitoring data, and protect the privacy of each unit.

[0061] (3) The present invention can perform more in-depth analysis of monitoring data in the cloud by performing calculations in an encrypted state, providing more accurate decision support. The collaborative analysis of the spatiotemporal attribute relationship between different monitoring points is of great significance for the planning, management and prediction of Internet of Things technology and engineering.

[0062] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0064] Figure 1 The overall flow chart of the Internet of Things collaborative perception method based on encryption protection provided by the present invention;

[0065] Figure 2 The application flow chart of the CKKS fully homomorphic encryption algorithm provided by the present invention;

[0066] Figure 3 This is a diagram of the Laplace matrix generation process of the GCNN model provided by the present invention. DETAILED DESCRIPTION

[0067] The following describes the embodiments of the present invention by 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 the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept 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.

[0068] The embodiment of the present invention will provide an application description of the implementation steps for water quality environment detection in the water conservancy Internet of Things collection scenario. The objects to be encrypted are various water conservancy data, including water level, water quality, rainfall and seepage pressure, etc. The data needs to be functionalized before encryption, and the numerical monitoring data is mapped to a finite-dimensional complex number field. Using the default polynomial representation of CKKS can more conveniently represent the changing laws of various monitoring data, and at the same time, the representation is more compact and easy to regenerate the original curve, which is conducive to decryption and restoration as well as multi-source data integration and analysis.

[0069] See also Figure 1-2 The present invention provides a system for fully homomorphic encryption using the CKKS algorithm. The system is mainly based on edge computing centralized procurement terminals, deploys encryption algorithms and model frameworks, and completes the training and prediction of IoT perception data under the collaboration of multiple monitoring points in combination with the cloud. It includes a data acquisition module, a data processing module, a homomorphic encryption module, and a homomorphic computing module, which are divided into two aspects: edge computing and cloud computing. A feature expression method for perceiving scenes in a homomorphic space is formed, and a fully homomorphic encryption algorithm (CKKS) suitable for IoT data is applied. At the same time, geographic information and data features of multiple IoT monitoring points are extracted to construct topological structures and parameter attributes, and a graph convolutional neural network model (GCNN) based on a topological space graph of monitoring points is trained.

[0070] The process steps of the above system are mainly divided into: encoding, encryption, homomorphic computing, decryption and decoding.

[0071] S1: First, specify the polynomial order N, which means that the polynomial only contains coefficients from order 0 to order N. For each monitoring value x, calculate its coefficient in the standard basis polynomial {1, x, x 2 , ..., x N The coefficients in} are combined to form a coefficient vector and a coefficient matrix C as the input quantization matrix of the encryption system.

[0072] S2: Given a security parameter λ, select two integer powers N, respectively The above is the key, error learning, and encryption setting distribution χ key , χ err , χ enc For a base integer p and level number L, and a base modulus q0, set the modulus q of the ciphertext l =p l q0(1≤l≤L), for the ring Here, k=2 is set, and then the parameter PP is randomly generated, and the output is: pp=(N, χ key , χ err , χ enc , L, q l ).

[0073] The choice of security parameters is related to the cost and capability of homomorphic computing. The number of levels L, i.e. setLevel, controls the number of bits of the encrypted number. The larger the value, the stronger the homomorphic computing capability, but the computing cost also increases. It is generally selected between 30 and 60, which defines the maximum precision depth supported by the system. The modulus length q generated by each level l lThat is, modulusQ, which describes the size of the modulus of the operation. The larger the value, the smaller the calculation error, but the calculation cost is also higher. Lower setLevel and modulus can speed up the calculation, but the homomorphic operation may be over-modularized and fail earlier. Higher parameters provide stronger homomorphic computing power and deeper calculations, but the cost will also increase exponentially. Therefore, in the actual process, it is necessary to dynamically adjust these two core parameters and other indicators according to the complexity of the model and the scale of the data to balance security, correctness and calculation cost.

[0074] S3: Key generation: KeyGen(params)→(pk, sk, evk), including:

[0075] Randomly generate s←χ key , and set the private key sk←(1,s).

[0076] Random Generation and error e←χ err , set the public key:

[0077] Random Generation and e′←χ err , set the evaluation key:

[0078] S4: Encryption: Encrypt(m, pk)→ct, including:

[0079] Randomly generate r←χ enc and e0, e1←χ err , output ciphertext c, and

[0080] c=r·pk+(m+e0,e1)(modq L )

[0081] S5: Decryption: Decrypt(c, sk)→m, including:

[0082] For ciphertexts of the same level l, decryption yields the plaintext result:

[0083] m′=〈c, sk〉(modq l )

[0084] S6: Addition: Add(c, c′) → c add , specifically including:

[0085] For the ciphertext ct, ct′, the ciphertext addition result is

[0086] c add =c+c′(modq l )

[0087] S7: Multiplication: Mult evk (c,c′)→c mult , including:

[0088] For ciphertext c, c′:

[0089]

[0090] (d0, d1, d2)=(c0c′0, c0c′1+c′0c1, c1c′1)(modq l )

[0091] The result of the ciphertext multiplication is

[0092]

[0093] S8: Use the public key to encrypt the plaintext functional coefficient matrix into a ciphertext tensor supported by the CKKS system, and then use the modular operation CKKS.Modulo in the CKKS system to perform modular truncation on the encrypted tensor. The main purpose is to control the accumulation of floating-point errors and maintain the reversibility of the calculation. Repeating this operation is an important means to control floating-point errors.

[0094] S9: In the key management part, the public-private key pair is generated by the collection device through a random algorithm. The public key is used by the collection device to encrypt the original data to be uploaded, while the private key is securely stored by the cloud computing analysis end. The corresponding private key is selected based on the device identification information and decrypted when needed. To ensure the long-term security of the key, the key can be updated and regenerated regularly.

[0095] See also Figure 3 , water conservancy monitoring points are distributed in different river basins. The water conservancy data uploaded to the cloud by the monitoring points contain time series data of important water quality indicators, such as dissolved oxygen, ammonia nitrogen, pH value, etc. The geographic information data of the basins of different monitoring points are collected, and a graph structure between monitoring points is constructed based on geographic information, where nodes represent monitoring points, edges represent the relationship between them, and the weight of the edge can reflect the similarity or distance between nodes. The water quality index values ​​obtained by the most recent monitoring of each monitoring point constitute the node feature vector, and the adjacency matrix is ​​used to represent the topological relationship.

[0096] Define the signal diagram of the monitoring point. The signal is represented by a vector, which contains characteristic indicators related to the water quality prediction of the monitoring point, and is represented by the characteristic matrix X N×T represents the signal of each monitoring point. The historical time series length is provided by T, that is, when T = t, X N×t It represents the set of current values ​​of different water quality indicators at N monitoring points in the selected monitoring basin at time t.

[0097] The graph convolutional neural network model (GCNN) uses the geographic location information, topological structure G and feature matrix X of the monitoring points to train and learn a complex mapping function to extract spatial features. The main steps are as follows:

[0098] M1. Construct a topological map D and an adjacency matrix A based on the geographical location of the monitoring points.

[0099] M2. Construct the feature matrix X using the water quality index information values ​​in the water conservancy monitoring data at each point.

[0100] M3, perform Laplace transform based on adjacency matrix A to generate Laplace transform matrix It can be expressed as:

[0101]

[0102] Among them, (A+I N ) is the self-connected adjacency matrix, I N is the identity matrix, is the degree matrix.

[0103] M4. Based on the graph convolution operation, the model will use the encrypted data of the adjacency matrix and the node feature matrix to perform convolution calculations and extract water quality features. The convolution operation formula in the encrypted space is as follows:

[0104]

[0105] Where σ(·) is the activation function, which is used to achieve nonlinear transformation. (i) represents the weight matrix of the i-th layer, H (i) represents the activation value of the i-th layer, and H(0) = X. In the homomorphic encryption space, the adjacency matrix and feature matrix X are encrypted, where Enc(·) represents the encrypted matrix or vector. The approximate activation function of the ciphertext space is used to realize nonlinear feature transformation, which is used for nonlinear processing of encrypted data. GCNN then dynamically captures the influence weights of different monitoring points through the Laplace matrix to obtain spatial features.

[0106] According to the above four steps, the geographical location information of the monitoring points in the target basin is used to construct a topological map, and the corresponding adjacency matrix, degree matrix and Laplace matrix are calculated and generated based on the topological map, as shown in Figure 3 shown.

[0107] M5. Complete model training in homomorphic encryption space to ensure that the original data will not be leaked when water quality prediction is executed in the cloud. Calculate the loss function in the ciphertext space and keep the privacy of the loss value through homomorphic encryption. Update the model weight W through gradient descent in the ciphertext space:

[0108]

[0109] Where η is the learning rate and Enc(L) is the encrypted loss function. After the model training is completed, the new encrypted water quality data is predicted to generate the predicted value of the future water quality state. After decryption, the prediction results are used for analysis and early warning.

[0110] Consider adding link prediction to the graph, analyzing the possible water quality transmission paths between different monitoring points, detecting abnormal changes, whether there is transmission and association between nodes, and identifying monitoring points that may be pollution sources. For example, if the predicted values ​​of individual monitoring points or regions have large deviations, they may be at the source of water pollution or within the affected area.

[0111] Specifically, in terms of implementation, CrypTen is a privacy-preserving machine learning framework based on PyTorch. Its goal is to allow machine learning practitioners to use secure computing technology, which can greatly simplify the development difficulty of homomorphic computing using CKKS, while meeting the scenario requirements of building neural networks and deep learning applications in homomorphic spaces on cloud servers. CrypTen provides an object-oriented programming model, uses Modules and Tensors to define the model structure, and has built-in common layers and operator implementations, such as convolutional layers, fully connected layers, etc. Using CrypTen in combination with CKKS can facilitate and efficiently perform deep learning predictions of water quality under homomorphic encryption.

[0112] This embodiment provides an IoT collaborative perception method based on encryption protection. The method realizes the propagation and fusion of hydrological characteristics in the entire spatial area by combining fully homomorphic encryption technology and graph convolutional neural network, and digs deeper into the inherent laws of water quality evolution from a global perspective. While protecting data security, it predicts the dynamic change laws of water quality and the sources of water pollution, etc., thereby promoting the realization of hydrological security sharing.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A collaborative perception method for the Internet of Things based on encryption protection, characterized in that: The method specifically comprises the following steps: S1: On the remote terminal control system device of the IoT monitoring station, communicate with various types of sensors, collect monitoring data in the current scenario at a preset frequency, and synchronize and fuse the monitoring data of multiple sensors to generate standardized time point monitoring data; S2: Convert and standardize the collected raw monitoring data, unify the timestamps and fill in missing values, aggregate the monitoring data by time series, build a multi-dimensional monitoring data set, downsample and compress the data set, and then perform deterministic verification on the processed data; S3: Time-sorting and dividing the pre-processed monitoring data set according to a preset storage strategy, and storing it in the local cache space of the remote terminal control system device; S4: Use the CKKS algorithm to generate a public-private key pair, and use the public key to homomorphically encrypt the cached data to generate encrypted ciphertext data. The public-private key pair needs to be updated regularly, and the private key is stored in the cloud server for decryption operations after subsequent calculations. At the same time, perform module length management and expansion on the ciphertext data to ensure that the encrypted data meets the numerical range requirements of homomorphic calculations. Among them, the CKKS algorithm represents a fully homomorphic encryption algorithm. S5: Transmit the homomorphically encrypted ciphertext data to the cloud server via a wireless network, encapsulate and encode it according to the communication protocol, and transmit the ciphertext data in real time to ensure that real-time data is transmitted first. In case of abnormality, the reliability of transmission is ensured through a disconnection and reconnection mechanism. S6: receiving the uploaded homomorphic encrypted ciphertext on the cloud server, and performing forward propagation and back propagation calculations on the monitoring data in the ciphertext domain through the GCNN model, completing model training or inference, and finally generating encrypted ciphertext of the prediction result; wherein GCNN represents graph convolutional neural network; S7: The ciphertext result obtained by cloud computing is transmitted to the terminal through the network, and the ciphertext is decrypted using the private key. The decrypted data is restored to the plaintext result, and the decryption result report is displayed on the front-end page. At the same time, the decryption log is recorded and potential improper use is monitored.

2. According to claim 1, the method for collaborative perception of the Internet of Things based on encryption protection is characterized in that: In step S3, the cache space uses a ring buffer or a fixed cache structure to manage the data of each time slice, regularly clears outdated data, and ensures the cache resource utilization of the embedded device.

3. According to the encryption protection-based collaborative sensing method of the Internet of Things according to claim 1, it is characterized in that: In step S4, the specific steps of encryption and decryption using the CKKS algorithm are: (1) Initialize basic parameters: Given the basic integer p and the initial modulus q0, initialize the security parameter λ, the power N of the polynomial modulus, and the noise distribution parameter χ key , χ err , χ enc ; Analyze the data characteristics of different types of data sources, including mean, variance, maximum and minimum values, so as to select the appropriate ciphertext modulus for each data type; (2) Define the adaptive parameter γ i : Set adaptive parameters γ according to the data characteristics of different data types i In order to balance noise growth and data accuracy during encryption; (3) Define the adaptive modulus q l : Based on the adaptive parameter γ i and the number of encryption layers l i , set the corresponding modulus sequence for different data sources l i It is the current calculation depth of a data source, and the modulus is dynamically adjusted according to its data characteristics and noise accumulation; (4) Generate public parameters PP: Adaptive modulus sequence based on each data source Construct adaptive public parameters PP, including adaptive moduli and basic encryption parameters; assuming the polynomial modulus N of the encryption ring and the ring size k = 2, the generated public parameters include: Where L represents the global maximum number of encryption layers; Represents the modulus sequence of different data sources, which is used for noise control and adaptive parameter selection in subsequent encryption and calculation processes; (5) Execute key generation, encryption and decryption operations: Use the defined adaptive public parameter PP to generate keys, encrypt data and decrypt data respectively: Key generation: Generate public key pk, private key sk and evaluation key evk through PP: KeyGen(PP) → (pk, sk, evk) Encryption: Use the public key pk to encrypt different data sources and generate adaptive modulus ciphertext ct i : Encrypt(m i ,pk)→ct i Among them, m i Represents different types of sensor data, and each type of data selects the corresponding module according to its characteristics Decryption: Use the private key sk to decrypt the ciphertext ct i Decrypt and get the original data m i : Decrypt(ct i ,sk)→m i (6) Adaptive modulus adjustment in addition and multiplication operations: When performing addition and multiplication operations between different data sources, the appropriate modulus q is selected based on the adaptive modulus of the data. l To control the noise growth.

4. The method for collaborative perception of the Internet of Things based on encryption protection according to claim 1 or 3, characterized in that: In step S4, in the CKKS algorithm, for the object to be encrypted, which is information data collected in different IoT scenarios, it is necessary to functionalize the data before encryption and map the numerical monitoring data to a finite-dimensional complex number field.

5. The method for collaborative sensing of the Internet of Things based on encryption protection according to claim 1 is characterized in that: In step S6, the GCNN model based on the topological relationship of multiple monitoring points is constructed in the homomorphic ciphertext space, which specifically includes the following steps: (1) Distribute IoT monitoring points at different locations in the collection area, collect geographic information data at different monitoring points, and build a graph structure between monitoring points based on geographic information, where nodes represent monitoring points, edges represent the relationship between them, and edge weights reflect the similarity or distance between nodes; Given N monitoring points, the entire topology information is represented by a graph G = (V, E), where V = (v0, v1, ..., v N-1 ) is the set of monitoring points, and E represents the set of edges between monitoring points; (2) The node feature matrix X is formed by the information data index values ​​obtained from the most recent monitoring of each monitoring point, and the adjacency matrix is ​​used to represent the topological relationship; the feature matrix X is expressed as: Among them, x i The feature vector representing monitoring point i contains the data values ​​of different monitoring indicators; (3) Use the following formula to calculate the distance d between two monitoring points, and use the reciprocal as the edge weight value, which is stored in the adjacency matrix A; Among them, Δh=|h1-h2|, Δφ=φ2-φ1, Δλ=λ2-λ1; Δh represents the height difference between the two monitoring points, η is the vertical distance weight factor, which reflects the influence of vertical distance in the calculation; Δφ and Δλ represent the latitude difference and longitude difference respectively; w represents the weight factor, which is used to dynamically adjust the importance of distance in a specific IoT scenario, according to the environmental characteristics of the monitoring point or the device priority setting; R is the radius of the earth; (4) The GCNN model uses the geographical location information, topological structure G and feature matrix X of the monitoring points to train and learn a complex mapping function to extract spatial features; In the GCNN model, node features are represented by aggregating the features of neighboring nodes and convolving them to generate high-level feature representations; in graph convolution, node v i The feature aggregation is expressed as: in, is the feature representation of node i at layer k; is the neighborhood of node i; d i and d j are the degrees of nodes i and j respectively; W (k) is the weight matrix of the kth layer, used to learn the feature map; σ is the activation function used to increase nonlinearity; w ij is the edge weight, which indicates the similarity relationship between nodes i and j. It is the inverse of the distance d, indicating that the smaller the distance, the greater the edge weight; The graph convolution operation is expressed as the convolution calculation of the node feature matrix X through the adjacency matrix A as follows: in, That is, add self-loops to the adjacency matrix, and I is the identity matrix; yes The degree matrix of (k) is the feature matrix of the kth layer; H represents the number of multi-heads, that is, the number of multi-head aggregations; and denote the adjacency matrix and adaptive edge weight coefficient of the h-th head respectively; W (k，h) Represents the weight matrix of the h-th head in the k-th layer. Each head captures different feature representations through an independent weight matrix.

6. The method for collaborative sensing of the Internet of Things based on encryption protection according to claim 5 is characterized in that: In step S6, the GCNN model training is completed in the homomorphic encryption space, specifically including: calculating the loss function in the ciphertext space and maintaining the privacy of the loss value through homomorphic encryption; updating the model weight W through gradient descent in the ciphertext space: Among them, η is the learning rate and Enc(L) is the encrypted loss function.

Images