A non-invasive load monitoring method with privacy protection in smart grid

Abstract

The present invention is a privacy-preserving non-invasive load monitoring method for smart grids, which decomposes smart meter readings from the household level to the appliance level. This method enables a user and server to jointly perform a non-invasive load monitoring task based on a deep neural network under privacy-preserving conditions. The server utilizes a pre-trained model to provide prediction services, and the user maintains their own household-level smart meter readings. The present invention utilizes lattice-based homomorphic encryption, secret sharing, and an oblivious transfer protocol to apply privacy-preserving techniques to the deep neural network (PSPNet). This enables the privacy-preserving PSPNet deep neural network to be used for NILM tasks, addressing the security requirements of both users and servers in NILM tasks.

Description

Technical Field

[0001] The present invention relates to a non-invasive load monitoring method under privacy protection in a smart grid. The method is based on secure multi-party computing technology and homomorphic encryption technology, and belongs to security and confidentiality technology. Background Art

[0002] The non-intrusive load monitoring (NILM) task in smart grids can decompose the meter readings of the entire household from the household level to the appliance level. Given the total household load L at time t, t , for I target device, NILM can decompose it into: in represents the load consumption of the i-th appliance at time t, γ t represents the residual error or unmonitored load at time t. Furthermore, if we set a threshold for each device's load, we can decompose the on / off status of each appliance using household meter readings. Implementing the NILM task has important real-world implications: For households, obtaining feedback on specific devices is more effective than obtaining aggregate feedback, as it can provide personalized recommendations to each household, allowing them to adjust their electricity usage based on their individual device usage. Furthermore, householders can use device-level feedback to promptly diagnose devices and conduct inspections if abnormalities in device usage occur. For R&D, businesses and academics can leverage the collected data on actual energy consumption across different appliances to strategically focus their efforts on targeted R&D to reduce energy consumption. For utilities, this data can facilitate quarterly load monitoring for power companies, providing guidance for energy procurement, power distribution, and power generation.

[0003] NILM was first proposed by Hart, who used combinatorial optimization (CO) to perform load decomposition. Since then, various machine learning models have been proposed to achieve high accuracy, primarily categorized as unsupervised and supervised learning algorithms. Commonly used unsupervised learning algorithms include the factorial hidden Markov model (FHMM) and cluster analysis. For supervised learning, models based on deep neural networks (DNNs) are the most representative architectures for NILM. PSPNet, a DNN model originally developed for image semantic segmentation, was chosen as the DNN model for NILM tasks due to its superior performance in learning the inherent features of the target.

[0004] In a privacy-preserving scenario, the DNN-based NILM task can be considered a privacy-preserving machine learning (PPML) task, consisting of two components: a user and a server. The user holds their own meter reading data, while the server possesses a trained DNN model. In this task, the user wishes to use their data as input to the DNN model and obtain the corresponding output, without exposing the server to any information about their input and output. The server wishes to use its trained DNN model to provide prediction services for the user, without exposing the user to any information about the model parameters, but the network structure is publicly available. The scenario addressed by this invention also includes a user and a server. The server has pre-trained its own PSPNet model and obtained a high-quality model. It wishes to use this model to provide NILM prediction services for profit, but does not want its relevant model parameters to be leaked. The user, which can be a household, business, or researcher, wishes to use its meter readings and appliance types as input to obtain decomposed appliance-level load consumption using the NILM prediction service, but does not want its meter readings and the load consumption of each appliance to be disclosed to the server.

[0005] The main technologies used in this solution include pyramid scenario analysis network, homomorphic encryption, secret sharing, etc.

[0006] The Pyramid Scene Parsing Network (PSPNet) was first proposed as a neural network for semantic segmentation. However, since its pyramid structure combines multi-scale features and has better performance in learning target features, we use it for the NILM task. Figure 1 The PSPNet network structure used in this invention can be divided into three modules: an extractor, a pyramid pooling module, and a classifier. The raw input data is processed using an extractor to generate a feature map. In the pyramid pooling module, the feature map is average-pooled with four different pooling kernel sizes to obtain four pooling results of different scales, representing features of the original input data at different scales. The four pooling results are then convolved, upsampled, and appended to the original feature map, achieving a combination of the original feature map and features of different scales. Finally, a classifier is used to obtain the output result.

[0007] Homomorphic encryption technology was proposed by Ron Rivest and others in 1978. Unlike general encryption schemes that focus on data storage security, homomorphic encryption focuses on data processing security, providing a method for processing encrypted data. Homomorphic encryption allows keyless parties to perform computations on ciphertext. This eliminates the need for key holders to perform computations, reducing communication costs and offloading computational tasks, thereby balancing computational costs among all parties. Homomorphic encryption also allows the decryptor to only obtain the final result, without accessing every ciphertext message, thereby improving information security. Homomorphic encryption satisfies the function Dec(f(Enc(m1), Enc(m2), …, Enc(mk))) = f(m1, m2, …, mk). Depending on the value of f, homomorphic encryption can be categorized as additive homomorphic, multiplicative homomorphic, and fully homomorphic.

[0008] Secret Sharing (SS) was first proposed by Shamir and Blakey in 1979. In a (n, t) threshold secret sharing scheme, a secret s is divided into n parts, each part being called a secret share and held by one participant. A sub-secret held by t or more participants can reconstruct s, while a sub-secret held by fewer than t participants cannot reconstruct s and cannot reveal any information about s. In this paper, we use (2, 2) additive secret sharing, where, for secret s, participant P1 holds s1 and participant P2 holds s2, where s1 + s2 = s.

[0009] In the pyramid pooling module, a key module in PSPNet, the operators that need to be jointly calculated include average pooling, convolution, and upsampling. In traditional privacy-preserving machine learning techniques, linear operations such as matrix multiplication and convolution are performed using Beaver triples. However, this method requires a time-consuming preprocessing phase and is not suitable for audiences such as households who only want to use the NILM prediction service to obtain decomposed electrical load consumption. To solve this problem, this solution uses homomorphic encryption technology to protect data, converting the huge communication overhead into computational overhead, and transferring most of the computational overhead to the server with more powerful computing power. For upsampling, the design uses additive secret sharing, allowing both parties to perform local calculations and reducing communication overhead. Summary of the Invention

[0010] The purpose of the present invention is to propose an efficient and secure method for two parties to jointly complete NILM tasks in response to the security and performance requirements in the smart grid scenario. It solves the problem of implementing NILM tasks based on the DNN model under the premise of privacy protection, and will not leak the privacy data of the user end and the server under the semi-honest model. At the same time, in view of the difference in computing power between the user end and the server in real scenarios, homomorphic encryption technology is used to protect the data, and the communication cost is converted into computing cost; the client only needs to encrypt the data and finally decrypt the result, and the complex homomorphic operations are performed by the server, which further realizes the transfer of computing costs to the server with stronger computing power. Moreover, the upsampling operation is implemented under privacy protection for the first time in this paper. Both parties can perform calculations locally without the need for communication, which further reduces communication overhead.

[0011] This paper combines privacy protection technologies such as machine learning and homomorphic encryption to implement a solution for two parties to jointly calculate NILM tasks. The technical solution is as follows.

[0012] The present invention involves two parties: a user and a server. The user holds the data, while the server holds the PSPNet model. Both parties jointly complete a neural network prediction, with the user inputting their own data and receiving a prediction result, and the server inputting model parameters. Neither party receives detailed information about the other's input. In the present invention, both parties collaborate layer by layer to perform privacy-preserving computations according to the PSPNet network structure. The specific process is as follows:

[0013] Preparation phase: In the preparation phase, both parties determine the number of extractor layers N based on the public network. E , Pyramid Pooling layer number N P , Classifier layer number N C , the calculation type of each layer and the related parameters of the homomorphic algorithm, where N E 、N P 、N C They are all natural numbers, and the specific values ​​depend on the selected network; the user then adds its own input for secret sharing and sends one share to the server.

[0014] Privacy-preserving computation extractor: Both parties have N in the extractor E The privacy protection calculation is performed layer by layer, and finally both parties obtain the additive secret share of the extracted feature map.

[0015] Privacy protection calculation pyramid pooling module: Both parties have N in the pyramid pooling module P The privacy protection calculation is performed layer by layer, where [T] c and [T]s Denote the secret shares of T held by the client and the server, respectively, satisfying T = [T] c +[T] s The same applies when T is replaced with other values. This module mainly consists of 3 steps.

[0016] (1) Average pooling layer: Four pooling operations with different pooling area sizes are required. For each pooling operation, its essence is to calculate the average value of all numbers in each pooling area. The operations involved are addition and scalar multiplication, both of which are linear operations. Both parties hold secret shares of the input value of average pooling. For additive secret sharing, performing linear operations on their secret shares and then reconstructing them is equivalent to performing the same linear operation on the secret. Therefore, for average pooling, both parties can locally perform average pooling on their secret shares, and ultimately both parties have additive secret shares of the pooling results.

[0017] (2) Convolution layer: Perform convolution operation on the four pooling results in the previous layer with the corresponding convolution kernel with a length and width of 1. The user end sends its own secret share [T] c After encryption using the RLWE homomorphic encryption algorithm, we get Enc([T] c ) is sent to the server, which uses the RLWE homomorphic algorithm to calculate the convolution and obtain the encrypted value Enc(T') of the convolution result. Then, [T'] is randomly sampled from the plaintext space. s As its own secret share, Enc(T')-Enc([T'] s )=Enc(T'-[T'] s ) is extracted to obtain the LWE ciphertext and sent to the user end, and finally the user end performs LWE decryption to obtain [T'] c =T'-[T'] s As your own secret share.

[0018] (3) Upsampling layer: After average pooling and convolution operations, features of four different dimensions are obtained. These features need to be merged with the original feature map. Therefore, the features of four different dimensions need to be upsampled to the same dimensions as the original feature map. The method used is bilinear interpolation. Bilinear interpolation only involves linear operations. Similar to the average pooling layer, both parties can locally perform bilinear interpolation on their secret shares.

[0019] Privacy-preserving computation classifier: Both parties have N C The privacy protection calculation is performed layer by layer, and finally both parties obtain the additive secret share of the final result.

[0020] Restoration phase: In the restoration phase, the server sends the shared value of the final result to the client, and the client adds the two shared values ​​to obtain the final prediction result.

[0021] To sum up the overall process of the present invention, when both parties are semi-honest, the user end can only obtain the final result and cannot obtain any information about the model parameters; while the server can only obtain the ciphertext form of the user data and cannot obtain other relevant information of the user input except the size of the user input data. Therefore, the protocol is semi-honest and secure.

[0022] Advantages and effects

[0023] The present invention is a two-party execution protocol for non-invasive load monitoring tasks in smart grids. The method integrates machine learning technology and privacy protection technologies such as homomorphic encryption, and the communication and computational overhead required are consistent with actual scenarios. Its advantages and functions are: 1) High security. During the entire process, the client cannot obtain any information about the server model parameters, and the server cannot obtain any other relevant information except the size of the user input data, thus protecting the privacy of both parties' inputs and meeting security requirements; 2) High performance. Using homomorphic encryption technology, communication overhead is converted into computational overhead, and the computational overhead is mainly transferred to the server with strong computing power. At the same time, the homomorphic algorithm used does not require homomorphic rotation operations, further reducing computational overhead; 3) No third party is required. Some common privacy protection schemes choose to introduce a trusted third party to perform auxiliary calculations. The protocol provided by the present invention is a two-party protocol that can be completed without the participation of a third party; 4) High scalability. The privacy-protected PSPNet network used in the present invention can be used for other machine learning tasks in addition to NILM tasks in smart grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is the PSPNet network structure. DETAILED DESCRIPTION

[0025] In order to make the above-mentioned objects, features and advantages of the present invention more obvious, the specific implementation of the non-invasive load monitoring method in the smart grid scenario will be described in detail below.

[0026] The relevant symbols are explained as follows:

[0027] (1) In the RLWE and LWE homomorphic encryption schemes, express polynomial on , where represents its kth coefficient;

[0028] (2) The extract of the RLWE scheme refers to the given RLWE ciphertext Can be extracted The kth coefficient is the LWE key sk The legal ciphertext (b, a);

[0029] (3)[T] c and [T] s Denote the secret shares of T held by the client and the server, respectively, satisfying T = [T] c +[T] s , the same applies when T is replaced with other values;

[0030] (4) In bilinear interpolation, f(x,y) represents the value of a two-dimensional vector at coordinates (x,y).

[0031] Initial condition: The security model is semi-honest security.

[0032] The specific implementation steps of the invention are as follows:

[0033] (1) Preparation stage

[0034] 1. Both parties agreed that the extractor used is a 5-layer neural network, which is a convolution layer, a ReLU layer, an average pooling layer, a convolution layer, and a ReLU layer; the pyramid pooling network includes average pooling with 4 different pooling kernel dimensions, convolution with 4 different convolution kernels, and upsampling; the classifier is a 3-layer neural network, which is a fully connected layer, a ReLU layer, and a fully connected layer. At the same time, both parties jointly confirmed the parameters of the homomorphic encryption scheme. The user secretly chooses the private key for homomorphic encryption Publicly disclose the public key associated with it in from Random sampling is obtained, The coefficients of are sampled from a discrete Gaussian distribution with standard deviation σ.

[0035] 2. The user has smart meter readings T and randomly samples [T] from the same space. c As your own secret share, s =T-[T] c Sent to the server as its share of the secret.

[0036] (2) Privacy-preserving computation extractor

[0037] 3. When computing the convolutional layer, the client sends its secret share [T] c According to the convolution kernel size, it is divided into several matrices {T γ,α,β ]c}, and then use the public key pk to encrypt it into RLWE ciphertext {CT γ,α,β} and then sent to the server. The server also sends [T] S Divide into {[T γ,α,β ] S}, locally calculates RLWE homomorphic addition and multiplication based on its own convolution kernel parameter K, and then uses the extract operation in the RLWE scheme on the result to obtain the LWE ciphertext at a specific position, that is, the homomorphic ciphertext CT' of the final result. The server randomly samples [T'] from the plaintext space s As its own secret share, CT'-Enc([T'] s =Enc(T'-[T'] s ) is sent to the user. Finally, the user performs LWE homomorphic decryption to obtain [T'] c , and satisfies [T'] c +[T'] s =T'=Conv(T,K).

[0038] 4. When computing the ReLU layer, for each element x in the input tensor T, calculate ReLU(x) = DReLU(x) x. The secure comparison protocol is used to compute DReLU(x), and the selection protocol is used to implement DReLU(x) x.

[0039] 5. When calculating the average pooling layer, the secret share of the user's input tensor T is [T] c , the server's input secret share is [T] s Due to the linear nature of average pooling, both parties only need to perform average pooling on their respective secret shares according to the common pooling kernel and step size, and then add the elements in the corresponding pooling kernel area and divide them by the pooling kernel size. Finally, both parties obtain [T'] c and [T'] s , satisfying [T′] c +[T′] s =T′=AvgPool(T).

[0040] (3) Privacy-preserving computational pyramid pooling module

[0041] 6. First, you need to calculate the average pooling of 4 different pooling kernels. Similar to step 5, when the input tensor size is W×H×C, use W×H, (W / 2)×(H / 2), (W / 3)×(H / 3), and (W / 6)×(H / 6) as four different pooling kernel sizes, and get 4 pooling results of sizes 1×1×C, 2×2×C, 3×3×C, and 6×6×C respectively.

[0042] 7. For the four pooling results obtained in step 6, apply four 1×1×C convolution kernels to perform convolution operations similar to those in step 3. Both parties obtain the additive secret shares of the four convolution results, with sizes of 1×1, 2×2, 3×3, and 4×4 respectively.

[0043] 8. Upsampling uses bilinear interpolation to upscale the smaller 2D vector to a W×H 2D vector of the same size as the original feature map. When f(x1,y1),f(x1,y2),f(x2,y1),f(x2,y2) are known, the value f(x,y) of a point (x,y) within the region formed by the four points is calculated as follows:

[0044]

[0045] where f(x i ,y j ),f(x,y) is the added secret share of both parties, x,y,x i ,y i In order to disclose the plaintext coordinate values, bilinear interpolation is also a linear operation. Both parties perform bilinear interpolation on their respective secret shares, and finally splice the obtained upsampling results with the original feature map.

[0046] (4) Privacy-preserving computing classifier

[0047] 9. When computing the fully connected layer, both parties hold the additive secret share [v] of the input tensor v c ,[v] s , the server holds the model parameters W,b, and jointly calculates u=W·v+b. First, the user sends its own secret share [v] c The RLWE ciphertext CT is obtained by homomorphic encryption with the public key pk and sent to the server. The server combines its own secret share [v] s And the model parameters are calculated homomorphically to obtain Enc(W·v), and then r is randomly sampled from the plaintext space, and [u] s = r + b mod p as the additive secret share of the output result held by oneself, calculate Enc(W·v)-Enc(r)=Enc(W·vr), then use the extract operation in the RLWE scheme on the calculation result to obtain the LWE ciphertext at a specific position and send it to the user end. The user performs LWE decryption on the received ciphertext to obtain W·vr as its own secret share [u] c .

[0048] 10. The calculation of ReLU in the classifier is the same as step 4.

[0049] 11. The calculation of the last fully connected layer is the same as step 9. In the end, both parties hold the added secret share of the final prediction result.

[0050] (5) Reduction stage

[0051] 12. The server will share the final result value [y] s Sent to the user, the user will share it with the shared value [y] c Add them together to get the final prediction result y.

Claims

1. A non-invasive load monitoring method under privacy protection in a smart grid, characterized by: The method involves two parties, namely a user and a server. The user holds the smart meter readings, and the server has the PSPNet model parameters. Both parties implement the following steps. Finally, the user uses the server's model parameters to perform neural network prediction to obtain a prediction result, but does not obtain the server's model parameters. The server cannot obtain information about the user input; Preparation phase: In the preparation phase, both parties determine the number of extractor layers N based on the public network. E , Pyramid Pooling layer number N P , Classifier layer number N C , the calculation type of each layer and the related parameters of the homomorphic algorithm, where N E 、N P 、N C are all natural numbers; the user then adds its own input to the secret share and sends one share to the server; Privacy-preserving computation extractor: Both parties have N in the extractor E The privacy protection calculation is performed layer by layer, and finally both parties obtain the added secret share of the extracted feature map; Privacy protection calculation pyramid pooling module: Both parties have N in the pyramid pooling module P The privacy protection calculation is performed layer by layer, where [T] c and [T] s Denote the secret shares of T held by the client and the server, respectively, satisfying T = [T] C +[T] S The same applies when T is replaced with other values: (1) Average pooling layer: Four pooling operations with different pooling area sizes are required. Both parties in the average pooling layer locally average pool their secret shares, and ultimately both parties possess the added secret share of the pooling result. (2) Convolutional layer: Convolution operation is performed on the four pooling results in the previous layer with the corresponding convolution kernel with a length and width of 1; the user end sends its own secret share [T] c After encryption using the RLWE homomorphic encryption algorithm, we get Enc([T] c ) is sent to the server, which uses the RLWE homomorphic algorithm to calculate the convolution and obtain the encrypted value Enc(T′) of the convolution result. Then, [T′] is randomly sampled from the plaintext space. s As its own secret share, Enc(T′)-Enc([T′] s ) = Enc(T′-[T′] s ) is extracted to obtain the LWE ciphertext and sent to the user end, and finally the user end performs LWE decryption to obtain [T′] c =T′-[T′] s As one's own secret share; (3) Upsampling layer: Upsamples the features of four different dimensions and expands them to the same dimension as the original feature map. Both parties can locally perform bilinear interpolation on their secret shares. Privacy-preserving computation classifier: Both parties have N C The privacy protection calculation is performed layer by layer, and both parties obtain the additive secret share of the final result; Restoration phase: The server sends the shared value of the final result to the client, and the client adds the two shared values ​​to obtain the final prediction result.

Images