Infectious disease flow regulation track generation system considering privacy protection
By generating an adversarial network and attention mechanism, combined with style migration technology, reconstructing the infectious disease epidemic regulation trajectory, the problem of security and availability balance in the privacy protection of infectious disease epidemic regulation trajectory is solved, and effective protection of private information and non-private information is achieved. Retention of
Patent Information
- Application Number
- CN202411022183.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-08
AI Technical Summary
The prior art has failed to effectively balance security and availability in the privacy protection of infectious disease epidemic regulation trajectory, and has failed to fully tap into space-time information, resulting in poor privacy protection effects.
Adopt an infectious disease epidemic regulation trajectory generation system based on a generative adversarial network. By inputting selectors, generators and discriminators, the attention mechanism and style migration components are used to reconstruct trajectories similar to the original trajectory, eliminate private information, and retain non-private information. .
It realizes effective protection of private information in the infectious disease epidemic regulation trajectory, while retaining non-private information, improving the availability and security of the trajectory.
Smart Images

Figure CN118919094A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of privacy protection technology, and more specifically to an infectious disease epidemiological trajectory generation system that takes privacy protection into consideration. Background Art
[0002] During an infectious disease outbreak, epidemiological surveys are one of the key tools used by public health departments to control the spread of infectious diseases. By tracking the activities and contact history of infected people, people who may be infected can be identified and appropriate isolation measures can be taken to prevent further spread. Although this data is crucial for infectious disease control, there is also a serious risk of privacy leakage in the process of infectious disease surveys.
[0003] At present, the privacy protection methods of infectious disease epidemiological trajectories mainly focus on how to more effectively prevent various attacks from attackers, but do not consider whether the availability of infectious disease epidemiological trajectories in downstream tasks is retained after the infectious disease epidemiological trajectories are modified. Such protection methods are relatively coarse-grained.
[0004] Secondly, although many privacy protection ideas from other fields have been transferred to the privacy protection of infectious disease epidemiological trajectories, these methods have not been able to fully exploit the spatiotemporal information unique to infectious disease epidemiological trajectories, and therefore will have some impact on the privacy protection of infectious disease epidemiological trajectories.
[0005] Finally, although some researchers have begun to try to use deep learning models to protect the privacy of infectious disease epidemiological trajectories, they have not been able to strike a good balance between the security and usability of the privacy of infectious disease epidemiological trajectories.
[0006] Therefore, how to remove private information from infectious disease epidemiological trajectories and retain non-private information to simultaneously ensure security and availability is an urgent problem that technicians in this field need to solve. Summary of the invention
[0007] In view of this, the present invention provides an infectious disease epidemiological trajectory generation system that takes privacy protection into consideration, which can reconstruct the original trajectory using a sampling trajectory similar to the original trajectory, and can eliminate the privacy information in the trajectory during the reconstruction process and retain the non-privacy information in the trajectory.
[0008] In order to achieve the above object, the present invention adopts the following technical solution:
[0009] A privacy-preserving infectious disease epidemiological trajectory generation system, including an input selector, a generator, and a discriminator:
[0010] The input selector is used to extract a plurality of sampled trajectories that meet preset similarity conditions with the current trajectory from a pre-constructed trajectory data set; and input the current trajectory and the sampled trajectories into the generator;
[0011] The generator is used to fuse the multiple sampling trajectories according to the current trajectory through the attention mechanism to obtain a fused trajectory; and is used to generate a final infectious disease epidemiological survey trajectory according to the fused trajectory;
[0012] The discriminator is used to determine the categories of the final infectious disease epidemiological survey trajectory and the current trajectory.
[0013] Preferably, the input selector comprises a trajectory encoding layer, a similarity measurement module and a selection module;
[0014] The trajectory encoding layer is used to encode the current trajectory to obtain a first vector; encode the trajectories in the trajectory data set to obtain a plurality of second vectors;
[0015] The similarity measurement module is used to calculate the similarity between each second vector and the first vector respectively;
[0016] The selection module is used to filter out the sampling tracks according to the similarity.
[0017] Preferably, the generator is further configured to transfer the style of the current track to the fused track by using an adaptive instance normalization operation.
[0018] Preferably, the generator comprises a preprocessing module, a first embedding layer, an attention mechanism module, a noise layer, a first adaptive instance normalization layer, a first LSTM layer, an output decoding layer and a second adaptive instance normalization layer connected in sequence;
[0019] The preprocessing layer is used to encode multiple infectious disease epidemiological survey trajectory information to generate feature vectors; the infectious disease epidemiological survey trajectory information includes location information, time information and functional category;
[0020] The first embedding layer is used to extract information from various feature vectors using a fully connected neural network, and concatenate the information extraction results to obtain an overall embedding vector, which is then input into the attention mechanism module;
[0021] The attention mechanism module fuses the overall embedding vector corresponding to the sampled trajectory according to the overall embedding vector corresponding to the current trajectory to obtain a fused trajectory;
[0022] The fused trajectory is input into the first adaptive instance normalization layer after random noise is added to the noise layer;
[0023] The first adaptive instance normalization layer transfers the style of the current trajectory to the infectious disease epidemiological survey trajectory after adding noise, and inputs the normalization result to the first LSTM layer;
[0024] The first LSTM layer is used to extract the time series in the input infectious disease epidemiological survey trajectory and output the vector corresponding to each time step;
[0025] The vector output by the first LSTM layer passes through the output decoding layer and is then style-transferred again by the second adaptive instance normalization layer to generate the final infectious disease epidemiological survey trajectory.
[0026] Preferably, the embedding layer works as follows:
[0027]
[0028] in, Respectively indicate location information Time of week information Hour Information and location point function information The corresponding representation vector; W and b represent the parameters to be learned of the corresponding fully connected neural network.
[0029] Preferably, the calculation process of the attention mechanism module is specifically as follows:
[0030]
[0031] Among them, h i Represents the vector representation of the i-th time step of the newly generated infectious disease epidemiological survey trajectory, The vector representation of the i-th time step of the current trajectory, and They represent the vector representation of different sampling trajectories at the i-th time step respectively; the att(·, ·) function represents the weight calculation process, and the weight calculation process is:
[0032]
[0033] Among them, q is 1 or 2, representing the two input sampling trajectories.
[0034] Preferably, the discriminator comprises a second embedding layer, a second LSTM layer and a classification output layer;
[0035] The second embedding layer is used to receive the current trajectory as a real sample or the final infectious disease epidemiological survey trajectory as a false sample to construct a sample representation vector;
[0036] The second LSTM layer receives the sample representation vector and inputs the vector output by the final hidden layer into the classification output layer;
[0037] The classification output layer is used to calculate the final judgment result.
[0038] Preferably, an optimization module is also included, and the optimization module is used to calculate the model loss according to a preset loss function and optimize the model parameters.
[0039] Preferably, the loss function includes:
[0040]
[0041] Among them, L co Represents the original trajectory Generate infectious disease epidemiological trajectories The mean square error loss on the position information, L d Represents the original trajectory Generate infectious disease epidemiological trajectories The cross entropy loss on the day of the week, L h Represents the original trajectory Generate infectious disease epidemiological trajectories The cross entropy loss on the hourly information, L ca Represents the original trajectory Generate infectious disease epidemiological trajectories In the cross entropy loss on the category, α, β, η and λ represent four hyperparameters used to adjust the size of the constraint term.
[0042] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a system for generating infectious disease epidemic investigation trajectories that takes privacy protection into consideration. It is based on a generative adversarial network architecture, integrates attention mechanisms and style transfer components, and uses two infectious disease epidemic investigation trajectories similar to the original trajectory to reconstruct the original trajectory. During the reconstruction process, it can eliminate privacy information in the infectious disease epidemic investigation trajectory and retain non-privacy information in the infectious disease epidemic investigation trajectory. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] 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 or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0044] Figure 1 A schematic diagram of the structure of an infectious disease epidemiological investigation trajectory generation system considering privacy protection provided by the present invention;
[0045] Figure 2 A schematic diagram of the structure of an input selector provided in an embodiment of the present invention;
[0046] Figure 3 A schematic diagram of the structure of a generator provided in an embodiment of the present invention;
[0047] Figure 4 Schematic diagram of the discriminator structure provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0048] 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.
[0049] like Figure 1 The embodiment of the present invention discloses a system for generating infectious disease epidemiological investigation trajectories with consideration of privacy protection, including an input selector, a generator, and a discriminator. The input selector is responsible for generating an infectious disease epidemiological investigation trajectory according to the current trajectory. And the overall trajectory dataset The generator is responsible for mining and modeling the non-user privacy information in the multiple input infectious disease epidemiological survey trajectories, and integrating them into a new infectious disease epidemiological survey trajectory. The discriminator is responsible for learning to distinguish real infectious disease epidemiological survey trajectory samples And the fake infectious disease epidemiological survey trajectory samples generated by the generator These three parts are interrelated and work closely together to effectively protect the privacy of infectious disease epidemiological investigation trajectories and effectively retain the non-privacy information of infectious disease epidemiological investigation trajectories.
[0050] The following is a detailed description of these three structures.
[0051] like Figure 2 The input selector is used to extract multiple sampled trajectories that meet preset similarity conditions with the current trajectory from the pre-constructed trajectory data set, and input the current trajectory and the sampled trajectories into the generator. The input selector includes a trajectory encoding layer, a similarity measurement module and a selection module; the trajectory encoding layer is used to encode the current trajectory to obtain a first vector; the trajectories in the trajectory data set are encoded to obtain multiple second vectors; the similarity measurement module is used to calculate the similarity between each second vector and the first vector respectively; the selection module is used to screen out sampled trajectories with higher similarity.
[0052] Specifically, the encoder part of the trained infectious disease epidemiological survey trajectory similarity calculation model can be used as the trajectory encoding layer to represent the input infectious disease epidemiological survey trajectory as a vector. In order to reduce the impact of sampling noise at the infectious disease epidemiological survey trajectory location, we use the Geohash algorithm to grid all infectious disease epidemiological survey trajectories and obtain the grid sequence corresponding to each infectious disease epidemiological survey trajectory. This process is the same as the processing in the infectious disease epidemiological survey trajectory similarity calculation model. Suppose the overall trajectory dataset is The number of infectious disease epidemiological survey trajectories in is M, then this process can be formally expressed as:
[0053] r i =encoder(Geohash(T i )), i = 0, 1, ..., M-1,
[0054] The encoder() represents the trained trajectory encoding layer, and r i It represents the vector representation corresponding to the i-th infectious disease epidemiological trajectory in the trajectory dataset.
[0055] So far, the trajectory of infectious disease epidemiology The spatiotemporal information in can be represented by the vector r i On this basis, the similarity between the epidemiological survey trajectories of infectious diseases can be expressed as:
[0056]
[0057] So, for the current trajectory We can get the similarity Sim(0,i) between it and each infectious disease epidemiological survey trajectory in the trajectory dataset, and then select the two infectious disease epidemiological survey trajectories with the highest similarity. and The infectious disease epidemiological survey trajectory selected by the input selector. It should be noted that in this process, we will not consider all Infectious disease epidemiological investigation trajectory of the same user.
[0058] Although there is no very accurate method to define or represent the privacy information contained in the epidemic survey trajectory, the privacy information of a user is basically hidden in his or her personalized travel mode, and will not appear in the travel mode of other users in the same way. Therefore, by inputting the selector, we hope to find the private information that is consistent with the current trajectory. The most similar infectious disease epidemiological survey trajectories, but not belonging to the same user and These infectious disease survey trajectories are used as input to the generator, guiding the generator to learn the common behavior patterns of people traveling as much as possible, and preventing the generator from reconstructing user privacy information. The entire input selector process can be expressed in the following pseudo code:
[0059] After selecting the input selector, you can get the current trajectory The most similar infectious disease epidemiological survey trajectories of other users and They will be used together as inputs to the model generator. This embodiment requires that the two infectious disease epidemiological survey trajectories must come from other users because it avoids using the infectious disease epidemiological survey trajectories of the same user and avoids introducing similar privacy information contained in other infectious disease epidemiological survey trajectories of the same user.
[0060] The generator uses the attention mechanism to fuse multiple sampled trajectories according to the current trajectory to obtain a fused trajectory; and to generate the final infectious disease epidemiological survey trajectory according to the fused trajectory. However, in the infectious disease epidemiology trajectory-infectious disease epidemiology trajectory attention mechanism, The relevant information is only used as a query vector to guide the weight calculation, and does not really participate in the generation of new infectious disease epidemiological survey trajectories. Therefore, the original trajectory The user privacy information in will not be learned and modeled by the generator, thus protecting the user's privacy.
[0061] In order to further implement the above technical solution, the generator is also used to transfer the style of the current trajectory to the fused trajectory by using an adaptive instance normalization operation.
[0062] As an implementation, style transfer can be achieved through an adaptive instance normalization layer. The adaptive instance normalization layer (Adaptive Instance Normalization, AdaIN) can adjust the mean and variance of the content features to values that match the style features at the feature level. This method is often used to implement style transfer tasks, where the goal is to apply the style of one image to another content image while maintaining the structure of the content image. The core idea of AdaIN is to calculate the mean and standard deviation of the feature map of the style image, and then use these statistics to adjust the corresponding statistics of the feature map of the content image.
[0063] In one embodiment, if Figure 3 The generator includes a preprocessing module, a first embedding layer, an attention mechanism module, a noise layer, a first adaptive instance normalization layer, a first LSTM layer, an output decoding layer, and a second adaptive instance normalization layer connected in sequence.
[0064] The preprocessing layer is used to encode various infectious disease epidemiological trajectory information and generate feature vectors; the infectious disease epidemiological trajectory information includes location information, time information and functional category; the first embedding layer is used to extract information from various feature vectors using a fully connected neural network, and the information extraction results are spliced to obtain the overall embedding vector and then input into the attention mechanism module; the attention mechanism module fuses the overall embedding vector corresponding to the sampling trajectory according to the overall embedding vector corresponding to the current trajectory to obtain a fused trajectory; the fused trajectory is added with random noise through the noise layer and input into the first adaptive instance normalization layer; the first adaptive instance normalization layer transfers the style of the current trajectory to the infectious disease epidemiological trajectory after noise is added, and inputs the normalization result into the first LSTM layer; the first LSTM layer is used to extract the time series in the input infectious disease epidemiological trajectory and output the vector corresponding to each time step; the vector output by the first LSTM layer passes through the output decoding layer and is again style-transferred by the second adaptive instance normalization layer to generate the final infectious disease epidemiological trajectory.
[0065] The input received by the generator is no longer completely random Gaussian noise, but the current real infectious disease epidemiological survey trajectory. The two infectious disease survey trajectories of other users that are most similar to the infectious disease survey trajectory and These three infectious disease epidemiological survey trajectories serve as the input of the generator to guide the generation process of new infectious disease epidemiological survey trajectories. In terms of specific functions, and As a sampling trajectory generated by the epidemiological survey trajectory of new infectious diseases, It is used to guide the trajectory of infectious disease epidemiology and impose certain constraints. Represents the position point corresponding to the i-th time step of the j-th input infectious disease epidemiological investigation trajectory, that is, the subscript represents the infectious disease epidemiological investigation trajectory number, and the superscript represents the time step number.
[0066] In the preprocessing process, the four types of information of each location point in the infectious disease epidemiological survey trajectory need to be vector-encoded. For location information (such as longitude and latitude), we use For the days of the week, we use To represent its corresponding one-hot encoding. For the number of hours, we use To represent its corresponding one-hot encoding. For the functional category of the location point, we use To represent its corresponding one-hot encoding. Therefore, the infectious disease epidemiological investigation trajectory input of the generator is the vector of these four types of information concatenated in sequence, namely:
[0067]
[0068] The three input infectious disease epidemiological survey trajectories after input preprocessing will be input into the embedding layer for further information mining. In the embedding layer, there is a fully connected neural network to extract the four types of information of the location point, that is, there are a total of four parallel fully connected neural networks, which map the input vector to a higher-dimensional vector. Similar to the discriminator, we still use the Leaky Relu activation function in the embedding layer, and the parameters to be learned of the fully connected neural network are represented by W and b. Therefore, the working mode of the embedding layer of the generator can be expressed as:
[0069]
[0070] After being processed by the embedding layer, the vector representations corresponding to different categories of information are still spliced together in order to obtain the spliced overall embedding vector The specific process is expressed as:
[0071]
[0072] Next, in the attention mechanism layer, the input is the embedding vectors corresponding to the three infectious disease survey trajectories (one current trajectory and two sampling trajectories) and In this layer, and Will be in Under the guidance of , through the attention mechanism, the new infectious disease epidemiological survey trajectory representation vector v = [v 0 , v 1 , ..., v m-1 ].
[0073] The input of the attention mechanism is three vector representations, called query, key, and value. The query and key are multiplied by matrix multiplication, and then scaled according to the vector dimension, and then processed by the Softmax function to obtain the weight value α, or probability value. The calculated weight value α is then used to weight the vectors in the value vector to obtain the final output. From a holistic perspective, the whole process can be described as calculating the weight according to the query vector under the guidance of the key value, and then weighted summing the value vector. In general time series, the key and value vectors are both vector representations of the entire sequence, and the query vector is the vector representation corresponding to the current time step. Therefore, the attention mechanism can find the most relevant information for the current time step in the entire time series and perform weighted summation.
[0074] In this embodiment, the present invention does not use the attention mechanism in the time series dimension, but the attention mechanism between different infectious disease epidemiological survey trajectories. Embedding vector for each time step are all used as query vectors, and the embedding vectors of each time step of similar infectious disease epidemiological survey trajectories are and It serves as both the key vector and the value vector. At each time step, we want to Under the guidance of and Perform weighted summation to obtain a vector representation v of a new infectious disease epidemiological survey trajectory at this time step i , the calculation process of the attention mechanism module is as follows:
[0075]
[0076] Among them, h i Represents the vector representation of the i-th time step of the newly generated infectious disease epidemiological survey trajectory, represents the vector representation of the i-th time step of the current trajectory, and They represent the vector representation of different sampling trajectories at the i-th time step respectively; the att(·, ·) function represents the weight calculation process, and the weight calculation process is:
[0077]
[0078] Among them, q is 1 or 2, representing the two input sampling trajectories.
[0079] Next, in the fused infectious disease epidemiological survey trajectory representation vector v, each time step will be spliced with a random noise z to obtain a new vector l i This is to make the generator more generalizable in the process of generating samples, and also to make the training process more robust and avoid the generation mode collapse. This process can be formally expressed as:
[0080] l i =v i ||z.
[0081] So far, in the current trajectory, that is, the original trajectory Under the guidance of , a new infectious disease epidemiological survey trajectory was fused through the sampling trajectory. However, since there are still many differences between the input sampling trajectory and the original trajectory, and they have different styles, the new infectious disease epidemiological survey trajectory at this time still has a certain difference in style from the original trajectory.
[0082] Furthermore, in order to reduce the difference in style, the first adaptive instance normalization layer is used to perform style transfer. This process is formally expressed as:
[0083]
[0084] where r i is the representation of the i-th time step in the vector representation obtained after the original trajectory in the input selector passes through the encoder, μ() and σ() represent the mean and variance of the vector respectively. It can be seen that the AdaIN layer can convert each time step l of the newly generated infectious disease epidemiological trajectory vector representation into i The mean and variance of the original trajectory vector represent each time step r i The mean and variance of the trajectory are aligned to achieve the purpose of migrating the style of the original trajectory to the new infectious disease epidemiological trajectory.
[0085] Next, the vector representation s i It will be input into a Long Short-Term Memory (LSTM) layer to extract the time series information of the newly generated infectious disease epidemiological survey trajectory. This process can be expressed as:
[0086] o i =LSTM(s i , o i-1 ),
[0087] Where LSTM() represents each processing unit in LSTM, o i is the vector output corresponding to the i-th time step after LSTM calculation, that is, o=[o 0 , o 1 , ..., o m-1 ].
[0088] Next, the vector output by the LSTM layer will be input into the output decoding layer to convert the vector representation into the initial infectious disease epidemiological trajectory information. The output decoding layer and the embedding layer at the input can be regarded as reciprocal operations, that is, the output decoding layer is also composed of four fully connected layers, which process the four types of information of the location points respectively. Slightly different from the embedding layer at the input, in the output decoding layer, the decoding layer of the location information has two fully connected layers, and the activation functions used in the decoding layers corresponding to different information are different. For floating-point information such as location information, we use Leaky Relu activation function and Tanh activation function. For the other three categories of information, we use Softmax activation function. The calculation process of the output decoding layer can be expressed as:
[0089]
[0090] Among them, we use W′ and b′ to represent the parameters to be learned in the fully connected layer in order to distinguish it from the embedding layer at the input. Its parameters are also learned independently and do not share weights with the embedding layer of the input.
[0091] In order to make the location information of the generated infectious disease epidemiological trajectory closer to the original trajectory, we decoded the location information Use the AdaIN layer again, this time the style information comes from the original trajectory input by the generator The position information of the corresponding time step in The formal expression is:
[0092]
[0093] At this point, the newly generated infectious disease epidemiological survey trajectory vector after fusion is:
[0094]
[0095] By decoding these four types of information, the new infectious disease epidemiological survey trajectory can be generated. At this point, the infectious disease epidemiological trajectory generation process of the entire generator is completed.
[0096] like Figure 4 ,In one embodiment, the discriminator includes a second embedding layer, a second LSTM layer, and a classification output layer;
[0097] The second embedding layer is used to receive the current trajectory as a real sample or the final infectious disease epidemiological survey trajectory as a false sample to construct a sample representation vector; the second LSTM layer receives the sample representation vector and inputs the vector output by the final hidden layer into the classification output layer; the classification output layer is used to calculate the final judgment result.
[0098] The input received by the discriminator is an infectious disease epidemiological survey trajectory, which can be a real infectious disease epidemiological survey trajectory. It can also be a false infectious disease epidemiological survey trajectory generated by the generator For the convenience of description, we will use Instead, the infectious disease epidemiological investigation trajectory length is represented by m, and the location point corresponding to the ith time step is represented by . Generally speaking, when sampling location points, in addition to the necessary longitude and latitude and timestamp, some functional attribute information of the location points is generally involved, such as the function of the location or the classification of the location. Making better use of this information can improve the model's ability to mine travel patterns. Based on this, we focus on the following four attributes of each location point: coordinates, the day of the week corresponding to the timestamp, the hours corresponding to the timestamp, and the functional type of the location point. For coordinates, in most cases, longitude and latitude are used to represent them, and they may also be gridded. For the sake of loss of generality, we use To represent the coordinate information of the location point. For the day of the week, since a week has seven days, we use a one-hot vector to encode it, using For the number of hours, since a day has 24 hours, we can still encode it using a one-hot vector, using For the functional types of location points, we can use the trajectory dataset The characteristics of the self-division are used. Because it represents the category, it is also encoded with a one-hot vector. Therefore, the position point x i In fact, it is composed of the vector representations of these four types of information spliced in order, namely:
[0099]
[0100] Among them, || represents the vector concatenation operation.
[0101] The input infectious disease epidemiological investigation trajectory after the above preprocessing is input into the second embedding layer of the discriminator to further extract and mine the infectious disease epidemiological investigation trajectory information. In order to better mine the above four types of information, the discriminator uses four different embedding layer networks to embed the four types of information respectively. Each embedding layer network is essentially a layer of fully connected neural network, which maps the corresponding vector to a higher-dimensional vector. The fully connected neural network is essentially a linear transformation of the vector plus the nonlinear processing of the activation function. The parameters to be learned are the weight matrix W and the scalar b of the linear transformation. Here, the activation function we use is the Leaky Relu activation function, and its function definition is as follows:
[0102]
[0103] Compared with activation functions such as Sigmoid or Tanh, the Leaky Relu activation function can better solve the gradient vanishing problem, because when x is greater than or equal to 0, the gradient is always 1 and there will be no saturation. Compared with activation functions such as Relu, the Leaky Relu activation function can avoid the problem of neuron death and increase the robustness of the model, because when x is less than 0, the gradient still exists, but the slope becomes smaller, so neurons can also be activated, with a more flexible activation range. Therefore, the embedding layer part of the discriminator can be expressed as:
[0104]
[0105] After being processed by the embedding layer, the embedding vectors corresponding to different categories of information are also concatenated in order to obtain the concatenated embedding vectors. This process can be expressed as
[0106]
[0107] Next, in order to further extract the spatiotemporal information in the infectious disease epidemiological trajectory sequence, the discriminator uses a long short-term memory network (LSTM) to model the time series information. The LSTM network is essentially a recurrent neural network that is suitable for processing time series information and allows information to be transmitted in time order. In a recurrent neural network, all processing unit structures and parameters are exactly the same, so the recurrent neural network can be understood as multiple copies of the same network.
[0108] The LSTM network mainly solves the problem of "long-term dependency" between sequence information. Each processing unit in the LSTM network can be divided into three gates: forget gate, input gate and output gate. Through the fine control of these gated units, the input of each time step will be remembered, and the information from a long time ago can also be smoothly passed on.
[0109] The forget gate mainly uses a Sigmoid function to calculate the current time step x. t and the hidden state h of the previous time step t-1 To decide how much information from the previous step needs to be forgotten, the formula is expressed as
[0110] f t =σ(W f ·[h t-1 , x t ]+b f ),
[0111] Among them, σ() represents the Sigmoid function, W f and b f represents the learnable parameters in the forget gate.
[0112] The input gate uses the Sigmoid function and the tanh function to determine how much information in the input of the current time step needs to be recorded. The formula is expressed as
[0113] i t =σ(W i ·[h t-1 , x t ]+b i ),
[0114]
[0115] Where σ() represents the Sigmoid function, W i , b i , W c and b cRepresents the learnable parameters in the input gate. On this basis, we can use the forget gate and the input gate to update the cell state of the current processing unit. The formula is expressed as
[0116]
[0117] Finally, we use the output gate to decide what hidden vector to output in the next time step, which can be expressed as
[0118] o t =σ(W o ·[h t-1 , x t ]+b o ),
[0119] h t =o t *tanh(C t ),
[0120] Where σ() represents the Sigmoid function, W o and b o represents the learnable parameters in the output gate.
[0121] In the discriminator of the model, we embed the vector v i As the input of the i-th time step of the LSTM layer, this process can be formally expressed as:
[0122] u i =LSTM(v i ,u i-1 )
[0123] Where LSTM() represents the processing unit in each LSTM network layer, which receives the input vector v at that time step i and the hidden vector u output at the previous time step i-1 , and output the hidden vector u to the next time step i Finally, we take the hidden vector outputted at the last time step as the output of the LSTM layer, denoted by u, i.e. u = u m-1 .
[0124] Finally, the vector u is input to the output layer to get the final result. The output layer is also a fully connected layer, but no activation function is required. In the original structure of the generative adversarial network, the final output layer in the discriminator needs to use the Sigmoid function to obtain the classification probability. However, our discriminator refers to the improvement made by WGAN and no longer calculates the probability, so there is no need to use the Sigmoid function. This step can be formally expressed as
[0125] y=Wd ·u+b d ,
[0126] Where W d and b d represents the learnable parameters in the fully connected layer, and y represents the final result given by the discriminator. For real samples, y should be as large as possible, while for fake samples generated by the generator, y should be as small as possible.
[0127] In order to further implement the above technical solution, the present invention improves the loss function during training, so as to better generate infectious disease epidemiological trajectories that can accurately protect user privacy.
[0128] Specifically, the present invention uses the loss function optimized by Wasserstein GAN as the basis, and then adds some constraints on generating infectious disease epidemiological trajectory information to obtain the following loss function:
[0129]
[0130] Among them, D represents the discriminator, G represents the generator, and L co Represents the original trajectory Generate infectious disease epidemiological trajectories The mean square error loss on the position information, L d Represents the original trajectory Generate infectious disease epidemiological trajectories The cross entropy loss on the day of the week, L h Represents the original trajectory Generate infectious disease epidemiological trajectories The cross entropy loss on the hourly information, L ca Represents the original trajectory Generate infectious disease epidemiological trajectories Cross entropy loss on the category. α, β, η and λ represent four hyperparameters used to adjust the size of the constraint terms.
[0131] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0132] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A system for generating infectious disease epidemiological survey trajectories with consideration of privacy protection, characterized in that: Including input selector, generator and discriminator: The input selector is used to extract a plurality of sampled trajectories that meet preset similarity conditions with the current trajectory from a pre-constructed trajectory data set; and input the current trajectory and the sampled trajectories into the generator; The generator is used to fuse the multiple sampled trajectories according to the current trajectory through an attention mechanism to obtain a fused trajectory; Used to generate a final infectious disease epidemiological survey trajectory based on the fusion trajectory; The discriminator is used to determine the categories of the final infectious disease epidemiological survey trajectory and the current trajectory.
2. According to the privacy protection-based infectious disease epidemiological survey trajectory generation system of claim 1, it is characterized in that: The input selector includes a trajectory encoding layer, a similarity measurement module and a selection module; The trajectory encoding layer is used to encode the current trajectory to obtain a first vector; encode the trajectories in the trajectory data set to obtain a plurality of second vectors; The similarity measurement module is used to calculate the similarity between each second vector and the first vector respectively; The selection module is used to filter out the sampling tracks according to the similarity.
3. According to the privacy protection-based infectious disease epidemiological survey trajectory generation system of claim 1, it is characterized in that: The generator is further configured to transfer the style of the current track to the fused track by using an adaptive instance normalization operation.
4. According to the privacy protection-based infectious disease epidemiological survey trajectory generation system of claim 3, it is characterized in that: The generator includes a preprocessing module, a first embedding layer, an attention mechanism module, a noise layer, a first adaptive instance normalization layer, a first LSTM layer, an output decoding layer and a second adaptive instance normalization layer connected in sequence; The preprocessing layer is used to encode multiple infectious disease epidemiological survey trajectory information to generate feature vectors; the infectious disease epidemiological survey trajectory information includes location information, time information and functional category; The first embedding layer is used to extract information from various feature vectors using a fully connected neural network, and concatenate the information extraction results to obtain an overall embedding vector, which is then input into the attention mechanism module; The attention mechanism module fuses the overall embedding vector corresponding to the sampled trajectory according to the overall embedding vector corresponding to the current trajectory to obtain a fused trajectory; The fused trajectory is input into the first adaptive instance normalization layer after random noise is added to the noise layer; The first adaptive instance normalization layer transfers the style of the current trajectory to the infectious disease epidemiological survey trajectory after adding noise, and inputs the normalization result to the first LSTM layer; The first LSTM layer is used to extract the time series in the input infectious disease epidemiological survey trajectory and output the vector corresponding to each time step; The vector output by the first LSTM layer passes through the output decoding layer and is then style-transferred again by the second adaptive instance normalization layer to generate the final infectious disease epidemiological survey trajectory.
5. According to the privacy protection-based infectious disease epidemiological survey trajectory generation system of claim 4, it is characterized in that: The embedding layer works as follows: in, Respectively indicate location information Time of week information Hour Information and location point function information The corresponding representation vector; W and b represent the parameters to be learned of the corresponding fully connected neural network.
6. According to the privacy protection-based infectious disease epidemiological survey trajectory generation system of claim 4, it is characterized in that: The calculation process of the attention mechanism module is specifically as follows: Among them, h i Represents the vector representation of the i-th time step of the newly generated infectious disease epidemiological survey trajectory, The vector representation of the i-th time step of the current trajectory, and They represent the vector representation of different sampling trajectories at the i-th time step respectively; the att(·, ·) function represents the weight calculation process, and the weight calculation process is: Among them, q is 1 or 2, representing the two input sampling trajectories.
7. According to the privacy protection-based infectious disease epidemiological survey trajectory generation system of claim 1, it is characterized in that: The discriminator includes a second embedding layer, a second LSTM layer and a classification output layer; The second embedding layer is used to receive the current trajectory as a real sample or the final infectious disease epidemiological survey trajectory as a false sample to construct a sample representation vector; The second LSTM layer receives the sample representation vector and inputs the vector output by the final hidden layer into the classification output layer; The classification output layer is used to calculate the final judgment result.
8. According to the privacy protection-based infectious disease epidemiological survey trajectory generation system of claim 1, it is characterized in that: It also includes an optimization module, which is used to calculate the model loss according to a preset loss function and optimize the model parameters.
9. The infectious disease epidemiological investigation trajectory generation system considering privacy protection according to claim 8 is characterized in that: The loss function includes: Among them, L co Represents the original trajectory Generate infectious disease epidemiological trajectories The mean square error loss on the position information, L d Represents the original trajectory Generate infectious disease epidemiological trajectories The cross entropy loss on the day of the week, L h Represents the original trajectory Generate infectious disease epidemiological trajectories The cross entropy loss on the hourly information, L ca Represents the original trajectory Generate infectious disease epidemiological trajectories In the cross entropy loss on the category, α, β, η and λ represent four hyperparameters used to adjust the size of the constraint term.
