A wireless positioning differential privacy federated learning method based on dynamic privacy budget

Abstract

A wireless positioning differential privacy federated learning method based on dynamic privacy budget comprises the following steps: (1) a terminal device preprocesses and differentially perturbs RSS fingerprint dataset data obtained from indoors, and sends the perturbed RSS data to a neighboring edge server; (2) the edge server aggregates the received RSS fingerprint data, and trains a local positioning submodel using the aggregated data; in each iteration process, the edge server uploads the processed local submodel parameters to a cloud server after dynamic privacy budget allocation using RDP technology; (3) the cloud server receives the submodel parameters shared by each edge server, updates global shared model parameters using a federated average optimization algorithm, and distributes the updated model parameters to each edge server for next round of iteration training until an optimal training model is obtained. The present application can obtain an indoor positioning model with higher precision and lower response time delay while protecting the data privacy of users and the parameter privacy of models.

Description

Technical Field

[0001] This invention belongs to the field of indoor positioning, and specifically relates to a method for protecting user data privacy and model parameter privacy when using location information for location services, specifically a wireless positioning differential privacy federated learning method based on dynamic privacy budget. Background Technology

[0002] With the development of network and mobile positioning technologies, more and more applications rely on location functionality. Outdoor positioning technology has matured, leading to a shift in focus towards indoor positioning. This has resulted in numerous technologies aimed at improving indoor positioning accuracy, including those based on infrared, ultrasonic, RFID, Wi-Fi, Bluetooth, sensor networks, ultra-wideband, electromagnetic signals, vision, and sound, as well as hybrid systems combining these technologies. While most of these technologies offer high levels of positioning accuracy, their applications have significant limitations, making it difficult to meet the requirements of indoor positioning and navigation. Fingerprint positioning technology based on Received Signal Strength (RSS) has become the mainstream trend for indoor positioning due to its advantages such as ease of implementation, low cost, low power consumption, and mature infrastructure. Faced with massive amounts of available RSS fingerprint data, some new studies have proposed using multi-sensor data, Channel State Information (CSI) data, and federated learning techniques to assist WiFi or Bluetooth fingerprints in indoor positioning in order to further improve indoor positioning accuracy and cope with the adverse effects of complex and variable indoor environments, wireless signal fading, and multipath effects. This has yielded higher positioning accuracy and robustness than traditional positioning methods and systems. However, in the process of providing indoor positioning technology to users, problems such as data processing and transmission delays, cloud server pressure, and network load have emerged. Deploying indoor positioning federated learning models to edge computing frameworks is an effective strategy to address the aforementioned problems of traditional cloud-based indoor positioning systems.

[0003] While edge computing environments and federated learning can effectively address many challenges in cloud architectures, the limited computing, communication, and storage resources of edge devices make it difficult to implement complex privacy protection strategies, thus privacy leaks remain a serious problem. Malicious attackers can exploit neural network parameters to obtain users' private information through techniques such as model inversion attacks and member inference attacks. Differential privacy techniques, by adding random noise that conforms to the data's distribution, can effectively solve the privacy leak problem in federated learning models during training and application, protecting user privacy. The amount of noise is controlled by the privacy budget parameter ε; the larger ε is, the less noise is added and the higher the data availability, and vice versa. Therefore, designing an efficient and dynamic privacy budget allocation strategy is currently a research hotspot and challenge in the field of deep learning privacy protection.

[0004] To address the privacy budget allocation problem in the training and application of federated learning models using differential privacy, many researchers have proposed improvements to this approach. Martin et al., focusing on privacy leakage during deep learning model training, proposed the DP-SGD algorithm based on (ε,δ)-differential techniques. This algorithm utilizes Mom-ents Account (MA) for refined analysis and tracking of privacy loss, achieving reliable model performance while maintaining fine-grained user privacy protection. However, it does not consider the impact of batch processing methods on privacy loss. To solve this problem, Yu et al. proposed a new dynamic privacy budget method based on centralized differential privacy, which improves model performance while providing stricter privacy protection. While these methods can achieve dynamic privacy budget allocation, they are still not directly applicable to highly complex and dynamic edge computing environments. They struggle to effectively address the dynamic adjustment and allocation problem of privacy budget in fingerprint indoor positioning privacy protection methods based on deep learning, a problem that has become one of the challenges hindering the rapid development and application of indoor positioning technology.

[0005] Given the high complexity and dynamism of edge computing environments, existing dynamic privacy budget allocation strategies are difficult to apply, prompting in-depth research in academia. Patent CN114462090A uses f-differential privacy technology based on hypothesis testing to allocate the privacy budget, enabling federated learning systems to maintain differential privacy while keeping accuracy within an acceptable threshold. However, this method does not consider real-time performance and cannot meet the demands of dynamic edge computing environments. Patent CN113642715A addresses the issue of excessive privacy budget consumption by considering noise addition to the optimal model after perturbing the gradient with correlation, thus improving privacy protection. However, this method allocates the privacy budget based on weight changes and cannot provide fine-grained control over differential privacy. Summary of the Invention

[0006] To address the above issues, this invention proposes a differential privacy federated learning method for wireless positioning based on dynamic privacy budget, which mainly solves two problems: 1. While protecting user data privacy and model parameter privacy, it achieves high indoor positioning model accuracy and low response latency; 2. It realizes fine-grained allocation of privacy budget during the indoor positioning model training process.

[0007] The technical solution adopted in this invention is as follows:

[0008] A differential privacy federated learning method for wireless positioning based on dynamic privacy budget is implemented according to the following steps:

[0009] Step A, Data Preprocessing and Differential Perturbation: The terminal device will preprocess and differentially perturb the Received Signal Strength (RSS) fingerprint dataset obtained from the indoor area, and then send the perturbed RSS fingerprint data to the nearby edge server. The data preprocessing involves using the Pearson Correlation Coefficient (PCC) to remove weakly correlated RSS fingerprint data, and then using (ε, δ)-differential privacy technology to differentially perturb the converted grayscale image. Finally, the perturbed grayscale image is sent to the edge server.

[0010] Step B, Local Sub-model Training: The edge server aggregates the received RSS fingerprint data and uses this aggregated data to train the local positioning sub-model. It uses Rényi Differential Privacy (RDP) technology to dynamically calculate the privacy budget loss and allocate the privacy budget, adaptively adjusts the degree of perturbation of the parameter gradient at the current stage, realizes fine-grained control of noisy data, and uploads the trained local sub-model parameters to the cloud server.

[0011] Step C, Model Parameter Aggregation: In each iteration, the cloud server receives model parameters shared by each edge server, updates the globally shared model parameters using the federated average optimization algorithm, and then distributes the aggregated model parameters to each edge server for the next iteration of training until the optimal training model is obtained.

[0012] Step A further includes:

[0013] A1. Divide the indoor area into N uniform small areas, each labeled as l. i , 1≤i≤N, K fingerprint data records R=[r1,r2,...,r K ], where r j= [r1, r2, ..., r m , l i ], 1≤j≤K indicates that the user is in region l i The j-th RSS fingerprint data collected by m APs;

[0014] A2. Calculate fingerprint data tag L K = [l1, l1, ..., l K ] and w n The PCC value c for 1≤n≤m, where w n This represents the nth column of the fingerprint dataset R. When |c|≤0.3, it indicates that the AP signal strength is unusable. Therefore, at this stage, we removed the fingerprint data column with |c|≤0.3.

[0015] A3. Calculate the Hadamard Product, HP, of the latest RSS fingerprint data record and c, and obtain the Hadamard moment HP; where HP = [HP1, HP2, ..., HP]. K ], for HP i After differential perturbation, it is converted into a grayscale image (img). i Finally, the perturbed grayscale image [img1, img2, ..., img] will be obtained. K Uploaded to the edge server.

[0016] Step B further includes:

[0017] The edge server receives a message from a nearby terminal device. i Disturbed RSS fingerprint image data Then, they are aggregated to form the dataset required for model training. N is the size of the aggregated dataset. express The corresponding labels, i≤j≤N, are then used with the dataset. Perform local sub-model training, and train the edge server E according to formula (1). i Optimize the local sub-model:

[0018]

[0019] Among them, f e (·,·) represents the model's prediction function. Let l represent the optimization objective of the model, and l(·,·) represent the optimization function of the model. Represents edge server E i Model parameters.

[0020] Step B, the local sub-model training process, further includes the following steps:

[0021] B1. Clipping the parameter gradient: Clip the parameter gradient g(img) i After cutting according to formula (2), we get g'(img) i ):

[0022]

[0023] Where C is the gradient clipping threshold, ||g(img) i )||2 is the gradient g(img) i The Euclidean distance of ).

[0024] B2. Add controllable differential private noise: Add controllable differential private noise g on the gradient according to formula (3). Gaussian (0, σ) 2 C 2 The new gradient g is obtained. t :

[0025]

[0026] Where B is the number of data points input into the model each time. For noise scale;

[0027] B3. Parameter Update: Update the model parameters using the perturbed parameter gradient according to formula (4) to enable the next iteration of training:

[0028]

[0029] in, For the updated model parameters, Here are the model parameters for the current stage, and α is the learning rate.

[0030] In step C, after the cloud server receives the model parameters uploaded by N edge servers, it aggregates the parameters according to formula (5) and then distributes them to each edge server for the next iteration of training.

[0031]

[0032] Among them, w global These are the aggregated model parameters. For edge server E i Uploaded model parameters.

[0033] This invention utilizes a combination of federated learning and (ε,δ)-differential privacy techniques to protect user data privacy and model parameter privacy during deep learning model training. It leverages existing Pearson correlation coefficients and differential privacy techniques for RSS data collection and preprocessing; and employs Rényi Differential Privacy (RDP) to dynamically track and analyze privacy losses during deep learning model training, measuring privacy losses at each stage to achieve fine-grained noise addition. This approach protects both user data privacy and model parameter privacy while achieving a high-accuracy indoor positioning model and low response latency. Attached Figure Description

[0034] Figure 1 This is a diagram of the architecture of the present invention;

[0035] Figure 2 This is a flowchart of the present invention;

[0036] Figure 3 This is a comparison chart of the model accuracy of the present invention and the No-DP federated learning method without differential privacy.

[0037] Figure 4 This is a comparison chart of the model loss of the present invention and the No-DP federated learning method without differential privacy;

[0038] Figure 5 This is a graph showing the changes in the adjusted privacy budget, i.e., sub-privacy ε, according to the present invention.

[0039] Figure 6 This is a graph showing the change of the optimal RDPorder in each round of model training according to the present invention. Detailed Implementation

[0040] The invention will be further described below with reference to the accompanying drawings.

[0041] Reference Figure 1 , Figure 1 The diagram illustrates the architecture of this invention, detailing the specific implementation process in its application scenario. First, the terminal device acquires and stores a local RSS fingerprint dataset from an indoor area. Then, it independently preprocesses and differentially perturbs the collected RSS dataset. Finally, it sends the perturbed RSS data to a nearby edge server. Second, the edge server performs data aggregation, local sub-model training, and service feedback tasks. Finally, it receives the shared sub-model parameters from each edge server, updates the globally shared model parameters using a federated average optimization algorithm, and distributes the updated model parameters to each edge server for the next round of iterative training until the optimal training model is obtained.

[0042] The specific design of the architecture is shown in Table 1. In terms of network architecture design, a simple indoor positioning network model was built using TensorFlow. The entire network structure contains 2 convolutional layers, 2 pooling layers and 3 fully connected layers, where the convolutional layers use a 3*3 kernel size.

[0043] Table 1 Network Architecture of the Invention

[0044]

[0045] Reference Figure 2 A differential privacy federated learning method for wireless positioning based on dynamic privacy budget is implemented according to the following steps:

[0046] Step A, Data Preprocessing and Differential Perturbation: The terminal device will preprocess and differentially perturb the Received Signal Strength (RSS) fingerprint dataset obtained from the indoor area, and then send the perturbed RSS fingerprint data to the nearby edge server. The data preprocessing involves using the Pearson Correlation Coefficient (PCC) to remove weakly correlated RSS fingerprint data, converting it into a grayscale image, and then using (ε, δ)-differential privacy technology to differentially perturb the converted grayscale image to protect the user's data privacy. Finally, the perturbed grayscale image is sent to the edge server.

[0047] The data preprocessing method involves calculating the Pearson correlation coefficient (PCC) between location points and RSS fingerprint data to remove APs with poor availability. The APs are then converted into grayscale images to improve the model's learning and generalization capabilities. Since access points (APs) closer to user terminals have stronger signals, while APs farther away have weaker or even non-existent signals, the preprocessing method employs a different approach.

[0048] Step A can be further broken down into the following steps:

[0049] A1. Divide the indoor area into N uniform small areas, each labeled as l. i , 1≤i≤N, K fingerprint data records R=[r1,r2,...,r K ], where r j = [r1, r2, ..., r m , l i ], 1≤j≤K represents the j-th RSS fingerprint data collected by the user from m APs in region lx.

[0050] A2. Calculate fingerprint data tag L K= [l1, l1, ..., l K ] and w n The PCC value c for 1≤n≤m, where w n This represents the nth column of the fingerprint dataset R. When |c|≤0.3, it indicates that the AP signal strength is unusable. Therefore, at this stage, we removed the fingerprint data column with |c|≤0.3.

[0051] A3. Calculate the Hadamard Product (HP) of the latest RSS fingerprint data record and c to obtain the Hadamard moment HP; where HP = [HP1, HP2, ..., HP...]. K ], for HP i After differential perturbation, it is converted into a grayscale image (img). i Finally, the perturbed grayscale image [img1, img2, ..., img...] is... K Uploaded to the edge server.

[0052] The data preprocessing algorithm in this invention is shown in Table 2. The algorithm first divides the indoor area into N uniform small areas and labels them; second, it removes data with unusable APs signal strength by calculating the PCC value; finally, it perturbs the data and uploads it to the cloud server.

[0053] Table 2 Data Preprocessing Algorithm of the Invention

[0054]

[0055]

[0056] Step B, Local Sub-model Training: The edge server aggregates the received RSS fingerprint data and uses this aggregated data to train the local positioning sub-model. It uses Rényi Differential Privacy (RDP) technology to dynamically calculate the privacy budget loss and allocate the privacy budget, adaptively adjusting the degree of perturbation of the parameter gradient at the current stage to achieve fine-grained control of noisy data. The trained local sub-model parameters are then uploaded to the cloud server.

[0057] Step B further includes: the edge server receiving data from a nearby terminal device u i Disturbed RSS fingerprint image data Then, they are aggregated to form the dataset required for model training. N is the size of the aggregated dataset. express The corresponding labels, i≤j≤N, are then used with the dataset. Perform local sub-model training, and train the edge server E according to formula (1).i Optimize the local sub-model:

[0058]

[0059] Among them, f e (·,·) represents the model's prediction function. Let l represent the optimization objective of the model, and l(·,·) represent the optimization function of the model. Represents edge server E i Model parameters.

[0060] Step B, the local sub-model training process, further includes the following steps:

[0061] B1. Clipping the parameter gradient: Clip the parameter gradient g(img) i After cutting according to formula (2), we get g'(img) i ):

[0062]

[0063] Where C is the gradient clipping threshold, ||g(img) i )||2 is the gradient g(img) i The Euclidean distance of ).

[0064] B2. Add controllable differential private noise: Add controllable differential private noise a on the gradient according to formula (3). Gaussian The new gradient g is obtained from (0, σ2C2). t :

[0065]

[0066] Where B is the number of data points input into the model each time. This is a noise scale.

[0067] B3. Parameter Update: Update the model parameters using the perturbed parameter gradient according to formula (4) to enable the next iteration of training:

[0068]

[0069] in, For the updated model parameters, Here are the model parameters for the current stage, and α is the learning rate.

[0070] The training methods for the local sub-models described above are shown in Table 3.

[0071] Table 3 Local Sub-model Training Method in this Invention

[0072]

[0073]

[0074] Step C, Model Parameter Aggregation: In each iteration, the cloud server receives model parameters shared by each edge server, updates the globally shared model parameters using the federated average optimization algorithm, and then distributes the aggregated model parameters to each edge server for the next iteration of training until the optimal training model is obtained.

[0075] After the cloud server receives the model parameters uploaded by N edge servers, it aggregates the parameters according to formula (5) and then sends them to each edge server for the next iteration of training:

[0076]

[0077] Among them, w global These are the aggregated model parameters. For edge server E i Uploaded model parameters.

[0078] This invention requires satisfying (ε, δ-differential privacy, as proven below:

[0079] Let the total privacy budget of the algorithm be ε. The privacy budget is only needed in the data preprocessing and edge server training stages. Therefore, the privacy budget for these two stages is divided into two sub-privacy budgets ε. c and ε e At this time, ε = ε c +ε e These two stages are in dataset D. c and D e The random algorithm above is And M c and M e The stochastic processes are relatively independent.

[0080] The derivation yields:

[0081]

[0082]

[0083] Therefore, these two stages respectively satisfy (ε) c ,δ)-differential privacy and (ε) e ,δ)-Differential privacy.

[0084] Because the data received by the cloud server has been processed, the aggregation phase of the model parameters by the cloud server also satisfies (ε). e ,δ) Differential privacy.

[0085] The prerequisite for satisfying (εe, δ) differential privacy across the terminal, edge server, and cloud server stages is that a randomized algorithm M must be used. c and M e It satisfies (ε, δ)-differential privacy. The proof is as follows:

[0086] During the data preprocessing stage, N is the number of participants, let P be... i In dataset PD i Random algorithm on Satisfy (ε) i δ i Differential privacy, and any two random algorithms The stochastic processes are independent of each other.

[0087] The derivation yields:

[0088]

[0089] Algorithm Satisfy (ε) c δ-differential privacy, then Random Algorithm M c The output is denoted as O = {r1, r2, ..., r...} N}, because any two algorithms Stochastic processes are independent of each other, therefore,

[0090] Depend on It can be seen that in all For i∈{1, 2, ..., N}, there is exactly one Other neighboring datasets j≠i.

[0091] The derivation yields:

[0092] for Make

[0093]

[0094] If and only if ε c ≥ε i , δ≥δ i Sometimes,

[0095]

[0096] Thus, the random mechanism algorithm M is proved. c Satisfy (ε) c δ) - Differential privacy. Similarly, we can obtain the random mechanism algorithm M e Satisfy (ε) e,δ)-Differential privacy.

[0097] During the model training phase, N represents the number of training rounds for the model, and E represents the training edge server in each round. i In dataset D i Random algorithm on Satisfy (ε) e δ i Differential privacy, and any two random algorithms The stochastic processes are independent of each other. Therefore, for It can be deduced that:

[0098]

[0099] Therefore, in each edge server E j During the model training process, (ε) is satisfied. e ,δ)-Differential privacy.

[0100] For indoor positioning federated learning methods that support dynamic privacy budget allocation under the entire edge computing framework, let the random algorithm be... The operations in the data preprocessing stage and the model training stage are M respectively. c and M e Therefore, M = {M} c M e Let the output of algorithm M be O = {r}. c r e}. Due to algorithm M c and M e The stochastic processes are independent of each other. Therefore, for It can be deduced that:

[0101]

[0102] because We can obtain:

[0103] Pr[M(D)=O]=e ε ×Pr[M(D′)=O]+δ

[0104] Therefore, it can be concluded that the present invention satisfies (ε, δ-differential privacy).

[0105] Safety Analysis of the Invention

[0106] The edge computing-based fingerprint indoor positioning federated learning dynamic privacy budget allocation method ADP-FLocEC described in this invention is based on federated learning and (ε,δ)-differential privacy technology. While ensuring service quality, it simultaneously protects user data privacy and model parameter privacy during model training. During data preprocessing, PCC and (ε,δ)-differential privacy technology are used to differentially perturb the converted grayscale image to protect user data privacy. During model training, this invention uses RDP to dynamically analyze and measure the privacy loss at each stage, achieving fine-grained noise addition. In the cloud server parameter aggregation stage, a federated averaging optimization algorithm is used to update the globally shared model parameters. In these three stages, this invention... The model satisfies (ε, δ)-differential privacy, making it difficult for attackers to obtain user data and model parameter privacy through malicious means. Given the high complexity and dynamism of edge computing environments, existing privacy budget allocation strategies are unsuitable for protecting indoor fingerprint positioning privacy in edge computing environments. This invention studies the dynamic allocation and adjustment of privacy budgets in edge computing environments and proposes an indoor positioning federated learning method that supports dynamic privacy budget allocation. This method achieves refined protection of user data privacy and model parameter privacy, effectively enabling refined allocation of privacy budgets during indoor positioning model training. While effectively resisting attacks such as differential analysis and Bayesian inference, it also improves the performance of indoor fingerprint positioning models.

[0107] Table 4 shows a comparison of the time overhead of this invention and the No-DP federated learning method (which does not employ differential privacy) at different stages. It can be seen that the time overhead of No-DP during model training is significantly less than that of this method. However, the time overhead of both methods during model testing is almost identical, both reaching 0.35 seconds. Although this method increases the time overhead during model training, it effectively protects user data privacy and parameter privacy, and has almost no impact on the time overhead during model application.

[0108] Table 4

[0109]

[0110] like Figure 3 , Figure 4As shown, we statistically analyzed the training accuracy, testing accuracy, training loss, and testing loss of the federated learning method No-DP (without differential privacy) and the method of our invention. It can be seen that the training accuracy and testing accuracy of both methods gradually increase with the number of iterations, while the training loss and testing loss gradually decrease. When the number of iterations reaches 200, the training accuracy, testing accuracy, training loss, and testing loss of the No-DP method reach 99.95%, 99.45%, 0.0021, and 0.0249, respectively, while those of our invention are 95.38%, 94.84%, 0.3085, and 0.4488, respectively. Furthermore, although our invention adds a significant amount of noise at the beginning of model training, resulting in a slower convergence speed compared to the No-DP method, it effectively resists differential attacks, inference attacks, and model inversion attacks, achieving privacy protection for both users and model parameters, even at the cost of 4.5% lower model accuracy.

[0111] like Figure 5 , Figure 6 As shown, we calculated the optimal RDP Order and adjusted privacy budget for each round during model training. It can be seen that the privacy budget gradually increases while the RDP Order gradually decreases with the number of iterations. When the number of iterations is 1 and 200, the privacy budget reaches 0.35367 and 3.1357, respectively. When the number of iterations is 1, the model's fitting ability is reduced due to the addition of a large amount of noise to the parameter gradient. As the number of iterations increases, the model's fitting ability gradually becomes stronger. At this point, adding a large amount of noise will severely affect the model's performance. Therefore, in order to achieve an effective trade-off between model performance and privacy protection strength, this invention dynamically adjusts the privacy budget parameter with the increase of the number of iterations, and finely controls the amount of noise added in each round to obtain efficient indoor positioning performance.

Claims

1. A wireless positioning differential privacy federated learning method based on dynamic privacy budget, characterized in that, Specifically, the following steps are implemented: Step A, data preprocessing and differential perturbation: the terminal device performs data preprocessing and differential perturbation on the received signal strength, RSS fingerprint data set based on the received signal strength obtained from the indoor area, and then sends the perturbed RSS fingerprint data to the adjacent edge server; the data preprocessing is to remove weakly correlated RSS fingerprint data by using the Pearson correlation coefficient, PCC, and then to use the - differential privacy technology is used to perform differential perturbation on the converted gray image, and finally the perturbed gray image is sent to the edge server; Step B, local sub-model training: the edge server aggregates the accepted RSS fingerprint data, uses the aggregated data to train the local positioning sub-model, and uses Differential Privacy Differential Privacy, RDP technology dynamically calculates the loss of privacy budget and allocates privacy budget, adaptively adjusts the degree of disturbance to the parameter gradient at the current stage, realizes fine-grained control of noisy data, and uploads the trained local sub-model parameters to the cloud server; Specifically: The edge server receives the RSS fingerprint data from the nearby terminal device Disturbed RSS fingerprint picture data After that, it is aggregated to form a data set required for model training , , , N is the size of the data set after aggregation, denotes the corresponding label, , then the local sub-model training is performed using the data set , and the local sub-model of the edge server is optimized according to formula (1): (1) wherein, represents a prediction function of the model, represents an optimization objective of the model, represents an optimization function of the model, represents a model parameter of the edge server . The local sub-model training process further includes the following steps: B1. clipping the parameter gradient: clipping the parameter gradient After clipping according to formula (2), we get : (2) wherein C is a gradient clipping threshold value, is the Euclidean distance of the gradient ​ B2. Add controllable differential private noise: Add controllable differential private noise on the gradient according to equation (3) Get new gradient : (3) where B is the number of data input into the model each time, is the noise scale; B3, parameter updating: the disturbed parameter gradient is used to update the model parameters according to formula (4) to make the next iteration training: (4) wherein, are the updated model parameters, are the model parameters of the current stage, is the learning rate; Step C, model parameter aggregation: in each iteration process, the cloud server receives the model parameters shared from each edge server, updates the global shared model parameters using the federated average optimization algorithm, and then distributes the aggregated model parameters to each edge server for the next iteration training until the optimal training model is obtained.

2. The wireless positioning differential privacy federated learning method based on dynamic privacy budget according to claim 1, characterized in that, Step A further includes: A1, divide the indoor area into N uniform small areas, and the label of each area is , , K pieces of fingerprint data records collected by the user in the indoor area wherein: , represents the jth RSS fingerprint data collected by m APs in the area ​ A2, compute the fingerprint data label with , the PCC value c of the APs, where c is the nth column data of the fingerprint dataset R; when the APs signal strength is not available, thus, in this phase, we eliminate the fingerprint data column of the APs; A3. Calculate the Hadamard Product (HP) of the latest RSS fingerprint data record and c to obtain the Hadamard moment. ;in ,right Converted to grayscale image after differential perturbation Finally, the perturbed grayscale image will be obtained. Uploaded to the edge server.

3. The method of claim 1, wherein, After the cloud server receives the model parameters uploaded by the N edge servers in step C, the parameters are aggregated according to formula (5) and distributed to each edge server for the next iteration training: (5) wherein, are the model parameters after the aggregation, are the model parameters after the aggregation, are the model parameters after the aggregation,

Images