Digital traffic sign horizontal federal recommended methods, electronic devices and storage media
By employing a horizontal federated learning system and differential privacy methods, the problems of user privacy protection and data silos were solved, enabling accurate and personalized digital traffic sign recommendations, thereby improving driving safety and user experience.
Patent Information
- Application Number
- CN202311181510.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-13
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-09-13
AI Technical Summary
Current technologies lack the ability to recommend accurate and personalized digital traffic signs to drivers while protecting user privacy, resulting in insufficient driving safety and user perception.
A horizontal federated learning system is adopted, and the Wide&Deep model is deployed through a client-server architecture. Differential privacy methods are used to encrypt model parameters and perform weighted averaging to achieve the fusion of multi-source data and personalized recommendations.
While protecting user privacy, the accuracy of the recommendation model has been improved, enabling precise and personalized digital traffic sign recommendations, thereby enhancing driving safety and user experience.
Smart Images

Figure CN117216792B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent transportation technology and road traffic safety technology, and in particular to a method for horizontal federal recommendation of digital traffic signs, electronic equipment, and storage medium. Background Technology
[0002] Digital traffic signs are a 1:1 replica of physical road traffic signs. On the 1:1 replica map, the geographical location of each physical road traffic sign has a corresponding digital traffic sign, thus enabling the display of all physical road traffic signs within the digital traffic sign navigation system. In the field of intelligent transportation technology, recommendation methods for digital traffic signs are lacking; only limited methods exist for sign location recommendation, attraction recommendation, travel service management, and location services. This can be achieved through the following methods:
[0003] (1) CN201810593593.5 A method for recommending the location of digital signage coupled with multiple source elements
[0004] By processing elements, dividing regions, recommending locations, and validating models, this invention automatically recommends locations for digital signage, achieving accurate site selection. This method comprehensively considers multiple factors, providing strong reference for recommended locations, high timeliness, high accuracy, and excellent impact on digital signage. It can meet the needs of numerous advertisers and media companies, realizing an efficient and scientific method for recommending digital signage locations that couples multiple factors. This method can be extended to various application areas with digital signage location selection requirements.
[0005] (2) CN201510938942.9 A Personalized Tourist Attraction Recommendation Method Based on Knowledge and Location
[0006] This invention addresses the shortcomings of relying solely on knowledge-based recommendations, which wastes communication bandwidth. The implementation steps are as follows: 1. Establish a complete recommendation system consisting of users, assisted positioning facilities, and a recommendation computing system; 2. Initialize this recommendation computing system; 3. Users send recommendation requests to the system from their current location; 4. The location processing module of the recommendation computing system responds to the user's request; 5. The recommendation computing system executes a query dominance algorithm on the candidate attractions; 6. The recommendation computing system finally outputs a set of recommended tourist attractions. By utilizing a dominance query algorithm in the database, the recommendation value of candidate attractions is comprehensively compared, reducing the communication volume between the user and the recommendation system. Furthermore, by incorporating the user's historical preferences, the effectiveness and reliability of personalized recommendation results are ensured, making it applicable to the field of tourist attraction or route recommendation.
[0007] (3) CN114742308A A method and system for managing personalized federal mobility services with privacy protection
[0008] This paper presents a privacy-preserving, federated personalized travel service management method and system. The system includes a basic data layer, a business response layer, a real-time data layer, an auxiliary training layer, and a user interaction layer. It can complete travelers' travel tasks and provide a visual interface for user operations and data results. Based on federated learning, the system can train travel models for different users, providing personalized travel solutions while protecting user privacy.
[0009] (4) CN110825955A Location-based Distributed Differential Privacy Recommendation Method
[0010] Traditional recommendation systems are ill-suited for location-based recommendation services and also suffer from privacy concerns. This paper addresses these issues by designing a distributed privacy-preserving recommendation framework and differential privacy theory. A singular value decomposition-based recommendation algorithm is developed using this framework. By employing an order-preserving encryption function to protect the user's requested location, both privacy and recommendation performance are achieved.
[0011] In summary, how to recommend accurate and personalized digital traffic signs to drivers while protecting user privacy, thereby improving driving safety and user experience, has become an urgent problem to be solved by those skilled in the art. Summary of the Invention
[0012] The purpose of this invention is to provide a method, electronic device, and storage medium for lateral federation recommendation of digital traffic signs, which can recommend accurate and personalized digital traffic signs to drivers while protecting user privacy, thereby improving driving safety and user perception.
[0013] To achieve the above objectives, the present invention provides the following solution:
[0014] A method for lateral federal recommendation of digital traffic signs, the method comprising:
[0015] Step S1: Obtain the client-server architecture of the horizontal federated learning system; the client-server architecture of the horizontal federated learning system includes a server and multiple clients;
[0016] Step S2: Deploy the Wide&Deep model on each of the aforementioned clients;
[0017] Step S3: Train the Wide&Deep model on the client using the local training dataset corresponding to the client to obtain the parameters of the Wide&Deep model after the h-th training on each client; the initial value of h is 1; the input of the Wide&Deep model after the h-th training on the client is the local training dataset, and the output is the digital traffic sign feature data sorted in descending order of recommendation priority; the local training dataset includes digital traffic sign feature data, driver social feature data, driving behavior feature data, and vehicle feature data;
[0018] Step S4: The parameters of the Wide&Deep model after the h-th training on each client are encrypted using the differential privacy method, and the encrypted parameters of the Wide&Deep model after the h-th training on each client are sent to the server using gRPC and ProtoBuf.
[0019] Step S5: Use a weighted average method to aggregate the parameters of the encrypted Wide&Deep model trained on each client for the hth time, which are received by the server, to obtain the global model parameters after the hth aggregation on the server.
[0020] Step S6: Encrypt the global model parameters after the h-th aggregation on the server using differential privacy, and send the encrypted global model parameters after the h-th aggregation on the server to each client using gRPC and ProtoBuf.
[0021] Step S7: Determine whether h is less than the preset number H; if yes, let h = h + 1, and update the Wide&Deep model on each client using the encrypted global model parameters after the hth aggregation on the server, and return to step S3; if no, update the Wide&Deep model on each client using the encrypted global model parameters after the hth aggregation on the server, and obtain the final recommendation model on each client.
[0022] Step S8: Input the real-time acquired road traffic sign feature data, driver social feature data, driving behavior feature data, and vehicle feature data into the final recommendation model to obtain digital traffic sign feature data arranged in descending order of recommendation priority;
[0023] Step S9: Recommend digital traffic signs based on the digital traffic sign feature data sorted in descending order of recommendation priority.
[0024] Optionally, the Wide&Deep model is deployed on each of the aforementioned clients, specifically including:
[0025] The server broadcasts the initialized Wide&Deep model, model parameters, and hyperparameters to each client in plaintext; the hyperparameters include the learning rate and the number of iterations.
[0026] Optionally, the Wide&Deep model on the client is trained using the local training dataset corresponding to the client to obtain the parameters of the Wide&Deep model after the h-th training on each client, which further includes:
[0027] Obtain the original local training dataset corresponding to each client;
[0028] Convert the original local training dataset into JSON data;
[0029] The JSON data is processed sequentially for missing values, outliers, and data standardization to obtain the local training dataset corresponding to each client.
[0030] Optionally, the digital traffic sign feature data includes sign code, sign name, sign category, sign content, sign shape, sign border color, sign background color, sign lining color, sign icon color, sign size, sign dynamic icon, sign static icon, sign latitude and longitude, location vector, geographical location of the road segment, latitude and longitude and location vector of the visible point, sign installation tilt angle, sign placement and left and right side position of the road segment, sign placement method, sign placement height, road segment curvature, and remarks.
[0031] Optionally, the driver's social characteristics data includes name, age, ID number, mobile phone number, emergency contact, driver's license validity period, driving experience, occupation, height, weight, and illness.
[0032] Optionally, the driving behavior characteristic data includes driving history, driving preferences and habits, driving behavior patterns, historical accident records, and historical traffic violation records.
[0033] Optionally, the vehicle characteristic data includes vehicle type, vehicle license plate, usage and driving status, vehicle type, vehicle condition, passenger capacity, empty vehicle weight, total vehicle weight, permissible gross weight, maximum load capacity, maximum axle load, vehicle width, vehicle height, wheelbase, track width, front overhang, rear overhang, minimum ground clearance, maximum speed, fuel tank capacity, vehicle inspection certificate, and vehicle city access code.
[0034] Optionally, the client is a terminal device; the terminal device includes augmented reality head-up display products, advanced driver assistance systems, mobile phones, in-vehicle audio-visual systems, traffic sign information displays for smart roads, and multi-functional display boards.
[0035] The present invention also provides the following solutions:
[0036] An electronic device includes a memory and a processor, the memory storing a computer program and the processor running the computer program to cause the electronic device to perform the digital traffic sign lateral federal recommendation method.
[0037] The present invention also provides the following solutions:
[0038] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the digital traffic sign lateral federation recommendation method.
[0039] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0040] This invention discloses a horizontal federated recommendation method, electronic device, and storage medium for digital traffic signs. It utilizes a horizontal federated learning method to improve the accuracy of the recommendation model while protecting user privacy. The method trains the horizontal federated learning system based on a client-server architecture, solving the data silo problem caused by data isolation between clients during digital traffic sign recommendation model training. This allows for the fusion of digital traffic sign feature data, driver social feature data, driving behavior feature data, and vehicle feature data from various clients for model training, achieving multi-source data-driven digital traffic sign recommendation. Because the model is trained using a large amount of data from various clients, the trained model has high accuracy and can provide personalized recommendations for different users, thus recommending accurate and personalized digital traffic signs to drivers, improving driving safety and user perception. Differential privacy is used to handle model parameters transmitted between the server and client, ensuring data security. gRPC is used for data transmission and reception between the client and server, and ProtoBuf is used as the service definition for gRPC to achieve data serialization, improving data compression rate and transmission efficiency. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 This is a flowchart of an embodiment of the digital traffic sign lateral federation recommendation method of the present invention;
[0043] Figure 2This is a diagram of the architecture of a horizontal federated learning system.
[0044] Figure 3 This is a functional diagram of each module of the present invention;
[0045] Figure 4 This is a schematic diagram illustrating the functions that the horizontal federation communication module of the present invention can achieve;
[0046] Figure 5 This is a schematic diagram of the communication structure between the server and the client of the present invention;
[0047] Figure 6 This is a schematic diagram illustrating the functions that the data preprocessing module of this invention can achieve;
[0048] Figure 7 This is a schematic diagram illustrating the functions that the horizontal federated encryption module of this invention can achieve;
[0049] Figure 8 This is a schematic diagram illustrating the functions that the horizontal federated digital traffic sign recommendation model training module of the present invention can achieve.
[0050] Figure 9 This is a diagram illustrating parameter updates during model training.
[0051] Figure 10 This is a schematic diagram of horizontal federated learning. Detailed Implementation
[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0053] The purpose of this invention is to provide a method, electronic device, and storage medium for lateral federation recommendation of digital traffic signs, which can recommend accurate and personalized digital traffic signs to drivers while protecting user privacy, thereby improving driving safety and user perception.
[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0055] Example 1
[0056] Figure 1 This is a flowchart of an embodiment of the digital traffic sign lateral federation recommendation method of the present invention. Figure 1 As shown, this embodiment provides a method for lateral federation recommendation of digital traffic signs, including the following steps:
[0057] Step S1: Obtain the client-server architecture of the horizontal federated learning system; the client-server architecture of the horizontal federated learning system includes a server and multiple clients.
[0058] Among them, the client is the terminal device; the terminal device includes augmented reality head-up display products, advanced driver assistance systems, mobile phones, in-vehicle audio-visual systems, traffic sign information display screens for smart roads, and multi-functional display boards.
[0059] Step S2: Deploy the Wide&Deep model on each client.
[0060] Step S2 specifically includes:
[0061] The server broadcasts the initialized Wide&Deep model, along with its parameters and hyperparameters, to each client in plaintext; the hyperparameters include the learning rate and the number of iterations.
[0062] Step S3: Train the Wide&Deep model on the client using the local training dataset corresponding to the client, and obtain the parameters of the Wide&Deep model after the h-th training on each client; the initial value of h is 1; the input of the Wide&Deep model after the h-th training on the client is the local training dataset, and the output is the digital traffic sign feature data sorted in descending order of recommendation priority; the local training dataset includes digital traffic sign feature data, driver social feature data, driving behavior feature data, and vehicle feature data.
[0063] The following steps precede step S3:
[0064] Obtain the original local training dataset for each client.
[0065] Convert the original local training dataset into JSON data.
[0066] The JSON data is processed sequentially for missing values, outliers, and data standardization to obtain the local training dataset for each client.
[0067] Specifically, digital traffic sign feature data includes sign code, sign name, sign category, sign content, sign shape, sign border color, sign background color, sign lining color, sign icon color, sign size, sign dynamic icon, sign static icon, sign latitude and longitude, location vector, geographical location of the road segment, latitude and longitude and location vector of the visible point, sign installation tilt angle, sign placement and left and right side position of the road segment, sign placement method, sign placement height, road segment curvature of the sign, and remarks.
[0068] Driver social characteristics data include name, age, ID number, mobile phone number, emergency contact, driver's license validity period, driving experience, occupation, height, weight, and medical conditions.
[0069] Driving behavior data includes driving history, driving preferences and habits, driving behavior patterns, historical accident records, and historical traffic violations.
[0070] Vehicle characteristic data includes vehicle type, vehicle license plate, usage and driving status, vehicle type, vehicle condition, passenger capacity, empty vehicle weight, total equipment weight, permissible gross weight, maximum load capacity, maximum axle load, vehicle width, vehicle height, wheelbase, track width, front overhang, rear overhang, minimum ground clearance, maximum speed, fuel tank capacity, vehicle inspection certificate, and vehicle city access code.
[0071] Step S4: Use differential privacy to encrypt the parameters of the Wide&Deep model after the h-th training on each client, and send the encrypted parameters of the Wide&Deep model after the h-th training on each client to the server using gRPC and ProtoBuf.
[0072] Step S5: Use a weighted average method to aggregate the parameters of the Wide&Deep model after the h-th training on each client, which are received by the server after encryption, to obtain the global model parameters after the h-th aggregation on the server.
[0073] Step S6: Use differential privacy to encrypt the global model parameters after the h-th aggregation on the server, and send the encrypted global model parameters after the h-th aggregation on the server to each client using gRPC and ProtoBuf.
[0074] Step S7: Determine if h is less than the preset number H; if yes, let h = h + 1, and update the Wide&Deep model on each client using the global model parameters after the hth aggregation on the encrypted server, and return to step S3; if no, update the Wide&Deep model on each client using the global model parameters after the hth aggregation on the encrypted server, and obtain the final recommendation model on each client.
[0075] Step S8: Input the real-time acquired road traffic sign feature data, driver social feature data, driving behavior feature data, and vehicle feature data into the final recommendation model to obtain digital traffic sign feature data arranged in descending order of recommendation priority.
[0076] Step S9: Recommend digital traffic signs based on the digital traffic sign feature data sorted in descending order of recommendation priority.
[0077] The technical solution of the present invention is illustrated below with a specific embodiment:
[0078] This invention presents a location-based lateral federated recommendation method for digital traffic signs. Based on vehicle location, driving information, and road traffic rules, and following national standards for road traffic signs and markings, it combines road traffic sign features (basic features, location features, and other features) set by traffic management departments, driver social and behavioral characteristics, and vehicle features to train a digital traffic sign recommendation model (i.e., the Wide&Deep model) to obtain the final recommendation model. This final model enables accurate and personalized recommendations of digital traffic signs, improving driving safety and user perception. The recommendation results are distributed from a cloud server to various client terminal devices, including but not limited to in-vehicle T-BOX, AR-HUD, ADAS systems, mobile phones, in-vehicle audio-visual systems, smart road traffic sign information displays, and multi-functional display boards.
[0079] The digital traffic sign horizontal federated recommendation method of this invention is based on a client-server architecture of a horizontal federated learning system. To more intuitively introduce the digital traffic sign horizontal federated recommendation method of this invention based on a client-server architecture of a horizontal federated learning system, the main technical means involved in the digital traffic sign horizontal federated recommendation method of this invention are described below in the form of modules:
[0080] This invention mainly comprises the following four parts:
[0081] (1) Horizontal Federation Communication Module
[0082] (2) Data preprocessing module
[0083] (3) Horizontal federated encryption module
[0084] (4) Training module for horizontal federated digital traffic sign recommendation model
[0085] The horizontal federated communication module, data preprocessing module, horizontal federated encryption module, and horizontal federated digital traffic sign recommendation model training module play different roles in the system (horizontal federated learning system). Specifically, the horizontal federated encryption module is mainly used to encrypt the parameters of the Wide&Deep model after the h-th training iteration on each client using differential privacy in step S4, and to encrypt the parameters of the Wide&Deep model after the h-th training iteration on the server using differential privacy in step S6. The horizontal federated communication module is mainly used to send the encrypted parameters of the Wide&Deep model after the h-th training iteration on each client to the server using gRPC and ProtoBuf in step S4, and to send the encrypted parameters of the Wide&Deep model after the h-th training iteration on the server to each client using gRPC and ProtoBuf in step S6. The data preprocessing module is mainly used to convert the original local training datasets corresponding to each client into JSON data, and to perform missing value processing, outlier processing, and data standardization on the JSON data to obtain the local training datasets corresponding to each client. The training module for the horizontal federated digital traffic sign recommendation model is mainly used to implement steps S1, S2, S3, S5 and S7.
[0086] The architecture diagram of the horizontal federated learning system (location-based digital traffic sign horizontal federated recommendation system) is as follows: Figure 2 As shown. See also Figure 2 The system functional architecture of this invention is planned as follows:
[0087] The system architecture involved in this invention consists of four parts: a horizontal federated communication module, a data preprocessing module, a horizontal federated encryption module, and a horizontal federated digital traffic sign recommendation model training module. The system adopts a client-server architecture. Figure 2 In the architecture shown, number 1 indicates that the server sends encrypted model parameters to the client, number 2 indicates that the server securely aggregates the client's gradients, number 3 indicates that the client sends back the gradients updated to the local model, and number 4 indicates that the client trains and updates the local model.
[0088] Figure 2 The system architecture diagram shown clearly defines the relationships between the horizontal federated communication module, the data preprocessing module, the horizontal federated encryption module, and the horizontal federated digital traffic sign recommendation model training module:
[0089] (1) Client servers communicate with each other through a horizontal federation communication module.
[0090] (2) Raw data such as vehicle feature data, driving feature data, and driver social feature data are processed by the data preprocessing module to provide data for recommendation model training.
[0091] (3) The client and server use a horizontal federated encryption module to encrypt the model gradient, protecting privacy data from being leaked during transmission between the client and server.
[0092] (4) Horizontal Federal Digital Traffic Sign Recommendation Model Training Module: The client uses this module to train the local model, and the server uses this module to aggregate and update the global model.
[0093] Figure 3 This is a functional diagram of each module of the present invention. See also: Figure 3 The modules are described in detail below:
[0094] (1) Horizontal Federation Communication Module
[0095] The horizontal federated communication module enables communication between the server and client. It employs gRPC and ProtoBuf to achieve end-to-end communication between the server and client, resulting in fast transmission speeds and reduced latency. gRPC enables remote function calls, which are performed as if calling local functions. ProtoBuf is a data serialization protocol; using ProtoBuf as the service definition for gRPC allows for data serialization, improving data compression rates and transmission efficiency.
[0096] gRPC is used to implement data transmission and reception between the client and the server. During node communication, the parameters sent by the client are converted into byte streams one by one and sent to the server. After receiving the data, the server converts the format of the byte streams one by one so that the server can recognize them locally.
[0097] The functions that the horizontal federation communication module can achieve are as follows: Figure 4 As shown:
[0098] ① The client calls a remote function on the server by initiating a function call request; the remote function is a method in a remote service, which is a service-defined method and is not specific to any particular method.
[0099] ② For the request information (including the function identifier and parameter values of the server response), use ProtoBuf to perform object serialization compression (each parameter is converted into a byte stream and sent to the server, which can recognize the byte stream).
[0100] ③ The client sends a request to the server.
[0101] ④ After receiving the request information from the client, the server decodes the request body, deserializes it, and processes the business logic (by calling the corresponding function on the server based on the function identifier and the passed parameter values, and obtaining the function call result), and returns the call result to the client; processing the business logic means processing the request according to the server-side method (which can be understood as a server-side custom function or method).
[0102] ⑤ Use ProtoBuf to perform object serialization and compression on the call result (response result, i.e., the result after the logical processing in the method after calling the server method) to obtain the response result.
[0103] ⑥ After receiving the response, the client decodes and deserializes it to obtain the result of the remote function call from the server.
[0104] The communication structure between the server and the client is as follows: Figure 5 As shown.
[0105] (2) Data preprocessing module
[0106] Data preprocessing is a separate module used to process raw feature data (the raw local training dataset) to provide training data for the horizontally federated digital traffic sign recommendation model training module.
[0107] The dataset (multi-source dataset, i.e. local training dataset) contains road traffic sign feature data, driver social feature data, driving behavior feature data, and vehicle feature data.
[0108] Road traffic sign feature data includes basic features, location features, and other features. Basic features include sign code, sign name, sign category, sign content, sign shape, sign border color, sign background color, sign lining color, sign icon color, sign size, dynamic icon, and static icon. Location features include sign latitude and longitude, location vector, geographical location of the road segment, and latitude and longitude and location vector of the visible point. Other features include sign installation tilt angle, sign placement relative to the left and right sides of the road segment, sign placement method, sign placement height, road segment curvature, and remarks.
[0109] Driver social characteristics data include name, age, ID number, mobile phone number, emergency contact, driver's license validity period, driving experience, occupation, height, weight, and medical conditions.
[0110] Driving behavior data includes driving history, driving preferences and habits, driving behavior patterns, historical accident records, and historical traffic violations.
[0111] Vehicle characteristic data includes vehicle type, vehicle license plate, usage and driving status, vehicle type, vehicle condition, passenger capacity, empty vehicle weight, total equipment weight, permissible gross weight, maximum load capacity, maximum axle load, vehicle width, vehicle height, wheelbase, track width, front overhang, rear overhang, minimum ground clearance, maximum speed, fuel tank capacity, vehicle inspection certificate, and vehicle city access code.
[0112] The functions that the data preprocessing module can perform are as follows: Figure 6 As shown:
[0113] ① Define a set of standard operators to unify data description and data transformation, and format the data as follows:
[0114] <Primitive data structure, primitive data type, target data structure, target data type>.
[0115] ②Use JSON as an intermediate medium for data conversion to convert the original data into JSON data, thus achieving heterogeneous data conversion.
[0116] ③ Parse JSON data to generate table structure.
[0117] ④ Design the data access according to the standard data types commonly used in JSON data.
[0118] ⑤ After handling missing values, outliers, and data standardization, the data preprocessing is complete, and the target data is obtained.
[0119] The commonly used standard data types for JSON data are shown in Table 1.
[0120] Table 1 Commonly Used Standard Data Types for JSON Data
[0121] Data types Example String {"name":"Marry"} number {"age":25} object {"sign":{"id":T001,"name":"Caution: Pedestrians","type":"Warning"}} array {"sign":[T001,"Caution: Pedestrians","Warning"]} Boolean {"overspeed":false} NULL {"remark":null}
[0122] The data preprocessing operations are shown in Table 2.
[0123] Table 2 Data Preprocessing Operations
[0124]
[0125]
[0126] (3) Horizontal federated encryption module
[0127] In the horizontal federated recommendation system for digital traffic signs, each client, as a participant, transmits parameters (gradients) to the server and uses gradient descent to update the parameters and gradients of the global digital traffic sign recommendation model. However, the parameters transmitted in plaintext during this process pose a security risk, as attackers could steal and misuse them to infer users' sensitive private data. Therefore, a horizontal federated encryption module is added to the digital traffic sign horizontal federated recommendation system to enhance the protection of user data security.
[0128] The horizontal federated encryption module aims to encrypt model parameters transmitted between the client and server, preventing third parties from stealing user privacy information. The functions that the horizontal federated encryption module can achieve include... Figure 7 As shown:
[0129] ① Model synchronization. The server broadcasts the initialized model to all participants.
[0130] ② Client-side model training (local model training on the client) and calculation of model parameters. Each of the N participants (clients) in the horizontal federated learning process has its own local private dataset. The model is trained and its parameters are calculated locally on each participant's machine.
[0131] ③ Client-side model parameter pruning. Set the pruning threshold as the sensitivity S, and use the median norm of the initial model parameters in the recommendation model as the threshold S. In the t-th iteration of training, if the updated model parameters are less than S, retain the new parameters; if the updated model parameters are greater than S, prune the model parameters according to the threshold.
[0132] Model parameters before cropping: w t
[0133] Model parameters after cropping:
[0134]
[0135] Where ||·||2 represents the I2 norm value, and S represents the number of model parameters selected by each participant in this iteration of training.
[0136] ④ Adding noise to client-side model parameters. After trimming, Gaussian noise is added to the local model parameter values to disrupt the locally generated model parameters before sending them to the server for aggregation.
[0137] Model parameters before adding noise:
[0138] Model parameters after adding noise:
[0139]
[0140] The model parameters before adding noise are updated using the model parameters after adding noise. The noise follows a Gaussian distribution (normal distribution), where 0 represents its mean and σ represents its mean. 2 S 2 N(0,σ) represents the variance, σ represents the dispersion of the data, and N(0,σ) represents the variance. 2 ,S 2 The probability density function of the noise function follows a Gaussian distribution with a mean of 0 and a variance of σ. 2 S 2 .
[0141] ⑤ Server model parameter aggregation. The server receives the model parameters, aggregates them, and broadcasts the aggregated model parameters to all participants.
[0142] Noisy model parameters After being transmitted to the server, the updated model parameters are obtained by weighted average aggregation on the server.
[0143]
[0144] Where K represents the number of participants (clients), that is, there are K participants in the training of the horizontal federated digital exchange recommendation model.
[0145] (4) Training module for horizontal federated digital traffic sign recommendation model
[0146] The horizontal federated digital traffic sign recommendation model training module receives model parameters encrypted by the horizontal federated encryption module. The client or server receives these encrypted model parameters to complete the recommendation model training task. In each model construction, each driver's user data is randomly divided into a 60% training set, a 20% validation set, and a 20% test set.
[0147] The horizontally federated digital traffic sign recommendation model training module enables the training of multiple recommendation models, including Wide&Deep and DeepFM. Taking the Wide&Deep model training as an example, the Wide part is mainly used to learn sparse discrete features (such as driver social characteristics data, name, age, driving experience, occupation, height, weight, diseases, etc.), memorizing and learning linear correlations among discrete features; the Deep part is a multilayer perceptron used to capture higher-order correlations between features and learn previously unseen combined features. Through joint training of Wide&Deep, both low-order and high-order combined features are learned simultaneously.
[0148] The functions that the horizontal federated digital traffic sign recommendation model training module can achieve are as follows: Figure 8 As shown:
[0149] ① The server issues a digital traffic sign recommendation model training task. Upon determining the model training task, the server broadcasts the initialized model and model parameters w0 in plaintext to all participating parties (clients), while simultaneously determining the same learning rate η, number of iterations, and other hyperparameters.
[0150] ② The client retrieves the model and model parameters from the server to train the model and update the local model. K participants execute the update in parallel. in These are the latest model parameters obtained from the server; k∈[1,K],k∈N, where k is the identifier for each participant: participant 1, participant 2, ..., participant K. Each client acts as a participant, using gradient descent as the unified optimization algorithm to train the recommendation model using local data samples.
[0151] Locally store data sample D k Divide the data into batches of size M (e.g., D1 represents the local data sample of participant 1), and perform S iterations with a batch size of B. Obtain the local model parameters based on the previous iteration. Update model parameters Where k represents the client, η represents the learning rate, b is the batch number, and i-1 indicates the number of iterations in the training process, i.e., the (i-1)th iteration in the training process. If the current iteration in the training process is the i-th iteration, then the previous iteration was the (i-1)th iteration.
[0152] Loss gradient calculation:
[0153] The detailed explanations of the above parameters are as follows:
[0154] (1) Local data sample D k =M×B, which means that each participant will conduct S iterations of model training with a batch size of M and a batch number of B (the number of iterations ranges from 1 to S, and each iteration is i).
[0155] (2) in This represents the latest model parameters obtained from the server, and the initial model parameters used in local model training on participant k. These are the latest model parameters obtained from the server. (k) identifies the participants, where k represents the k-th training iteration. This represents the local model parameters trained in the k-th iteration with batch number 1 for participant k, i.e., the local model parameters trained in the first batch of models for participant k; each participant will perform S iterations of model training with batch size M and batch quantity B, i.e., local data samples D k =M×B. This represents the local model parameters for participant k in the (i-1)th iteration of training with batch number B (each participant trains in batches B). Specifically, it represents the local model parameters for participant k in the last iteration of training with batch number B. This can be understood as the initial values of the local model parameters for the current batch being the results of the previous batch. i-1 indicates the number of iterations trained by the participant; that is, the (i-1)th iteration. If the current iteration is the i-th, then the previous iteration was the (i-1)th.
[0156] (3) This indicates that the parameters for the next batch of local models are calculated using the learning rate, loss gradient, and parameters from the previous batch. Here, b represents the batch number, ranging from 0 to B-1; and i represents the iteration number.
[0157] (4) k represents the client (participant), and b represents the batch sequence number. This represents the loss gradient for participant k corresponding to batch number b. k Let F be the loss gradient calculation function for participant k. k (w t ) indicates that the loss gradient of participant k is calculated based on the local model parameters.
[0158] Taking the Wide&Deep model as an example for the recommendation model training task:
[0159] The expression for the predicted value of the Wide&Deep model is:
[0160]
[0161] Where Y is the result label, Y∈{0,1}, and σ(·) is the sigmoid activation function. This represents the wide part of the model's weight parameters, x = [x1, x2, x3, ... x]. n ] represents n eigenvectors, It is a cross feature, a lf This is the activation function of the Deep part, where b represents the bias parameter, specifically bias(bi). P represents the model's predicted probability value; P(Y=1|x) represents the probability that Y=1 occurs given that x has occurred. The Wide&Deep model predicts a probability value of Y=1 given that x has occurred. The Wide&Deep model is a weighted combination of the Wide and Deep models, with the final output calculated using the sigmoid activation function, representing the probability of Y=1. This represents the weight parameters of the Wide model part. This represents the weight parameters of the Deep model.
[0162] The client-side model forward propagates by using local training data to perform calculations through linear transformations and nonlinear activation functions in each layer of the neural network, and updates the local model using a stochastic gradient-based model update method.
[0163] The client uses backpropagation to update the weight and bias parameters of each layer in Wide&Deep. After the local model completes one round of training, the gradient is calculated based on the chain rule to update the local model parameters.
[0164] ③ The client will update the model parameters The data is sent to the server, where the model parameters are encrypted using the horizontal federated encryption module.
[0165] ④ The server receives the model parameters from each client, aggregates them, and calculates the new global model parameters using a weighted average method. Here, t represents the round number of each global model update.
[0166] Model parameter update calculation:
[0167] Model parameter weighted average calculation:
[0168] In the formula, This represents the global model parameters after the model parameters from the k participants are aggregated to the server in round t+1, i.e., the latest global model parameters on the server. n represents the number of weight parameters in the wide part of the model. k This represents the number of weight parameters in the Wide part of the model on the k-th client.
[0169] ⑤ The server will aggregate the model parameters Broadcast to all clients.
[0170] ⑥ Repeat steps ② through ⑤. The training of the digital traffic sign recommendation model is completed when the global model completes the update round. Figure 9 This diagram illustrates the parameter updates during model training.
[0171] In addition to the horizontal federation communication module, data preprocessing module, horizontal federation encryption module, and horizontal federation digital traffic sign recommendation model training module, this invention also includes a digital traffic sign recommendation module. The digital traffic sign recommendation module is mainly used to implement steps S8 and S9.
[0172] This invention addresses the shortcomings of existing technologies, such as the lack of recommendation methods for digital traffic signs, the inability to achieve accurate recommendations based on multi-source data features including vehicles, drivers, and traffic signs, and the inability of existing technologies to solve privacy protection issues and data silos caused by data isolation (failing to address the leakage of user geolocation privacy, user preference information, and user driving information in the field of digital traffic sign recommendation). Furthermore, the inability to achieve multi-source data fusion results in insufficient training samples for recommendation models and low model accuracy. This invention provides a horizontal federated recommendation method for digital traffic signs that integrates multi-source data. It improves the training efficiency and accuracy of digital traffic sign recommendation models while protecting users, thus resolving the privacy protection, data silo, and accuracy recommendation issues in the field of digital traffic sign recommendation.
[0173] Compared with the prior art, the advantages of the present invention are as follows:
[0174] (1) This invention provides a digital traffic sign recommendation method that integrates multi-source data, including a digital traffic sign recommendation model training method that integrates multi-source data, which can recommend digital traffic signs based on vehicle positioning, road sign features (basic features, location features, other features) within a set range of driving segments, driver social features and driving behavior features, and vehicle features.
[0175] (2) This invention provides a digital traffic sign recommendation method based on horizontal federated learning. The horizontal federated learning method solves the data silo problem caused by data isolation among multiple sources during the training of the digital traffic sign recommendation model. Furthermore, the use of horizontal federated learning can effectively alleviate the cold start problem in the training of the digital traffic sign recommendation model.
[0176] Horizontal federated learning can be understood as federated learning based on sample partitioning. While the features are the same, the users differ, increasing the total number of training samples. In horizontal federated learning, data from each participant is merged horizontally, increasing the global model's training sample size. Model training is performed through model aggregation rather than data aggregation, achieving "the model runs without the data running." Each participant trains its local model in parallel, improving training efficiency and balancing performance and efficiency while preventing sensitive data leakage. Training data contains a large amount of sensitive user privacy data. Attackers can steal training data during model training or use the model's predictions as inference to obtain information about the participants building the model. Data theft attacks, attribute inference attacks, and member inference attacks all pose significant threats to user privacy and security. Horizontal federated learning can meet the needs of recommendation algorithms, enabling data sharing between different data sources, solving the data silo problem caused by data isolation between different data sources, and simultaneously meeting the requirement of protecting user personal data privacy. Furthermore, horizontal federated learning can alleviate the cold start problem to some extent. Figure 10 A schematic diagram of horizontal federated learning is shown.
[0177] (3) The present invention uses differential privacy to process the model parameters transmitted between the server and the client, thus ensuring data security.
[0178] Example 2
[0179] Embodiment 2 of the present invention provides a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable an electronic device to perform the digital traffic sign lateral federal recommendation method of Embodiment 1.
[0180] The aforementioned electronic device may be a server.
[0181] Example 3
[0182] Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the digital traffic sign lateral federation recommendation method of Embodiment 1.
[0183] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0184] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for lateral federated recommendation of digital traffic signs, characterized in that, The method includes: Step S1: Obtain the client-server architecture of the horizontal federated learning system; the client-server architecture of the horizontal federated learning system includes a server and multiple clients; Step S2: Deploy the Wide&Deep model on each of the aforementioned clients; Step S3: Train the Wide&Deep model on the client using the local training dataset corresponding to the client to obtain the parameters of the Wide&Deep model after the h-th training on each client; the initial value of h is 1; the input of the Wide&Deep model after the h-th training on the client is the local training dataset, and the output is the digital traffic sign feature data sorted in descending order of recommendation priority; the local training dataset includes digital traffic sign feature data, driver social feature data, driving behavior feature data, and vehicle feature data; Step S4: The parameters of the Wide&Deep model after the h-th training on each client are encrypted using the differential privacy method, and the encrypted parameters of the Wide&Deep model after the h-th training on each client are sent to the server using gRPC and ProtoBuf. Step S5: Aggregate the encrypted parameters of the Wide&Deep model after the h-th training on each client received by the server using a weighted average method to obtain the global model parameters after the h-th aggregation on the server; wherein, the server receives the model parameters of each client, aggregates them, and calculates the new global model parameters using a weighted average method. This indicates the iteration number of each global model update; model parameter update calculation: Model parameter weighted average calculation: In the formula, This represents the global model parameters after the model parameters of the k-th participant are aggregated to the server in the (t+1)-th round. This represents the aggregated global model parameters. Indicates the learning rate. Let represent the loss gradient of the k-th participant; n represents the number of weight parameters in the wide part of the model. k This represents the number of weight parameters in the Wide part of the model on the k-th client; Step S6: The parameters of the global model after the h-th aggregation on the server are encrypted using the differential privacy method, and the encrypted global model parameters after the h-th aggregation on the server are sent to each of the clients using gRPC and ProtoBuf. Step S7: Determine whether h is less than the preset number H; if yes, let h = h + 1, and update the Wide&Deep model on each client using the encrypted global model parameters after the hth aggregation on the server, and return to step S3; if no, update the Wide&Deep model on each client using the encrypted global model parameters after the hth aggregation on the server, and obtain the final recommendation model on each client. Step S8: Input the real-time acquired road traffic sign feature data, driver social feature data, driving behavior feature data, and vehicle feature data into the final recommendation model to obtain digital traffic sign feature data arranged in descending order of recommendation priority; Step S9: Recommend digital traffic signs based on the digital traffic sign feature data sorted in descending order of recommendation priority.
2. The digital traffic sign lateral federation recommendation method according to claim 1, characterized in that, The Wide&Deep model is deployed on each of the aforementioned clients, specifically including: The server broadcasts the initialized Wide&Deep model, model parameters, and hyperparameters to each client in plaintext; the hyperparameters include the learning rate and the number of iterations.
3. The digital traffic sign lateral federation recommendation method according to claim 1, characterized in that, The Wide&Deep model on the client is trained using the local training dataset corresponding to the client, to obtain the parameters of the Wide&Deep model after the h-th training on each client. This process also includes: Obtain the original local training dataset corresponding to each client; Convert the original local training dataset into JSON data; The JSON data is processed sequentially for missing values, outliers, and data standardization to obtain the local training dataset corresponding to each client.
4. The digital traffic sign lateral federation recommendation method according to claim 1, characterized in that, The digital traffic sign feature data includes sign code, sign name, sign category, sign content, sign shape, sign border color, sign background color, sign lining color, sign icon color, sign size, sign dynamic icon, sign static icon, sign latitude and longitude, location vector, geographical location of the road segment, latitude and longitude and location vector of the visible point, sign installation tilt angle, sign placement and left and right side position of the road segment, sign placement method, sign placement height, road segment curvature, and remarks.
5. The digital traffic sign lateral federation recommendation method according to claim 1, characterized in that, The driver's social characteristics data include name, age, ID number, mobile phone number, emergency contact, driver's license validity period, driving experience, occupation, height, weight, and illnesses.
6. The digital traffic sign lateral federation recommendation method according to claim 1, characterized in that, The driving behavior characteristic data includes driving history, driving preferences and habits, driving behavior patterns, historical accident records, and historical traffic violations.
7. The digital traffic sign lateral federation recommendation method according to claim 1, characterized in that, The vehicle characteristic data includes vehicle type, vehicle license plate, usage and driving status, vehicle type, vehicle condition, passenger capacity, empty vehicle weight, total equipment weight, permissible gross weight, maximum load capacity, maximum axle load, vehicle width, vehicle height, wheelbase, track width, front overhang, rear overhang, minimum ground clearance, maximum speed, fuel tank capacity, vehicle inspection certificate, and vehicle city access code.
8. The digital traffic sign lateral federation recommendation method according to claim 1, characterized in that, The client is a terminal device; the terminal device includes augmented reality head-up display products, advanced driver assistance systems, mobile phones, in-vehicle audio-visual systems, traffic sign information displays for smart roads, and multi-functional display boards.
9. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the digital traffic sign lateral federal recommendation method as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the digital traffic sign lateral federal recommendation method as described in any one of claims 1-8.
Citation Information
Patent Citations
Knowledge and position based individualized scenic spots recommendation method
CN105447185A
A method for recommending the location of digital signage that couples multiple source elements
CN108898244B
Distributed differential privacy recommendation method based on position service
CN110825955A
Federal personalized travel service management method and system for privacy protection
CN114742308A
Federal recommendation model training method and system
CN114741611A
