Intelligent networked automobile panoramic perception method based on federal learning
Through multi-task-aware federated learning network model and secret sharing method, the problems of CAV's long multi-tasking processing time and server failure in driving scenarios are solved, efficient driving scenario understanding and system security are achieved, and model quality and reliability are improved. .
Patent Information
- Application Number
- CN202411122093.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2024-11-26
AI Technical Summary
Existing intelligent connected vehicles (CAVs) are difficult to handle multiple tasks in parallel in driving scenarios, resulting in long execution times, and low-quality local models and server failures during federated learning will affect the quality of the global model and system reliability.
A multi-task-aware federated learning network model was designed, combining YOLO algorithm and secret sharing method, using cloud edge architecture for model training and aggregation, and screening high-quality models through a vehicle selection method based on reputation value to ensure the quality of the model and System security.
Multitasking parallel processing is realized, which shortens execution time, improves model performance and system reliability, avoids single point failure and inference attacks of the central server, and ensures model reusability and rapid convergence.
Smart Images

Figure CN119030976A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of mobile communications and relates to a panoramic perception method for intelligent networked vehicles based on federated learning. Background Art
[0002] With the rapid development of mobile communication technology, intelligent connected vehicles and Internet of Vehicles technology have gradually become research hotspots. The Internet of Vehicles combined with deep learning technology has greatly promoted the development of intelligent driving. CAV can collect a large amount of driving scene image data during driving, and extract and process features through convolutional neural network (CNN), so that CAV can better understand and cope with complex driving scenes. In driving scene understanding, traffic target detection can provide the location and size information of traffic participants, helping CAV to make accurate and timely decisions during the driving stage. Semantic segmentation can provide rich and accurate road information, such as drivable area division and lane line detection. At present, there are many methods that can handle these tasks separately. For example, Faster R-CNN and YOLO series algorithms can be used to handle traffic target detection problems. UNET, SegNet and pspNet can be used to handle drivable area division problems. SCNN and LaneNet can be used to handle lane line detection problems. However, the diversity and complexity of driving scenes pose severe challenges to single-task models. It is difficult for a single model to fully handle various tasks in driving scenes, and serial execution of multiple tasks will increase waiting delays, which is unacceptable for autonomous driving scenarios with high real-time requirements.
[0003] The performance of deep learning models usually depends on a large amount of high-quality data, but the data collected by a single CAV in the Internet of Vehicles is often not enough to train a model with better performance. Federated learning enables multi-vehicle data sharing and collaborative training while ensuring data localization, breaking down the barriers of information islands and improving the generalization ability of the model. The basic idea of federated learning is that each vehicle trains a model independently locally, and then uploads the local model to the federated learning server for aggregation to form a global model with better performance. However, federated learning technology still faces some challenges in practical applications. First, low-quality local models may seriously affect the quality of the global model, resulting in a decline in overall performance. Second, as a centralized node, if the federated learning server is attacked or suddenly crashes, it will cause the failure of the entire system and cause immeasurable losses.
[0004] In response to these problems, this study proposed a panoramic perception method for intelligent connected vehicles based on federated learning. This method designs a multi-task perception federated learning network model, which enables CAV to use one model to complete multiple driving tasks in parallel and understand complex driving scenarios at the same time. Compared with the single-task model, the multi-task model has a shorter execution time and is more suitable for scenarios with high real-time requirements. In order to improve the model quality in the federated learning process, this study introduced a reputation-based CAV selection method to screen out high-quality CAV local models. In addition, the secret sharing method is applied to the aggregation process of federated learning to prevent system failures caused by server attacks or downtime. Finally, the trained model is uploaded to the blockchain for storage to ensure that the model can reuse the global model in subsequent identical tasks, thereby accelerating the model convergence. Summary of the invention
[0005] In view of this, the purpose of the present invention is to provide a panoramic perception method for intelligent connected vehicles based on federated learning, which is used for CAV to perceive driving scenes during driving, so as to provide CAV with automatic driving auxiliary decision-making. CAV can cooperate with other vehicles to jointly train a global model with better performance when data is scarce.
[0006] In order to achieve the above object, the present invention provides the following technical solutions:
[0007] In a first aspect, the present invention considers the complexity of driving scenarios and the scarcity of CAV data, and provides a panoramic perception method for intelligent connected vehicles based on federated learning, the method comprising the following steps:
[0008] S1: Multi-task-aware federated learning method in cloud-edge-device architecture;
[0009] S2: Multi-task perception model based on YOLO algorithm;
[0010] S3: Vehicle selection method based on reputation value;
[0011] S4: Model aggregation method based on secret sharing.
[0012] On the second aspect, an embodiment of the present invention establishes a multi-task perception model under a cloud-edge-end architecture in S1, which consists of three network layers: 1) cloud layer, responsible for initializing the global model; 2) edge layer, including edge computing nodes such as roadside units, responsible for aggregating the global model; 3) terminal layer, composed of vehicles participating in federated learning, responsible for training local models.
[0013] RSUs are edge nodes equipped with servers deployed near roads, providing communication, storage and computing resources for vehicles. The driving scenario considered in the example of the present invention includes multiple RSUs and multiple CAVs. RSU is responsible for receiving service requests from CAVs, aggregating local model slices, and sending service request notifications to other RSUs. CAVs are responsible for sending or responding to service requests and training local models using collected data.
[0014] The RSUs at the edge layer jointly maintain the blockchain network based on the PFBT consensus, and the global model after each round of training will be stored in the blockchain.
[0015] In the third aspect, in S2, an embodiment of the present invention designs a multi-task learning model, which is divided into two parts: an encoder and a decoder. The encoder realizes feature extraction of image data. In the present invention, the target detection algorithm YOLO backbone extraction network is used to extract features of local image data of CAV. The deep feature map obtained after the backbone extraction network has strong semantic information. As the number of convolutions increases, the receptive field gradually increases, and the overlapping area between the receptive fields also increases. At this time, a pixel represents the information of an area in the original image. Therefore, the deep feature map pays more attention to the global information of the image, but due to its low resolution, the perception of details is also poor. The shallow feature map of the image contains more pixel information, which can ensure that the network captures more detailed information of the image data. Therefore, after passing through the backbone extraction network, it is also necessary to upsample the obtained feature map, and then merge it with the shallower feature map, combining the advantages of the deep and shallow feature maps to improve the detection performance of the model.
[0016] The detection head part needs to process the feature map output by the backbone extraction network to perform specific tasks. There are three different types of detection heads in the present invention, which respectively complete the tasks of traffic target detection, drivable area division and lane line detection. In the traffic target detection module, due to the influence of the relative distance of CAV, in addition to the normal-sized targets to be detected in the near distance, there are also relatively small targets to be detected in the distance in the image data collected by CAV. In order to accurately detect CAVs of various scales, in addition to deep and shallow feature fusion, three different scales of detection heads are used to detect CAVs of different sizes. Since the drivable area division task is relatively simple, it is only necessary to directly detect the feature map output by the backbone extraction network after several upsamplings. However, the lane lines in the image are often slender and difficult to detect, so in the lane line detection module, not only the feature map after feature fusion is used, but also the attention mechanism is added. The attention mechanism enables the model to focus on the lane line information in the image, thereby improving the detection accuracy of the lane line.
[0017] In the fourth aspect, an embodiment of the present invention provides a CAV selection method based on reputation value in S3. After two-stage selection, CAVs that can provide high-quality local models are screened out, thereby accelerating the convergence speed of the global model. In the first stage, all CAVs participate and evaluate each other based on the quality of the CAV local model. Since the local model provided by malicious CAVs is poor, they will receive a lower score in this stage. In the second stage, the concept of vehicle model age is proposed. For CAVs that have not participated in model aggregation for too long, their model age is older and the probability of being selected to participate in model aggregation is lower. In addition, the similarity between the local model and the global model is also considered. The local model with lower similarity to the global model has a lower probability of participating in model aggregation.
[0018] In the fifth aspect, in S4 of the embodiment of the present invention, before uploading the local model, CAV needs to use a secret sharing method to cut the local model into multiple parts and send them to different RSUs. After receiving the slices of the CAV local model, the RSU aggregates the slices of each CAV. And send the aggregated results to the P-CAV participating in the training, and the P-CAV recovers the global model based on the slices. In the traditional federated learning scenario, there is usually only one central server to perform the model aggregation operation. Once a single point failure occurs in the central server, the entire system will not be able to operate. In addition, the central server is generally trustworthy and will perform the model aggregation operation normally, but at the same time it may launch an inference attack on the local model sent by the CAV, thereby inferring the privacy information of the CAV. The model aggregation process based on the secret sharing method allows each RSU to have only partial slices of each local model, so that it is impossible to launch an inference attack on the CAV local model. In addition, due to the characteristics of secret sharing, even if some RSUs in the scene are down, P-CAV can still recover the correct global model from the aggregation results sent by other RSUs.
[0019] The beneficial effects of the present invention are:
[0020] (1) Process multiple driving tasks in parallel to shorten execution time, improve efficiency, and meet real-time requirements.
[0021] (2) Screen high-quality local models to prevent malicious models from affecting the global model and improve model performance and convergence speed.
[0022] (3) Avoid single point failure of the central server, prevent reasoning attacks, and improve system reliability and security.
[0023] (4) Ensure model reusability, accelerate model convergence, and improve training efficiency.
[0024] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:
[0026] Figure 1 This is a diagram of the CAV panoramic perception architecture based on federated learning;
[0027] Figure 2 A schematic diagram of a vehicle selection method based on reputation value;
[0028] Figure 3 Execute a flowchart for the federated learning based panoptic perception method. DETAILED DESCRIPTION
[0029] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0030] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.
[0031] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.
[0032] Figure 1 FIG. 1 shows a possible architecture diagram of CAV panoramic perception based on federated learning involved in an embodiment of the present invention. Figure 1 As shown in the figure, the system mainly consists of three parts. RSUs are responsible for aggregating local model slices sent by P-CAVs, and CAVs train local models based on perception data. The blockchain is mainly responsible for storing the global model and sending the initial global model to CAVs.
[0033] 1. Multi-task perception model based on YOLO algorithm
[0034] In the Internet of Vehicles scenario, when CAV is driving on the road, it will face various complex and changeable traffic conditions. In order to ensure the safe completion of autonomous driving tasks, CAV needs to use a variety of deep learning models that perform different tasks to assist autonomous driving. However, it takes a long time to wait for multiple models to be trained, which is unacceptable for complex and changeable Internet of Vehicles scenarios. In order to solve this problem, Figure 1 As shown, the present invention designs a multi-task learning model for processing multiple different types of tasks in parallel. First, the backbone extraction network in the target detection algorithm YOLO algorithm is used to extract features from the image data perceived by the CAV, and feature fusion is performed at the neck layer. On this basis, three different types of detection heads are added to perform traffic target detection tasks, drivable area division tasks, and lane line detection tasks, respectively. Traffic target detection enables CAV to identify potential collision risks and take preventive actions quickly to ensure driving safety. Drivable area division provides a detailed understanding of the environment for effective path planning. Lane line detection enables CAV to achieve lane keeping. These tasks provide CAV with accurate scene understanding for autonomous driving, enabling it to make correct decisions during driving. For the multi-task perception model, different detection heads share image features extracted by the backbone extraction network to improve the generalization ability of the model and speed up image processing.
[0035] Considering that the image data collected by CAV during driving is not necessarily of fixed size, a spatial pyramid pooling (SPP) layer is added to the neck layer. The SPP layer uses multiple sliding window poolings of different sizes to sample the feature map output by the last layer of the backbone extraction network, which can make the CAV input an image of any size to produce a fixed-size output. In addition, although the deep features contain a lot of semantic information, their resolution is low and the details are less preserved, so the detection of small targets is not satisfactory. The feature map of the shallow layer has a higher resolution and contains more detail information and the location information of the target. Therefore, the feature pyramid network (FPN) is used in the neck layer to fuse the low-resolution, high-semantic deep features with the high-resolution, low-semantic shallow features, so that the features at all scales have rich semantic information. The FPN module is a top-down feature pyramid that passes down the strong semantic features of the deep layer, but does not pass on the location information of the target. Therefore, a path aggregation network (PAN) is added after the FPN. The PAN module is a bottom-up feature pyramid that can complement FPN and pass on rich positioning information upward. After fusion of features at different levels, traffic targets of different sizes in the image can be detected.
[0036] In the traffic target detection module, in addition to using the feature map after feature fusion for detection, three different sizes of fused feature maps are used to complete the detection of large, medium and small traffic targets respectively. This ensures the accuracy of detection. Since the task of drivable area division is relatively simple, the detection can be completed after upsampling the feature map extracted by the backbone extraction network. The lane line is relatively slender in the driving scene, and its detection difficulty is higher than the drivable area division task. Therefore, the feature map after neck layer feature fusion is used in the lane line detection head for detection, which can effectively improve the lane line detection performance.
[0037] 2. Vehicle selection based on reputation
[0038] In federated learning, if there are malicious CAVs that intentionally upload low-quality local models and the system lacks a corresponding model screening mechanism, the global model convergence speed will slow down and the model accuracy will decrease. i From all vehicles V = {v1, v2, ..., v m} randomly select K vehicles V with high reputation value k = {v1, v2, ..., v k , ..., v K} Participate in indirect scoring. v i Using V k The public key of the vehicle in V encrypts its local model and broadcasts it to RSUs. k The vehicle in obtains the ciphertext from the RSUs and decrypts it using its own private key to obtain v i Then, V k The vehicles in v use the data they have collected i The local model is tested and the accuracy of the model output is used to evaluate v i Finally, V k The vehicles in broadcast the scoring results to RSUs. RSU can calculate the indirect score s1 for each vehicle.
[0039] For model poisoning attacks, the local model provided by the malicious CAV is of poor quality, so it will receive a low score in the first stage. In addition, the CAVs participating in the indirect scoring are randomly selected, so to a certain extent, it can avoid the phenomenon of multiple malicious CAVs working together to give each other high scores.
[0040] The concept of vehicle model age is introduced in the second stage of vehicle selection, which can be calculated by the following formula:
[0041] a i (r)=(a i (t)+1)(1-b i (r)) (1)
[0042] Among them, a i (r) is the v in the rth round of training i The model age, b i (r) represents v in the rth round of training i Whether it participates in the federated learning aggregation process. i (r) = 1, indicating v i Participate in the rth round of global aggregation, otherwise it is equal to 0. In order to ensure the timeliness of the model, RSU prefers to select CAVs with younger model ages to participate in global model aggregation. In addition, in the process of federated learning, local models that are consistent with the convergence trend of the global model are more conducive to the convergence of the global model. Therefore, v i The model similarity between its local model and the global model of the previous round will be calculated and sent to the RSU. The model similarity can be calculated by the following formula:
[0043]
[0044] Finally, the model age and model similarity of CAVs are comprehensively considered to calculate the direct score, and based on this, the P-CAVs participating in model aggregation are selected. i Direct rating in the rth round of training It can be calculated by the following formula:
[0045]
[0046] Among them, sim i (r) is the v in the rth round of training i The similarity between the local model and the global model. Based on the indirect and direct scores of CAV, the final score of CAV training in the rth round can be obtained:
[0047] s(r)=αs1(r)+βs2(r) (4)
[0048] Where s1(r) is the first-stage score of CAV, s2(r) is the second-stage score of CAV, s(r) is the final score of CAV in the rth round of training, and α and β are weight parameters. RSU will select the N vehicles with the highest final scores V = {v1, v2, ..., v N Participate in model aggregation, and after model aggregation, update the reputation value of each CAV according to its performance in this round of federated learning:
[0049]
[0050] Where C(r) is the reputation value of CAV after the rth round of federated learning, C(r-1) is the reputation value of CAV after the r-1th round of federated learning, and γ is the weight parameter.
[0051] 3. Model aggregation based on secret sharing
[0052] In traditional federated learning scenarios, all nodes participating in federated learning send local models to a central server for aggregation. Once a single point failure occurs in the central server, the entire system will not be able to operate. In addition, when performing model aggregation operations, some RSUs may launch inference attacks based on the model parameters sent by the CAV, from which they infer the privacy information of the CAV, resulting in privacy leakage.
[0053] In order to solve the above problems, the present invention uses a secret sharing method to perform global model aggregation. Secret sharing is a secret partitioning storage technology, which aims to resist multi-party collusion and intrusion to a certain extent. The core of secret sharing is to split the secret into n parts and send them to participants P1, ..., P n Secret sharing can be divided into n-out-of-n secret sharing and t-out-of-n secret sharing according to different decryption methods. n-out-of-n secret sharing requires all participants to combine to recover the secret. t-out-of-n secret sharing only requires at least t participants to combine to recover the secret.
[0054] In order to ensure that the system can work normally after some RSUs are down, the present invention uses Shamir secret sharing in the t-out-of-n secret sharing method. Even if RSUs are down, as long as there are no less than t RSUs working normally in the system, CAV can correctly restore the global model from the aggregated fragments sent by the RSUs.
[0055] Assume that there are M RSUs and N CAVs participating in the federated learning training in the scenario. CAV randomly generates a large prime number p and randomly and uniformly samples between [0, p] to obtain t-1 (t≤n) random numbers {a1, a2, ..., a t-1} and construct a secret polynomial:
[0056] f(x)=a1x+a2x 2 +...+a t-1 x t-1 +M i (6)
[0057] Among them, M i Yes i The local model of CAV is shown in Figure 2. CAV calculates f(1), f(2), ..., f(M) and sends them to M RSUs respectively.
[0058] After receiving the local model slices sent by N CAVs, the RSU performs model aggregation operations on them:
[0059]
[0060] When the CAV receives the aggregated results from at least t RSUs, it uses the Lagrangian interpolation method to recover the secret polynomial and obtain the global model parameters:
[0061]
[0062] Finally, the global model M global =g(0).
[0063] 4. System Process
[0064] Figure 3 The following is a flowchart of the panoramic perception method based on federated learning. The specific steps are as follows:
[0065] Steps 301-305: initialization phase;
[0066] Step 301: Algorithm initialization;
[0067] Step 302: The vehicle sends a service request to the RSU;
[0068] Step 303: The RSU broadcasts a service request to other RSUs, and other RSUs broadcast a service request to vehicles within their communication range;
[0069] Step 304: The remaining vehicles broadcast their own public keys;
[0070] Step 305: The blockchain sends the initial global model to the vehicle;
[0071] Steps 306-315: two-stage vehicle selection phase;
[0072] Step 306: The vehicle trains a local model using the collected image data;
[0073] Step 307: The vehicle randomly selects a vehicle with a high reputation value to score its own model;
[0074] Step 308: The vehicle encrypts its own local model using the public key of the vehicle selected to participate in the indirect scoring;
[0075] Step 309: The vehicle sends the encrypted local model to the RSU;
[0076] Step 310: The selected vehicle obtains the ciphertext from the RSU and decrypts it using the private key;
[0077] Step 311: The selected vehicle uses the collected image data to test the decrypted model and gives an indirect score;
[0078] Step 312: The vehicle calculates the similarity between the local model and the global model of the previous round and sends it to the RSU;
[0079] Step 313: RSU obtains a direct scoring result according to the model similarity and model age;
[0080] Step 314: The RSU selects vehicles to participate in model aggregation according to the results of the indirect scoring and the direct scoring;
[0081] Step 315: The RSU updates the reputation value of each vehicle according to the vehicle selection result;
[0082] Steps 316-323: Model aggregation phase;
[0083] Step 316: The vehicles participating in the global model aggregation construct secret polynomials locally;
[0084] Step 317: The vehicle slices the local model using the secret polynomial;
[0085] Step 318: The vehicle sends the local model slices to different RSUs;
[0086] Step 319: The RSU aggregates the local model slices sent by each vehicle;
[0087] Step 320: The RSU sends the aggregation result to the vehicle;
[0088] Step 321: After receiving the aggregation results from the RSU, the vehicle restores the global model through the Lagrange interpolation method
[0089] Step 322: If the global model converges, execute S323, otherwise execute step 306;
[0090] Step 323: The algorithm ends.
[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.
Claims
1. A panoramic perception method for intelligent connected vehicles based on federated learning, characterized by: The method comprises the following steps: S1: Multi-task-aware federated learning in cloud-edge architecture; S2: Establish a multi-task perception model based on the YOLO algorithm; S3: Vehicle selection based on reputation; S4: Model aggregation based on secret sharing.
2. The method for panoramic perception of intelligent connected vehicles based on federated learning according to claim 1, characterized in that: In S1, a multi-task perception model based on federated learning is established under a cloud-edge-end architecture, which consists of three network layers: a cloud layer, an edge layer, and a terminal layer; the edge layer consists of multiple roadside units (RSUs), which jointly maintain a single-layer blockchain based on a practical Byzantine Fault Tolerance (PBFT) consensus; RSUs obtain the initial global model from the blockchain and send it to connected and autonomous vehicles (CAVs), and CAVs use the global model obtained by training based on local data and upload updated local models. RSU aggregates CAV local models to generate a global model.
3. The method for panoramic perception of intelligent connected vehicles based on federated learning according to claim 2, characterized in that: In S2, the backbone extraction network in the target detection algorithm YOLO algorithm is first used to extract features from the image data perceived by the CAV, and feature fusion is performed at the neck layer. On this basis, three different types of detection heads are added to perform traffic target detection tasks, drivable area division tasks, and lane line detection tasks, respectively. Traffic target detection enables the CAV to identify potential collision risks and take preventive actions quickly to ensure driving safety. A spatial pyramid pooling (SPP) layer is added to the neck layer; the SPP layer uses multiple sliding window poolings of different sizes to sample the feature map output by the last layer of the backbone extraction network, so that CAV inputs images of any size will produce a fixed-size output; the feature pyramid network (FPN) is used in the neck layer to fuse the deep features of low resolution and high semantic information with the shallow features of high resolution and low semantic information, so that the features at all scales have rich semantic information; the FPN module is a top-down feature pyramid that passes down the deep strong semantic features, and a path aggregation network (PAN) is added after FPN; the PAN module is a bottom-up feature pyramid that supplements FPN and passes rich positioning information upward; after the fusion of features at different levels, the detection of traffic targets of different sizes in the image is realized; In the traffic target detection module, in addition to using the feature map after feature fusion for detection, three different sizes of fused feature maps are also used to complete the detection of traffic targets of large, medium and small sizes respectively; the feature map extracted by the backbone extraction network is upsampled to complete the detection; the feature map after neck layer feature fusion is used in the lane line detection head for detection.
4. The method for panoramic perception of intelligent connected vehicles based on federated learning according to claim 3 is characterized in that: In S3, first, the vehicle v i From all vehicles V = {v1,v2,…,v m } randomly select K vehicles V with high reputation value k ={v1,v2,…,v k ,…,v K } Participate in indirect scoring; v i Using V k The public key of the vehicle in V encrypts its local model and broadcasts it to RSUs; k The vehicle in obtains the ciphertext from the RSUs and decrypts it using its own private key to obtain v i The local model of k The vehicles in the v i The local model is tested and the accuracy of the model output is used to evaluate v i Finally, V k The vehicles in broadcast the scoring results to RSUs; RSU calculates the indirect score s1 for each vehicle; The concept of vehicle model age is introduced in the second stage of vehicle selection and is calculated by the following formula: a i (r)=(a i (t)+1)(1-b i (r)) (1) Among them, a i (r) is the v in the rth round of training i The model age, b i (r) represents v in the rth round of training i Whether it participates in the federated learning aggregation process; when b i (r) = 1, indicating v i Participate in the rth round of global aggregation, otherwise it is equal to 0, v i Calculate the model similarity between its local model and the global model of the previous round and send it to the RSU; the model similarity is calculated by the following formula: Finally, the model age and model similarity of CAV are comprehensively considered to calculate the direct score, and based on this, the P-CAV participating in the model aggregation is selected; i Direct rating in the rth round of training Calculated by the following formula: Among them, sim i (r) is the v in the rth round of training i The similarity between the local model and the global model; Based on the indirect score and direct score of CAV, the final score of CAV training round r is obtained: s(r)=αs1(r)+βs2(r) (4) Where s1(r) is the first-stage score of CAV, s2(r) is the second-stage score of CAV, s(r) is the final score of CAV in the rth round of training, α and β are weight parameters; RSU will select the N vehicles with the highest final scores V = {v1, v2, …, v N Participate in model aggregation, and after model aggregation, update the reputation value of each CAV according to its performance in this round of federated learning: Where C(r) is the reputation value of CAV after the rth round of federated learning, C(r-1) is the reputation value of CAV after the r-1th round of federated learning, and γ is the weight parameter.
5. The method for panoramic perception of intelligent connected vehicles based on federated learning according to claim 4, characterized in that: In S4, a secret sharing method is used to perform global model aggregation. Secret sharing is a secret partitioning and storage technology, which aims to resist multi-party collusion and intrusion to a certain extent. The core of secret sharing is to split the secret into n parts and send them to the participants P1, ..., P n ; Secret sharing is divided into n-out-of-n secret sharing and t-out-of-n secret sharing according to different decryption methods; n-out-of-n secret sharing requires all participants to combine to recover the secret; t-out-of-n secret sharing only requires at least t participants to combine to recover the secret; To ensure that the system can work normally after some RSUs fail, Shamir secret sharing in the t-out-of-n secret sharing method is used; if RSU fails, as long as there are no less than t RSUs working normally in the system, CAV can correctly recover the global model from the aggregated fragments sent by the RSU; Assume that there are M RSUs and N CAVs participating in the federated learning training in the scenario; CAV randomly generates a large prime number p, and randomly and uniformly samples t-1, t≤n random numbers {a1, a2, …, a t-1 } and construct a secret polynomial: f(x)=a1x+a2x 2 +...+a t-1 x t-1 +M i (6) Among them, M i Yes i The local model of CAV is: CAV calculates f(1), f(2), …, f(M) and sends them to M RSUs respectively; After receiving the local model slices sent by N CAVs, the RSU performs model aggregation operations on them: When the CAV receives the aggregated results from at least t RSUs, it uses the Lagrangian interpolation method to recover the secret polynomial and obtain the global model parameters: Finally, the global model M global =g(0).
Citation Information
Cited By
Self-driving automobile environment sensing method based on federal learning and automobile and road cloud cooperation
CN120543994A
Vehicle control method, device, equipment, medium and system based on vehicle key
CN120544300A
