Federal learning system and method applied to Internet of Vehicles

Abstract

The invention relates to a federated learning system and method applied to Internet of Vehicles, the system comprises a plurality of CAVs, RSUs and a cloud server, the plurality of CAVs are respectively in communication connection with the RSUs, the RSUs are in communication connection with the cloud server, and the CAVs are used for compressing and encrypting collected data and then transmitting the data to the RSUs; the RSU is used for training a local model, compressing and encrypting the trained model and then transmitting the model to the cloud server; an aggregation model issued by the cloud server is received, and an available global model is obtained through decryption and reconstruction operation and used for the next round of model training; and the cloud server is used for receiving the models uploaded by the RSUs and performing aggregation processing to obtain an aggregation model. Compared with the prior art, the method can solve the defects of insufficient vehicle computing power and insufficient privacy protection of the federated learning scheme in the existing Internet of Vehicles, and realizes an efficient and privacy-protected federated learning scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Vehicles, and in particular to a federated learning system and method applied to the Internet of Vehicles. Background Art

[0002] Machine Learning (ML) has recently been widely used in connected vehicle applications such as autonomous driving, road safety prediction, and vehicle target detection. The current trend of using machine learning in vehicle networks focuses on centralized algorithms, called centralized learning (CL), where powerful learning algorithms (usually neural networks) are trained on large datasets collected from vehicle edge devices. The neural network model provides a nonlinear mapping between inputs and outputs containing vehicle sensor data. This mapping is learned by training the neural network with data collected from the vehicle, which is a supervised learning scheme. Once the model training is completed, the model parameters are sent back to the edge device for prediction. However, in order to successfully train a better performing neural network architecture, the scale of data generated is huge. It can be said that the centralized learning approach will bring huge overhead to the data transmission between the parameter server and the on-board edge device. In addition, training the model in a reliable way to transmit data from the vehicle to the cloud center may be too expensive in terms of bandwidth, while generating unacceptable delays and infringing user privacy.

[0003] In the prior art, Federated Learning (FL) has been introduced into vehicle networks, with the goal of making full use of edge devices for training neural networks. In FL, edge devices only send the gradients of learnable parameters derived from these local datasets to the cloud server instead of their local datasets. The cloud server aggregates these gradients and updates the model parameters, which are then transmitted back to the edge device. This process is iterated until convergence. The training process is similar to CL, except that FL does not involve the transmission of the entire dataset. This not only reduces the transmission overhead, but also provides a certain degree of privacy. Therefore, it enables more nodes to participate in model training, so that the model can better adapt to local changes in the global dataset and enhance the generalization of the final trained model.

[0004] As mentioned above, although FL has advantages over CL in terms of communication overhead and privacy, FL requires computing resources from edge devices to perform model calculations. Therefore, not all vehicles can undertake the task of model training due to insufficient computing power. At the same time, since attackers can restore the original training data through model parameters, directly transmitting unprotected model parameters between edge devices and cloud servers still poses privacy risks.

[0005] In summary, in the Internet of Vehicles, federated learning requires CAV (Connected Autonomous Vehicle) to perform local training of the model, so CAV must have strong computing power. However, this cannot be met in practice, that is, a large number of CAVs are excluded from the federated learning system due to insufficient computing power. At the same time, direct transmission of model parameters still brings privacy risks, and attackers can recover the original training data through model parameters. As the model becomes more complex, the transmission of model parameters will also bring higher transmission costs and unacceptable delays. The above problems will affect the application of federated learning in the Internet of Vehicles. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a federated learning system and method applied to the Internet of Vehicles, which can solve the defects of insufficient vehicle computing power and insufficient privacy protection faced by the existing federated learning schemes in the Internet of Vehicles, and realize an efficient and privacy-protecting federated learning scheme.

[0007] The object of the present invention can be achieved by the following technical solutions: A federated learning system applied to an Internet of Vehicles, comprising a plurality of CAVs (Connected Autonomous Vehicles), RSUs (Road Side Units) and a cloud server, wherein the plurality of CAVs are respectively connected to the RSUs for communication, the RSUs are connected to the cloud server for communication, and the CAVs are used to compress and encrypt the collected data and transmit them to the RSUs;

[0008] The RSU is used to perform local model training, and transmit the trained model to the cloud server after compression and encryption; and receive the aggregated model sent from the cloud server, and obtain a usable global model through decryption and reconstruction operations for the next round of model training;

[0009] The cloud server is used to receive the models uploaded by each RSU and perform aggregation processing to obtain an aggregated model.

[0010] Furthermore, the CAV is communicatively connected to the RSU that is closest to it.

[0011] Furthermore, the CAV is provided with a first compressed sensing module for compressing and sampling the data collected by the CAV and performing encryption processing.

[0012] Furthermore, a reconstruction module and a second compressed sensing module are provided on the RSU, and the reconstruction module is used to decrypt and reconstruct the data uploaded by the CAV and the aggregation model sent by the cloud server to obtain the original data and the available global model, and the original data is used for local model training;

[0013] The second compressed sensing module is used to compress and encrypt the trained local model.

[0014] Furthermore, an aggregation module is provided on the cloud server for aggregating the models uploaded by each RSU to obtain an aggregated model.

[0015] A federated learning method applied to an Internet of Vehicles includes the following steps:

[0016] S1. Each CAV collects data, compresses and encrypts the collected data, and transmits it to the nearest RSU;

[0017] S2, RSU decrypts and reconstructs the received data to obtain the original training data, and uses the original training data to perform local model training;

[0018] S3. Each RSU compresses and encrypts the trained model and transmits it to the cloud server. The cloud server aggregates the received models to obtain the aggregated model and sends it to each RSU.

[0019] S4. RSU decrypts and reconstructs the received aggregate model to obtain a usable global model, which is used for the next round of local model training.

[0020] Furthermore, the data collected by the CAV in step S1 includes image data.

[0021] Furthermore, in step S2, the RSU specifically uses a reconstruction algorithm to decrypt and reconstruct the received data;

[0022] In the step S4, the RSU specifically uses a reconstruction algorithm to decrypt and reconstruct the received aggregation model.

[0023] Furthermore, the reconstruction algorithm includes an iterative hard threshold algorithm and a greedy iterative algorithm.

[0024] Furthermore, in step S3, the process of the cloud server aggregating the received models is as follows:

[0025] The cloud server receives the compressed and encrypted model parameters submitted by each RSU, and performs weighted aggregation operations on the received ciphertext data, where the weight is determined by the amount of data used by the corresponding RSU to train the model.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] In the present invention, multiple CAVs are designed to communicate with RSUs respectively, and RSUs are communicated with cloud servers. CAVs are used to compress and encrypt the collected data and transmit them to RSUs; RSUs are used to perform local model training, and the trained models are compressed and encrypted and transmitted to cloud servers, as well as to receive aggregated models sent from cloud servers, and obtain available global models through decryption and reconstruction operations for the next round of model training; cloud servers are used to aggregate the models uploaded by each RSU to obtain an aggregated model. As a result, CAVs are no longer terminal devices for model training, but the training tasks are transferred to RSUs with stronger computing power, which can reduce the computing overhead of CAVs and enable more CAV data to be used for model training. Compression encryption technology is used on both CAVs and RSUs to achieve compressed sampling and encryption of data, which can further reduce the communication overhead of the data transmission process and provide privacy protection for the transmitted data, thus realizing an efficient and privacy-protecting federated learning solution, ensuring that more vehicle data can be used for model training, and enhancing the availability of traditional vehicle network federated learning.

[0028] In the present invention, the CAV is connected to the nearest RSU by communication, so that the transmission delay is lowered and the efficiency of the whole system is ensured.

[0029] In the present invention, on the one hand, CAV compresses and encrypts the collected data and transmits it to RSU, and on the other hand, RSU also compresses and encrypts the trained model using compressed sensing technology. This can not only effectively encrypt common image data in the field of vehicle networks and achieve privacy protection of private data, but also reduce the communication overhead of model transmission and achieve privacy protection of model parameters.

[0030] In the present invention, the cloud server receives the compressed and encrypted model parameters submitted by each RSU, and performs weighted aggregation operations on the received ciphertext data, wherein the weight is determined by the amount of data used by the corresponding RSU for training the model, which can fully ensure the reliability of the aggregation operation and obtain an accurate aggregation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0032] Figure 2 It is a schematic diagram of the method flow of the present invention;

[0033] Figure 3 A schematic diagram of a federated learning framework constructed for an embodiment;

[0034] Description of the marks in the figure: 1. CAV, 2. RSU, 3. Cloud server. DETAILED DESCRIPTION

[0035] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example

[0037] like Figure 1 As shown, a federated learning system applied to the Internet of Vehicles includes multiple CAV1, RSU2 and a cloud server 3, multiple CAV1 are respectively connected to the RSU2 for communication, RSU2 is connected to the cloud server 3 for communication, and CAV1 is used to compress and encrypt the collected data and transmit it to RSU2;

[0038] RSU2 is used to perform local model training, and transmit the trained model to the cloud server 3 after compression and encryption; and receive the aggregated model sent from the cloud server 3, and obtain a usable global model through decryption and reconstruction operations for the next round of model training;

[0039] The cloud server 3 is used to receive the models uploaded by each RSU 2 and perform aggregation processing to obtain an aggregated model.

[0040] Among them, CAV1 is in communication connection with the RSU2 which is closest to it, and a first compressed sensing module is provided on CAV1 for compressing sampling and encrypting the data collected by CAV1;

[0041] RSU2 is provided with a reconstruction module and a second compressed sensing module. The reconstruction module is used to decrypt and reconstruct the data uploaded by CAV1 and the aggregation model sent by the cloud server 3 to obtain the original data and the available global model. The original data is used for local model training.

[0042] The second compressed sensing module is used to compress and encrypt the trained local model.

[0043] In addition, an aggregation module is provided on the cloud server 3 for aggregating the models uploaded by each RSU 2 to obtain an aggregated model.

[0044] Based on the above system, a federated learning method applied to the Internet of Vehicles is implemented, such as Figure 2 As shown, the following steps are included:

[0045] S1. Each CAV collects data, compresses and encrypts the collected data, and transmits it to the nearest RSU;

[0046] S2, RSU decrypts and reconstructs the received data to obtain the original training data, and uses the original training data to perform local model training;

[0047] S3. Each RSU compresses and encrypts the trained model and transmits it to the cloud server. The cloud server aggregates the received models to obtain the aggregated model and sends it to each RSU.

[0048] S4. RSU decrypts and reconstructs the received aggregate model to obtain a usable global model, which is used for the next round of local model training.

[0049] This embodiment applies the above solution to build Figure 3 The federated learning framework shown in the figure has the following working processes:

[0050] Step 1: CAV collects data sets;

[0051] Step 2: CAV compresses and encrypts the data needed for model training, such as image data, and transmits it to RSU;

[0052] Step 3: RSU decrypts and reconstructs the received data to obtain the original training data;

[0053] Step 4: RSU uses the original data for model training;

[0054] Step 5: RSU compresses and encrypts the model update obtained through training and transmits it to the cloud server;

[0055] Step 6: The cloud server receives all data transmitted by the RSU;

[0056] Step 7: The cloud server aggregates the received model updates;

[0057] Step 8: The cloud server sends the aggregated model update to each RSU;

[0058] Step 9: RSU decrypts and reconstructs the received aggregated model update to obtain a usable global model for the next round of training.

[0059] In summary, the federated learning system involved in this solution includes three types of entities, namely, networked autonomous vehicles (CAVs), roadside units (RSUs), and cloud servers. These three types of entities have different computing and storage capabilities and assume different functions in the entire system: CAVs have certain computing and storage capabilities and are mainly used for data collection and preprocessing;

[0060] RSU has strong computing and storage capabilities and is mainly responsible for the training of deep learning models;

[0061] The cloud server mainly aggregates the received model updates and updates the global model.

[0062] CAV will transmit the collected data to the nearest RSU. At the same time, in order to reduce the communication overhead of transmitted data and protect the privacy of transmitted data, CAV will use compressed sensing technology to achieve compression and privacy protection of transmitted data.

[0063] RSU will use the data received from CAV transmissions to regularly perform local model training, and transmit the trained model updates to the cloud server for model aggregation. RSU will also use compressed sensing technology to reduce the communication overhead caused by complex model transmission and achieve privacy protection of model parameters.

[0064] The cloud server aggregates the measurement values ​​after model update compression submitted by the RSU, but does not reconstruct the measurement values.

[0065] In this solution, CAV is no longer a terminal device for model training, but transfers the training task to the RSU with stronger computing power. CAV transmits the collected data to the nearest RSU for local model training. This not only reduces the computing overhead of CAV, so that more CAV data can be used for model training, but also because the distance between RSU and CAV is relatively close, the transmission delay is lower, ensuring the efficiency of the entire system. In order to further reduce the communication overhead of the process and provide privacy protection for the transmitted data, this solution proposes to use compressed sensing technology to achieve data compression sampling and encryption. This technology can not only significantly reduce the transmission cost, but also as a lightweight encryption scheme, it can effectively encrypt common image data in the field of vehicle networks to achieve privacy protection of private data. RSU will first use the reconstruction algorithm to reconstruct and decrypt the received data to obtain the original data, and then use the data for model training. Common reconstruction algorithms include iterative hard threshold algorithm, greedy iterative algorithm, etc. When the local model training is completed, RSU will also use compressed sensing technology to compress and encrypt the model update, so as to reduce the communication overhead of model transmission and privacy protection of model parameters. The cloud server will receive the compressed and encrypted model parameters submitted by each RSU, and perform weighted aggregation on the received ciphertext data, with the weight determined by the amount of data used by the RSU to train the model. The aggregated model is sent to each RSU. The RSU then reconstructs and decrypts the received aggregated model to obtain a usable aggregated model for the next round of model training.

[0066] This solution optimizes and improves the federated learning solution in vehicle networks, and implements an efficient and privacy-preserving federated learning framework, ensuring that more vehicle data can be used for model training, enhancing the usability of traditional vehicle network federated learning, and effectively reducing transmission overhead and the computational burden of servers. At the same time, it achieves privacy protection of sensitive data by transmitting model parameters.

Claims

1. A federated learning system applied to Internet of Vehicles, characterized in that: It includes a plurality of CAVs, RSUs and a cloud server, wherein the plurality of CAVs are respectively connected to the RSUs in communication, the RSUs are connected to the cloud server in communication, and the CAVs are used to compress and encrypt the collected data and transmit them to the RSUs; The RSU is used to perform local model training, and compress and encrypt the trained model before transmitting it to the cloud server; And receive the aggregated model sent from the cloud server, and obtain a usable global model through decryption and reconstruction operations for the next round of model training; The cloud server is used to receive the models uploaded by each RSU and perform aggregation processing to obtain an aggregated model.

2. A federated learning system applied to Internet of Vehicles according to claim 1, characterized in that: The CAV is in communication connection with the RSU closest to it.

3. The federated learning system applied to the Internet of Vehicles according to claim 1, characterized in that: The CAV is provided with a first compressed sensing module for compressing and sampling the data collected by the CAV and performing encryption processing.

4. The federated learning system applied to the Internet of Vehicles according to claim 1, characterized in that: The RSU is provided with a reconstruction module and a second compressed sensing module, wherein the reconstruction module is used to decrypt and reconstruct the data uploaded by the CAV and the aggregation model sent by the cloud server to obtain the original data and the available global model, and the original data is used for local model training; The second compressed sensing module is used to compress and encrypt the trained local model.

5. The federated learning system applied to the Internet of Vehicles according to claim 1, characterized in that: The cloud server is provided with an aggregation module for aggregating the models uploaded by each RSU to obtain an aggregated model.

6. A federated learning method applied to an Internet of Vehicles, implemented based on a federated learning system applied to an Internet of Vehicles as claimed in claim 1, characterized in that: The following steps are involved: S1. Each CAV collects data, compresses and encrypts the collected data, and transmits it to the nearest RSU; S2, RSU decrypts and reconstructs the received data to obtain the original training data, and uses the original training data to perform local model training; S3. Each RSU compresses and encrypts the trained model and transmits it to the cloud server. The cloud server aggregates the received models to obtain the aggregated model and sends it to each RSU. S4. RSU decrypts and reconstructs the received aggregate model to obtain a usable global model, which is used for the next round of local model training.

7. The method for federated learning applied to Internet of Vehicles according to claim 6, characterized in that: The data collected by the CAV in step S1 includes image data.

8. The method for federated learning applied to Internet of Vehicles according to claim 6, characterized in that: In step S2, the RSU specifically uses a reconstruction algorithm to decrypt and reconstruct the received data; In the step S4, the RSU specifically uses a reconstruction algorithm to decrypt and reconstruct the received aggregation model.

9. The method for federated learning applied to Internet of Vehicles according to claim 8, characterized in that: The reconstruction algorithm includes an iterative hard threshold algorithm and a greedy iterative algorithm.

10. The method for federated learning applied to Internet of Vehicles according to claim 6, characterized in that: The process of the cloud server aggregating the received models in step S3 is as follows: The cloud server receives the compressed and encrypted model parameters submitted by each RSU, and performs weighted aggregation operations on the received ciphertext data, where the weight is determined by the amount of data used by the corresponding RSU to train the model.

Images