An intelligent vehicle networking stream data privacy protection method and system based on federated learning

Abstract

The application provides a kind of intelligent vehicle networking flow data privacy protection method and system based on federal learning, by in the intelligent vehicle networking framework based on federal learning, the replacement of new and old training samples is carried out to the flow data generated in the framework using pair similarity strategy, and the total size of local training data is kept unchanged using dynamic data storage strategy;Among them, the pair similarity strategy is to replace the data points in the newly generated flow data to the training sample as the data points that need to be replaced, which satisfy the preset similarity condition;Dynamic data storage strategy is to store the newly generated flow data and the replaced old data points in the buffer area.The application uses pair similarity strategy and dynamic data storage strategy to replace similar samples in the original data set using samples in the flow data and keep the number of training samples unchanged, solves the problem of catastrophic forgetting by keeping the data distribution unchanged, and also ensures the effective convergence of the federal learning training process.

Description

Technical Field

[0001] This invention belongs to the field of vehicle networking technology, specifically relating to a method and system for protecting the privacy of intelligent vehicle networking streaming data based on federated learning. Background Technology

[0002] Intelligent Vehicle-to-Vehicle (IoV) is a key component of the Fourth Industrial Revolution, fundamentally transforming vehicle communication through interconnected devices to enhance productivity and efficiency. Simultaneously, IoV networks require intelligent information processing frameworks to handle big data. The widespread adoption of Artificial Intelligence (AI) enables vehicles in IoV to train models locally, allowing them to efficiently perform tasks across various domains, from autonomous driving to smart cities and real-world intelligent vehicle-to-vehicle applications in transportation. In this context, Federation Learning (FL) has significantly enhanced IoV applications, particularly in the areas of data privacy and scalability. More precisely, FL enables vehicle nodes in IoV to collaboratively train globally shared models without directly exchanging data, thus achieving privacy protection and substantial scalability. In practice, researchers are actively involved in addressing and enhancing various challenges related to robustness, privacy, and 6G within the FL domain.

[0003] However, existing research seems to overlook a fundamental challenge within FL-enabled IoV frameworks: the processing and learning of IoV streaming data. Specifically, IoV streaming data can be defined as a continuous stream of real-time data generated or collected by IoV nodes while running locally. This category includes a range of streaming data, including but not limited to real-time surveillance video captured by cameras and instantaneous traffic conditions recorded by intelligent transportation devices. Unlike traditional static data types, streaming data is dynamic, presenting challenges in capturing fixed distribution patterns. Importantly, existing research emphasizes that streaming data can lead to catastrophic forgetting of trained models or hinder their convergence. Therefore, for FL-enabled IoV frameworks to effectively adapt to complex, dynamic IoV environments, streaming data must be integrated with local data. This integration enables continuous participation in model training, ensuring the sustained effectiveness of business operations. Summary of the Invention

[0004] In view of this, the present invention aims to provide a method and system for protecting the privacy of intelligent vehicle network streaming data based on federated learning, specifically for IoV streaming data scenarios, to improve the overall performance of IoV models, while skillfully dealing with streaming data challenges in the IoV environment, including catastrophic forgetting and convergence-related challenges.

[0005] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0006] In a first aspect, the present invention provides a method for protecting the privacy of intelligent vehicle network streaming data based on federated learning. In the intelligent vehicle network framework based on federated learning, the streaming data generated in the framework is replaced with new and old training samples using a pairwise similarity strategy, while a dynamic data storage strategy is adopted to keep the total size of the local training data unchanged.

[0007] Among them, the pairwise similarity strategy is to select data points in the old streaming data that meet the preset similarity conditions with the data points in the newly generated streaming data as the data points to be replaced, and replace the corresponding data points in the newly generated streaming data in the training samples.

[0008] The dynamic data storage strategy utilizes buffers to store newly generated streaming data as well as replacing old data points.

[0009] Furthermore, the pairwise similarity strategy uses the following formula to determine the similarity probability between data samples, as follows:

[0010] p ij =(p j∣i +p i∣j ) / 2,

[0011]

[0012] In the formula, p ij p represents the similarity probability between samples i and j. j∣i and p i∣j Let represent the similarity from j to i and from i to j, respectively; κ is the temperature hyperparameter; z i z j and z h Let n represent the feature representations of data samples i, j, and h, respectively. k It is the number of samples at the k-th node.

[0013] Furthermore, the training process for the federated learning-based intelligent vehicle-to-everything (V2X) framework includes:

[0014] IoV nodes obtain the initial global model generated by the central server during the initialization phase;

[0015] The initial global model is trained based on local training samples to obtain a local model. During the training process, a pairwise similarity strategy and a dynamic data storage strategy are used to continuously update the training samples.

[0016] IoV nodes upload the parameters of the trained local models to the central server for model aggregation.

[0017] Furthermore, the initialization phase includes:

[0018] The central server performs parameter configuration;

[0019] Select evaluation criteria, which are used to evaluate the effectiveness of the global model;

[0020] Initialize the global model and deliver the initialized global model to each IoV node.

[0021] Furthermore, during training, a pairwise similarity strategy is used for data replacement, including:

[0022] IoV nodes continuously collect new data during model training;

[0023] A pairwise similarity strategy is used to identify similar data between new data and local training samples;

[0024] Replace local training samples with similar data.

[0025] Furthermore, during training, a dynamic data storage strategy is employed to continuously update the training samples, including:

[0026] A dedicated buffer is used to store the replaced data and new data in the local training samples;

[0027] The new training samples are obtained by replacing similar data from the new data with the local training samples.

[0028] Furthermore, model aggregation involves aggregating the local model parameters from the IoV nodes to obtain the parameters of the new global model, thereby updating the global model.

[0029] Furthermore, the parameter updates of the global model are based on the following formula:

[0030]

[0031] In the formula, w represents the local model parameters, and t represents the federated learning round. It is the learning rate of client k. It is the gradient on the local data of client k, n′ k Let N be the number of samples after data replacement at the k-th node, and N′ be the total number of samples at present.

[0032] Furthermore, in each training interval of the IoV node, a pairwise similarity strategy is used for data replacement.

[0033] Secondly, the present invention also provides a federated learning-based intelligent vehicle network streaming data privacy protection system, comprising: a central server and IoV nodes, wherein the central server and IoV nodes constitute a federated learning-based intelligent vehicle network framework, wherein the streaming data generated in the framework is replaced with new and old training samples using a pairwise similarity strategy, while a dynamic data storage strategy is adopted to keep the total size of the local training data unchanged.

[0034] The pairwise similarity strategy involves selecting data points in the old streaming data that meet a preset similarity condition with the newly generated streaming data from the old streaming data used as training samples as the data points to be replaced, and replacing the corresponding data points in the newly generated streaming data with the training samples.

[0035] The dynamic data storage strategy utilizes buffers to store newly generated streaming data as well as replacing old data points.

[0036] In summary, this invention provides a method and system for protecting the privacy of streaming data in intelligent vehicle networks based on federated learning. Within the federated learning-based intelligent vehicle network framework, a pairwise similarity strategy is used to replace old and new training samples in the generated streaming data, while a dynamic data storage strategy maintains the total size of the local training data. Specifically, the pairwise similarity strategy involves selecting data points in the old streaming data that meet a preset similarity condition with the newly generated streaming data as replacement data points, and then replacing these data points in the new streaming data with the corresponding data points. The dynamic data storage strategy utilizes a buffer to store the newly generated streaming data and the replaced old data points. This invention employs both pairwise similarity and dynamic data storage strategies, using samples from the streaming data to replace similar samples in the original dataset while maintaining a constant number of training samples. This addresses the problem of catastrophic forgetting by preserving the data distribution, and ensures effective convergence of the FL training process by maintaining the invariance of the local training sample size. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0038] Figure 1 A schematic diagram of the FedStream framework provided in an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of a pairwise similarity strategy provided in an embodiment of the present invention;

[0040] Figure 3 Performance graphs of FedStream and FedAvg on the CIFAR10 dataset provided in this embodiment of the invention;

[0041] Figure 4 Performance graphs of FedStream and FedAvg on the MNIST dataset provided in this embodiment of the invention;

[0042] Figure 5 Performance graphs of FedStream and baseline during training on the CIFAR-10 dataset provided in this embodiment of the invention;

[0043] Figure 6 Performance graphs of FedStream and baseline during training on the MNIST dataset, provided for embodiments of the present invention. Detailed Implementation

[0044] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0045] The following section will first provide a further introduction to the existing technology.

[0046] As mentioned earlier, within an IoV framework that supports FL, the challenges of processing and learning streaming data are as follows:

[0047] (C1. Catastrophic Forgetting): In IoV scenarios, effectively learning a FL (Flexible Learning) model from locally generated real-time data streams via IoV devices is a significant challenge. The continuous generation of streaming data creates a major obstacle, as it forces the FL model to constantly adapt to new data distributions, potentially causing it to forget previously encountered distributions. Therefore, repeated training on streaming data samples characterized by varying distributions can easily lead to overfitting or catastrophic forgetting within the FL model. This phenomenon significantly impacts the generalization performance of IoV models.

[0048] (C2. Difficulty in Convergence): Furthermore, within the IoV framework supporting FL (Fluid Flow), the inclusion of unlearned samples introduces the problem of non-independent and identically distributed (non-IID) data, hindering the effective convergence of the FL training process. This means that when using the FL optimization method to train IoV devices, the continuously generated real-time data stream is consistently integrated into the local training data of the IoV nodes, resulting in significant non-IID problems. As the degree of non-IID intensifies, the training cost of IoV nodes continues to rise, profoundly impacting the progress of FL optimization and posing significant challenges to the computational power and convergence of IoV.

[0049] To address the aforementioned issues, lifelong learning proposes a dynamic task ranking method aimed at mitigating the impact of task ranking on model performance and forgetting rate. This method aims to adaptively adjust the task sequence, enabling the lifelong learning system to better adapt to new tasks and reduce forgetting. Kirkpatrick et al. used a technique called "Elastic Weight Combining" (EWC) to overcome the catastrophic forgetting problem in neural networks. Mendez et al. primarily used neural network-based methods to achieve lifelong learning with complex, progressively built structures. Through neural networks, the structure can continuously expand to adapt to new inputs and tasks, maintaining flexibility and adaptability. When processing real-time data streams, traditional neural network methods merge new and old data and then use this data to retrain the model, leading to a significant waste of time and space resources.

[0050] Significant progress has been made in the field of continuous learning. Researchers have proposed various methods and algorithms, including memory-based approaches, incremental learning, forgetting suppression, and meta-learning. They have also addressed challenges in continuous learning, such as the forgetting problem, and proposed corresponding solutions. A replay buffer is a set of past samples stored in memory. When a model needs to be trained on a new task, these past samples are used along with samples from the current task to train the model. By revisiting and utilizing these past samples during training, the model retains knowledge from previous tasks, thus mitigating the "catastrophic forgetting" caused by learning new tasks. Based on this idea, Matthew Riemer et al. proposed ER (Replay Buffer), which focuses on regulating model updates. However, the above scheme requires frequent access to data, which may infringe on the system's data privacy.

[0051] Existing incremental learning methods, such as knowledge extraction and retaining a small number of old classes, struggle to scale to a large number of classes. Wu et al. proposed a simple and effective method to address the data imbalance problem: correcting fully connected layers using two bias parameters and a linear model, achieving significant results. Castro et al. addressed the issue of current neural network architectures requiring the entire dataset to update models by proposing a method that utilizes only new data and a small number of samples from the old classes. Their method is based on a loss function that includes a distillation metric to retain knowledge from the old classes and a cross-entropy loss to learn the new classes. Zhang et al. proposed a class-based incremental learning paradigm called Deep Model Merging (DMC), which first trains a single model only for the new classes and then combines the two separate models using a double distillation training objective to merge the existing models. This approach overcomes the difficulty of unavailable original training data.

[0052] To address the aforementioned technical issues, this invention provides a method and system for protecting the privacy of intelligent vehicle network streaming data based on federated learning. Specifically designed for IoV streaming data scenarios, it improves the overall performance of IoV models while skillfully handling streaming data challenges in the IoV environment, including catastrophic forgetting and convergence-related challenges.

[0053] The following is a detailed description of an embodiment of the intelligent vehicle network streaming data privacy protection method based on federated learning according to the present invention.

[0054] This embodiment provides a method for protecting the privacy of streaming data in intelligent vehicle networks based on federated learning. In the intelligent vehicle network framework based on federated learning, the streaming data generated in the framework is replaced with new and old training samples using a pairwise similarity strategy, while a dynamic data storage strategy is adopted to keep the total size of the local training data unchanged.

[0055] Among them, the pairwise similarity strategy is to select data points in the old streaming data that meet the preset similarity conditions with the data points in the newly generated streaming data as the data points to be replaced, and replace the corresponding data points in the newly generated streaming data in the training samples.

[0056] The dynamic data storage strategy utilizes buffers to store newly generated streaming data as well as replacing old data points.

[0057] It should be noted that this embodiment is a privacy protection strategy specifically designed for intelligent vehicle networking. It aims to process continuously generated streaming data through the federated learning (FL) framework, while ensuring data privacy, efficient use of resources, and maintaining the model's continuous learning capability.

[0058] In the context of intelligent connected vehicles, vehicles continuously generate massive amounts of data, which are of significant value for improving driving safety and optimizing route planning. However, directly aggregating and centrally processing this data poses a risk of privacy breaches. Therefore, federated learning has become an ideal choice, allowing data to be processed locally and sharing only model updates, rather than the original data itself.

[0059] To address the challenges of processing and learning streaming data within an IoV framework that supports FL (Flexible Interpretation), this embodiment first employs a pairwise similarity strategy to ensure the model can promptly absorb new streaming data information while avoiding the storage of outdated or redundant data. This strategy requires the system to monitor newly generated streaming data and search for data points in older training samples that match its data points under a preset similarity criterion. Once a match is found, the corresponding data point in the old sample is replaced with the new data point, ensuring the model can learn the latest data patterns.

[0060] Furthermore, this embodiment employs a dynamic data storage strategy. This strategy efficiently manages storage space within the constraints of limited onboard resources. It requires the use of buffers to temporarily store newly generated streaming data and older data points that are about to be replaced. This strategy ensures that even with a continuous influx of new data, the total size of the local training dataset remains essentially constant, avoiding the problem of unlimited data growth.

[0061] The federated learning architecture itself avoids the centralization of raw data, and combined with the strategies mentioned above, further reduces the risk of data leakage. The dynamic data storage strategy rationally allocates limited onboard storage resources, ensuring the system operates efficiently under resource constraints.

[0062] The method provided in this embodiment has the following advantages:

[0063] 1. Addressing catastrophic forgetting: By replacing rather than accumulating new and old data, maintaining a consistent data distribution helps mitigate the problem of models forgetting old knowledge when learning new knowledge, thereby reducing overfitting during training. In the dynamic environment of IoV streaming data, this refinement significantly helps enhance model generalization and overall performance.

[0064] 2. Ensuring Convergence: By using a dynamic data storage strategy, the number of local training samples remains constant, ensuring that the model is trained at a uniform computational cost. This facilitates stable updates of model parameters during the federated learning process, promotes effective model convergence, and prevents the convergence speed of FL under IoV streaming data from slowing down due to increased computational overhead.

[0065] Please see Figure 1 and 2 , Figure 1 This paper presents a FL-based IoV streaming data learning framework (i.e., FedStream); Figure 2The process of implementing data replacement using a pairwise similarity strategy is illustrated.

[0066] The following combination Figure 1 and 2 Some embodiments of the present invention will be described.

[0067] like Figure 1 As shown, the key entities in this framework include: a central server, responsible for coordinating FL training and executing FL training configurations, such as model delivery, parameter configuration, and model initialization; IoV nodes, such as industrial machines, robots, and sensors, responsible for performing local training, data collection, and data processing; a local model, the model trained by the IoV node using its local data; a global model, the initial global model provided by the central server to the IoV nodes, which is updated during model iterations; and a data buffer, used to store newly collected IoV stream data.

[0068] Key functionalities of the framework include: IoV streaming data acquisition: IoV nodes continuously collect newly generated data using sensors, such as traffic flow and GPS data. Pairwise similarity strategy: Used to calculate pairwise similarity between two sets of samples for effective data replacement. Dynamic data storage strategy: Utilizes a data buffer to continuously store new data and replace old data. Model update: IoV nodes perform local training and obtain new model updates. Model aggregation: A central server collects and uses model aggregation rules to aggregate model updates and obtain a new global model.

[0069] FedStream enables seamless interaction between nodes and a central server in the context of IoV streaming data, mitigating catastrophic forgetting and convergence difficulties during training. Specifically, this invention introduces the concept of pairwise similarity and constructs an efficient streaming data training method using this strategy to address the challenges posed by streaming data in IoV. When IoV nodes perform local training, the proposed FedStream framework allows for the efficient replacement of newly collected data with locally older data and its merging into the local training set, thus making local training more robust. Furthermore, this invention uses a dynamic data storage strategy to maintain a fixed size for the local dataset, thereby reducing training overhead and accelerating convergence. Therefore, FedStream provides a practical approach for IoV streaming data processing and learning.

[0070] The following section introduces the problems of streaming data processing and learning within the framework, based on the problem definition.

[0071] First, consider an IoV network with K IoV nodes and a central server S. These nodes are responsible for locally collecting, processing, and learning real-time streaming data. Specifically, the nodes use the FL framework to train a shared global model without sharing the data. More specifically, the IoV nodes interact with the central server and seek collaboration to find a global model parameterized by ω, thereby minimizing empirical risk. Therefore, the research problem is specifically defined as follows:

[0072]

[0073] Where F(w) represents the global loss, f k (w) represents the local dataset on the i-th client. The local loss. Let n k It refers to the number of samples at the k-th node, l i Here, represents the local loss function, and w represents the local model parameters. Therefore:

[0074]

[0075] In the context of IoV, IoV nodes generate and collect dynamic sensor data in real time, which means that the node's local training dataset D i Based on the dynamic changes of data collected in real time, this invention assumes that N′=n1+...+n' k +...+n k This represents the total number of current data samples, where n is the number of samples due to the new data at IoV node k. k And N become n′ respectively k and N′. In some embodiments, the central server update is calculated as follows by aggregating updates from IoV node k:

[0076]

[0077] in, It is the gradient on the local data of client k. This is the learning rate of client k. Therefore, in the IoV streaming data scenario, the problem studied by FL undergoes some subtle changes, namely, the local data is dynamic, and in this case, the global model needs to seek effective convergence.

[0078] Based on the above problem definition, this invention introduces an efficient streaming data training strategy based on pairwise similarity, such as... Figure 2As shown. Pairwise similarity is a method used to measure the degree of similarity between two samples. In the method of this invention, a symmetric t-distributed random neighbor embedding (t-SNE) similarity measure is introduced to capture the similarity distribution between global features. Furthermore, the similarity measure of symmetric SNE is based on a Gaussian kernel function to describe the similarity between objects. Therefore, for two given samples i and j, this invention uses a Gaussian kernel function based on the Euclidean distance between objects to calculate the similarity probability, as shown below:

[0079] p ij =(p j∣i +p i∣j ) / 2,

[0080]

[0081] In the formula, p ij p represents the similarity probability between samples i and j. j∣i and p i∣j Let represent the similarity from j to i and from i to j, respectively; κ is the temperature hyperparameter; z i z j and z h Let n represent the feature representations of data samples i, j, and h, respectively. k This represents the number of samples at the k-th node. It's worth noting that the sum of the similarity probabilities of all objects equals 1.

[0082] In some embodiments, it is assumed that each IoV node initially has n k N samples. After each S training interval, the IoV node generates N′ new samples. This invention calculates the pairwise similarity between the old and new training samples, and then selects N′ samples from the initial sample set of each node that have the same label as the new samples and high pairwise similarity to replace them. In this way, this invention uses a pairwise similarity scheme to replace the features of the old and new training samples. Furthermore, this invention keeps the number of samples at each node constant during training. Therefore, this strategy addresses the problem of catastrophic forgetting by replacing training samples with pairwise similarity, thereby maintaining the data distribution unchanged.

[0083] Furthermore, this invention emphasizes that using a pairwise similarity strategy to construct an efficient streaming data training strategy can improve model performance and maintain the diversity of the training sample set. Simultaneously, the characteristic of replacing data based on pairwise similarity ensures that redundant or old data is continuously replaced, mitigating overfitting caused by repeated training samples. Moreover, the invariance of the number of local training samples at a node also ensures effective convergence of the FL training process.

[0084] In some embodiments, based on a pairwise similarity strategy and a dynamic data storage strategy, the FedStream training process includes three key stages: model initialization, data replacement, and model aggregation. The present invention will now elaborate on these three key stages as follows:

[0085] Phase 1: Initialization

[0086] Step 1: Parameter Configuration. The central server is responsible for setting multiple parameters, including the selected test dataset, the global model, the number of iterations, and the batch size. The number of iterations and the batch size are adjusted according to the size of the dataset. The key task in this stage is to ensure the rationality and consistency of the parameters.

[0087] Step 2: Develop evaluation criteria. To evaluate the effectiveness of FedStream, this invention selects accuracy as the primary evaluation criterion for IoV streaming data scenarios.

[0088] Step 3: Initial Global Model Deployment. After completing the above configuration, the central server delivers its initialized global model to each IoV node. Note that this model is the starting point for FL, providing a unified foundation for further local training in the local environment.

[0089] Phase 2: Data Replacement

[0090] Step 1: Continuous Data Collection During Training. In the second phase, nodes continuously collect new samples during the ongoing training process. This ensures that the model can adapt to dynamically changing data distributions, enhancing its sensitivity to real-time information.

[0091] Step 2: Application of Pairwise Similarity Techniques. When new data is introduced, nodes use pairwise similarity techniques to identify the set of historical data most similar to the new data for replacement. This dynamic similarity matching and data replacement helps select the most representative and relevant prior data.

[0092] Step 3: Data Filtering and Temporary Storage. After similarity matching is completed, this invention filters out the data with the highest similarity to the new data and temporarily stores it in a dedicated buffer. The purpose of this step is to avoid these already highly similar data from being repeatedly used in subsequent training, thereby improving the diversity and robustness of the model.

[0093] Step 4: Data Merging and Subsequent Training. The remaining data is then merged with the new data to participate in the subsequent training process. This step ensures that the global data distribution is maintained while the model is continuously updated. The entire process aims to fully utilize the new data while preserving the model's effectiveness and performance.

[0094] Phase 3: Model Aggregation

[0095] Step 1: Aggregate Local Model Updates. During iterative training, IoV nodes upload model updates to a central server. The central server aggregates these updates using an aggregation algorithm such as FedAvg. Note that the updated global model is then redistributed to each node. Furthermore, during iterative training, the second-stage data replacement strategy and dynamic data storage strategy need to be continuously executed until no new data influxes in and the model reaches stable convergence.

[0096] The effectiveness of the method proposed in this invention is verified through experiments.

[0097] Datasets. To evaluate the effectiveness of the proposed FedStream framework, comprehensive simulation experiments were conducted using the CIFAR-10 and MNIST datasets. Throughout the training phase, the CIFAR-10 and MNIST training sets were divided into two complete components: a client portion consisting of either a CIFAR-10 training set of 40,000 samples or an MNIST training set of 50,000 samples, and a server portion, each consisting of 10,000 samples specified for inclusion in the training scheme. A Dirichlet sampling method was employed to generate statistically heterogeneous partitions for CIFAR-10 and MNIST, with particular attention paid to setting the parameter of the Dirichlet function to 0.3.

[0098] Model Architecture. For the model architecture, this invention uses a Convolutional Neural Network (CNN), characterized by two convolutional layers, two max-pooling layers, and two fully connected layers, with ReLU as the specified activation function. The convolutional layers use a 5×5 kernel size, with the first layer having 32 kernels and the second layer having 64 kernels. After each convolutional layer, a pooling operation is performed using a (2,2) kernel. After the two convolutional layers, two additional pooling operations are performed, each using two (2,2) kernels. Finally, two fully connected layers, each with 512 and 10 neurons respectively, are used to produce the final classification output.

[0099] Hyperparameters. This invention considers a set of 10 IoV nodes, each of which uses the stochastic gradient descent (SGD) algorithm to iteratively train a local model. Throughout the experimental scheme, this invention maintains a fixed batch size m = 256, a local training epoch E = 5, and a learning rate η = 0.01. To ensure a comprehensive evaluation of the performance of the FedStream framework in complex scenarios, this invention conducted a total of 300 epochs. During training, new data is introduced every S epochs, where S takes values ​​from the set {1, 5, 10, 15, 20, 25}. Correspondingly, the number of training samples to be added is denoted as N′, where N′ takes values ​​from the set {3, 10, 30, 50, 50, 80}. Furthermore, this invention conducted two sets of ablation experiments to evaluate the effectiveness of the framework. The first set of experiments increased the sample size while keeping the epoch period constant. The second set of experiments explored variations in epoch duration while keeping the number of samples added constant.

[0100] Overall performance comparison of FedStream and FedAvg. First, this invention uses the classic FedAvg as a baseline to compare the overall performance of FedStream and the baseline scheme. For example... Figure 3 and Figure 4 As shown, across various parameter configurations (i.e., S and N), the introduction of the FedStream framework consistently outperforms the standard FedAvg method. This observation underscores the effectiveness of the FedStream framework in mitigating the detrimental effects of catastrophic forgetting associated with the introduction of new training samples, thereby maintaining model accuracy. Furthermore, when considering the frequency S of introducing new training samples, it is clear that the improvement in accuracy is not strictly proportional to the frequency of added training samples. This implies a significant trade-off between learning from existing data and learning from newly introduced data. Furthermore, this invention delves into the impact of increasing the amount of data N′ on framework performance. While the FedAvg method exhibits relatively stable accuracy on the CIFAR-10 and MNIST datasets as N′ increases, the FedStream framework demonstrates a greater improvement, indicating its powerful ability to utilize additional data. Notably, on the CIFAR-10 dataset, FedStream achieves the most significant performance improvements when S is set to 20 and N′ to 50, and when S is 25 and N′ is 80. Furthermore, on the MNIST dataset, FedStream achieved optimal results when S was set to 15 and N′ to 50, and when S was set to 25 and N′ to 80.

[0101] Figure 3 and Figure 4It also shows that the rate of accuracy improvement may exhibit diminishing returns as the number of periods S increases. This observation suggests that the model could have absorbed as much information as possible from the existing data, reaching a saturation point in its learning capacity. These insights are invaluable for subsequent experiments aimed at optimizing the frequency and magnitude of data addition to maximize learning efficiency and improve overall model performance. In conclusion, the experimental results confirm the effectiveness of the FedStream framework in efficiently learning knowledge from IoV streaming data.

[0102] A comparison of the convergence performance of FedStream and FedAvg. Next, this invention delves into a comparison of the convergence performance between the proposed FedStream framework and FedAvg. For example... Figure 5 and Figure 6 As shown, the accuracy of the FedStream algorithm on both the CIFAR-10 and MNIST datasets exhibits a rapid upward trend in the early training phase. The CIFAR-10 dataset reaches a relatively stable high accuracy level after approximately 50 training epochs, while the MNIST dataset reaches a relatively stable low accuracy level after approximately 25 training epochs, demonstrating de facto convergence. Conversely, the accuracy of the FedAvg algorithm gradually improves on both datasets, stabilizing only around 150 epochs on the CIFAR-10 dataset and around 50 epochs on the MNIST dataset. Therefore, in terms of convergence performance, the FedStream algorithm outperforms the FedAvg algorithm.

[0103] Ablation experiments. Finally, two sets of ablation experiments were conducted on the CIFAR-10 and MNIST datasets to investigate the impact of hyperparameters N′ and S on framework performance. Specifically, the number of periods S was kept constant to analyze the impact of the additional data volume N′ on accuracy. Similarly, the additional data volume N was fixed to evaluate the influence of the number of periods S on accuracy. The results are summarized in Tables 1 and 2. The experimental results show that the accuracy performance of the proposed method remains stable as the data sample size N′ increases. This stability highlights the robustness of the proposed method in various scenarios involving different data volume growth. Conversely, regardless of the fluctuation of the value of S, the proposed method exhibits relatively stable accuracy when 50 data samples are merged in each instance.

[0104] Table 1: Ablation experimental results on the MNIST dataset

[0105]

[0106] Table 2: Ablation experiments on the CIFAR-10 dataset

[0107]

[0108] The following describes an embodiment of an intelligent vehicle network streaming data privacy protection system based on federated learning according to the present invention.

[0109] This embodiment provides a federated learning-based intelligent vehicle network streaming data privacy protection system, including: a central server and IoV nodes. The central server and IoV nodes constitute a federated learning-based intelligent vehicle network framework. The streaming data generated in the framework is replaced with new and old training samples using a pairwise similarity strategy, while a dynamic data storage strategy is used to keep the total size of the local training data unchanged.

[0110] The pairwise similarity strategy involves selecting data points in the old streaming data that meet a preset similarity condition with the newly generated streaming data from the old streaming data used as training samples as the data points to be replaced, and replacing the corresponding data points in the newly generated streaming data with the training samples.

[0111] The dynamic data storage strategy utilizes buffers to store newly generated streaming data as well as replacing old data points.

[0112] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for protecting the privacy of intelligent vehicle network streaming data based on federated learning, characterized in that, In the federated learning-based intelligent vehicle network framework, the streaming data generated in the framework is replaced with new and old training samples using a pairwise similarity strategy, while a dynamic data storage strategy is adopted to keep the total size of the local training data unchanged. The pairwise similarity strategy involves selecting data points in the old streaming data used as training samples that satisfy a preset similarity condition with the newly generated streaming data as data points to be replaced, and then replacing the corresponding data points in the newly generated streaming data with the training samples. The dynamic data storage strategy involves using a buffer to store newly generated stream data as well as replaced old data points; The pairwise similarity strategy uses the following formula to determine the similarity probability between data samples, as follows: ； In the formula, This represents the similarity probability between samples i and j. and Let represent the similarity from j to i and from i to j, respectively. It is a temperature hyperparameter, z i z j and z h Let n represent the feature representations of data samples i, j, and h, respectively. k It is the number of samples at the k-th node.

2. The method for protecting the privacy of intelligent vehicle network streaming data based on federated learning according to claim 1, characterized in that, The training process of the federated learning-based intelligent vehicle networking framework includes: IoV nodes obtain the initial global model generated by the central server during the initialization phase; The initial global model is trained based on local training samples to obtain a local model. During the training process, the pairwise similarity strategy and the dynamic data storage strategy are used to continuously update the training samples. IoV nodes upload the parameters of the trained local model to the central server for model aggregation.

3. The method for protecting the privacy of intelligent vehicle network streaming data based on federated learning according to claim 2, characterized in that, The initialization phase includes: The central server performs parameter configuration; Select an evaluation criterion, which is used to evaluate the effectiveness of the global model; The global model is initialized, and the initialized global model is delivered to each IoV node.

4. The method for protecting the privacy of intelligent vehicle network streaming data based on federated learning according to claim 2, characterized in that, During training, the pairwise similarity strategy is used for data replacement, including: IoV nodes continuously collect new data during model training; The pairwise similarity strategy is used to determine similar data between the new data and the local training samples; The local training samples are replaced based on the similar data.

5. The method for protecting the privacy of intelligent vehicle network streaming data based on federated learning according to claim 4, characterized in that, During training, the dynamic data storage strategy is used to continuously update the training samples, including: A dedicated buffer is used to store the replaced data and the new data in the local training samples; The similar data in the new data are replaced in the local training samples to obtain new training samples.

6. The method for protecting the privacy of intelligent vehicle network streaming data based on federated learning according to claim 2, characterized in that, The model aggregation involves aggregating the local model parameters from the IoV nodes to obtain new parameters for the global model, thereby updating the global model.

7. The method for protecting the privacy of intelligent vehicle network streaming data based on federated learning according to claim 6, characterized in that, The parameters of the global model are updated based on the following formula: ； In the formula, For local model parameters, For federal learning rounds, It is the learning rate of client k. It is the gradient on the local data of client k. Let be the number of samples at the k-th node after data replacement. This represents the current total number of samples.

8. The method for protecting the privacy of intelligent vehicle network streaming data based on federated learning according to claim 2, characterized in that, In each training interval of the IoV node, the pairwise similarity strategy is used for data replacement.

9. A data privacy protection system for intelligent vehicle-to-everything (V2X) networks based on federated learning, characterized in that, include: The central server and IoV nodes constitute an intelligent vehicle network framework based on federated learning. In this framework, the streaming data generated in the framework is replaced with new and old training samples using a pairwise similarity strategy, while a dynamic data storage strategy is used to keep the total size of the local training data constant. The pairwise similarity strategy involves selecting data points in the old streaming data used as training samples that meet a preset similarity condition with the newly generated streaming data as data points to be replaced, and then replacing the corresponding data points in the newly generated streaming data with the training samples. The dynamic data storage strategy involves using a buffer to store newly generated stream data as well as replaced old data points; The pairwise similarity strategy uses the following formula to determine the similarity probability between data samples, as follows: ； In the formula, This represents the similarity probability between samples i and j. and Let represent the similarity from j to i and from i to j, respectively. It is a temperature hyperparameter, z i z j and z h Let n represent the feature representations of data samples i, j, and h, respectively. k It is the number of samples at the k-th node.

Images