OES-Fed: a federated learning framework in vehicular network based on noise data filtering

Federated learning and isolated data islands

As an emerging information and communication technology widely used in everyday transportation, the IoV generates a large amount of data and information interaction during its operation. However, there is an increasing focus on data privacy, leading to the isolated data island in IoV. To address this problem, many scholars have incorporated some federated learning frameworks into the IoV network. Pokhrel & Choi (2020b) proposed a new efficient communication and privacy-preserving federated learning framework for further improving the efficiency of the IoV training data. However, the framework lacks an analysis of the loss and accuracy incurred in training the network model data. Bao et al. (2021) used distributed approaches to designate some vehicles as edge vehicles and used the edge vehicles as federated learning clients for local model training, resulting in an efficient deep learning network architecture. Also, Zhao et al. (2020) advanced federated learning and local differential privacy (LDP), as well as four LDP mechanisms to scramble the gradients generated by the vehicles, to provide high accuracy with a small privacy budget. The literature (Pokhrel & Choi, 2020b; Bao et al., 2021; Zhao et al., 2020) amply illustrates that the combination of vehicular networking and federated learning can be a solution to the problem of isolated data islands in the IoV. Although the models in the literature can improve the training accuracy of the network models, the global models are susceptible to noise data, leading to biases that make the global models inefficient to learn and slow to converge.

Solutions of the noise data

To face the noise data problem, a reasonable data filter is needed to increase the training efficiency of network models, which is one of the important aspects of network model training. Data filters play an important role when dealing with large-scale datasets with a large number of samples and features (Toet, 2016; Goudarzi & Rahmani, 2021; Li, Yang & Wen, 2021). We found that the noise data can be filtered from both the current and the historical perspectives of the noise data.

• Current perspective: data outlier

Data outliers are the most used and powerful method for noise data processing. In data analysis, outliers are deviated and unexpected observations. Reunanen, Räty & Lintonen (2020) proposed a method to optimize anomaly detection integration using a limited number of outlier samples, which improves the efficiency of outlier detection by defining the limited outliers. However, this method requires some manual adjustment of the parameters or setting them according to the rule of thumb. Besides, Goh, Chiew & Foo (2020) introduced a novel method based on combined distances, capable of high accuracy outlier detection. But this did not work for outliers in high-density regions. Based on these problems, Li, Wang & Guan, 2019 proposed a graph-based outlier detection method that can significantly improve the performance of existing outlier detection methods. The method does not distinguish between local outliers and global outliers. Therefore, a new outlier detection algorithm K-Nearest Neighbor-Local Outlier Factor (KNN-LOF) was proposed by Xu et al. (2022). The K-nearest neighbour algorithm is used to region the outlier attributes, calculate the average sequence distance from data objects in the hierarchy, and redefine the reachable distance of the objects to introduce new local outlier factors. The outlier detection method is convenient and efficient and can be used for most data, which is superior to other methods. The detection results of the outlier detection method are constrained by the clustering method itself. There is no universally applicable clustering method for different data sets, so corresponding adaptations are required for the IoV scenario and data parameters.

• Historical perspective: data iteration

Kalman filter and exponential smoothing are often used to filter noise data from data iteration. Seth, Swain & Mishra (2018) used the traditional Kalman filter to estimate the position and trajectory of a single target in motion, and obtained the actual trajectory by connecting the center of the obtained moving object image. For the processing of vehicle-related data in the IoV, Zhang et al. (2018) used Kalman filter for the selection of data from simple inertial navigation and data from various positioning system sources with different errors, which can effectively improve accuracy and reliability. Exponential smoothing is a way to simplify the classification process. Therefore, Rahajoe (2019) adopted a new feature matrix as the basis for selection and used exponential smoothing to encapsulate it with the Genetic Algorithm Support Vector Machine (GASVM) method, which significantly improved the accuracy and the number of filtered parameters. In the big data era of the IoV, multiple perspectives of analysis may yield completely different results due to the diversity of data sources. Therefore, we needed to combine the Kalman filter and exponential smoothing to ensure the accuracy of the filtering results according to the actual situation of the IoV.

Since federated learning is an emerging field, its use in handling noise data is rarely covered. So this article refers to Ahmed et al. (2020), Ye et al. (2020), Li, Wang & Guan (2019), Xu et al. (2022), Seth, Swain & Mishra (2018), Zhang et al. (2018) for a comparative analysis of federated learning algorithms applied to different domains with the mechanism proposed in this article, as shown in Table 1.

System model

This article sets up a federated learning framework composed of vehicles participating in local ML parameters sharing and the central server to simulate the real scene of safe and reliable IoV. As shown in Fig. 2, the system model includes three entities: central server, access point (AP) and vehicles. The specific functions of the different entities are as follows.

1. The central server is the core processing node of IoV, with high-speed computing capability and good scalability, which can run reliably for a long time and can undertake any level of computing work in the network. Meanwhile, the server does not receive the original data collected by the vehicle, but only the local ML model parameters transmitted by the access point (AP), which are the parameters of the ML model formed by the vehicle through training the local dataset of the vehicle client. This article assumes that the server has infinite computing power.

2. The AP is the base station or roadside unit, which is equipped with communication and computational processing. In addition, the AP adjusts the training resources of the vehicle according to the instructions of the server.

3. The vehicle will generate a large amount of driving data and related pictures at the vehicle equipment during the driving process, and the driving data will be processed and saved locally, forming the local dataset: $DataLoca{l}_{n,\mathrm{i}}=\left\{{\mathrm{x}}_{\mathrm{n},\mathrm{i}}\in {\mathrm{T}}^{\mathrm{i}},{\mathrm{y}}_{\mathrm{n},\mathrm{i}}\in {\mathrm{T}}^{\mathrm{i}}\right\},{\mathrm{x}}_{\mathrm{n},\mathrm{i}}$ is the input sample vector for vehicle n to participate in training, and ${\mathrm{y}}_{\mathrm{n},\mathrm{i}}$ is the label of the input sample vector. When the server has started some task, the vehicle will train the ML model for the base and upload the local ML model parameters to the server via the AP.

Problem description

In this section, considering that multiple vehicles conduct federated learning in the Internet of Vehicles to provide the data model for the server, the instability of vehicle high speed and on-board equipment, transmission bandwidth, time delay and other factors affect the transmission of the model in the real case, so only a fraction of compliant vehicles participate in the federated learning algorithm to share local ML model parameters to improve models’ quality and efficiency. The impact of vehicle selection on the performance of the federated learning algorithm is described in the following sessions.

In the above proposed model, when the central server needs to request the vehicle dataset: $Vehicle=\left\{V1,V2,V3,V4,...Vn\right\}$, the vehicle privacy data of $n\in N$ for a particular road study, the central server posts the task and the nearby edge node AP receives the task to send to the vehicle dataset: $Vehicle=\left\{V1,V2,V3,V4,...Vn\right\}$, and waits for the feedback of vehicle resource information. From the feedback, the central server selects the appropriate vehicles to participate in the training model. The local dataset for each vehicle: $\mathrm{D}\mathrm{a}\mathrm{t}\mathrm{a}\mathrm{L}\mathrm{o}\mathrm{c}\mathrm{a}{\mathrm{l}}_{n,i}=\left\{X_{n,i}\in T^i,Y_{n,i}\in T^i\right\},X_{n,i}$ is the input sample vector for vehicle n to participate in training, $Y_{n,i}$ is the label for the input sample vector. The amount of data for all participating local training is $Data={\sum }_{n=1}^N{\sum }_{i=1}^I\mathrm{D}\mathrm{a}\mathrm{t}\mathrm{a}\mathrm{L}\mathrm{o}\mathrm{c}\mathrm{a}{\mathrm{l}}_{n,i}$, $i\in I$ is the i-th sample of the input. $n\in I$ is the participating vehicles and N is the set of vehicles participating in data sharing.

After the vehicle client has trained the local model, it transmits the parameters of the local model: $\mathrm{P}\mathrm{a}\mathrm{r}\mathrm{a}\mathrm{m}\mathrm{e}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{L}\mathrm{o}\mathrm{c}\mathrm{a}{\mathrm{l}}_{n,i}=F_{n,i}\in T^i$ to the server via the AP. $i\in I$ is the i-th parameter of the input and $n\in I$ is the participating vehicle. The server receives the parameters of the local model from each vehicle client and stores them in $\mathrm{P}\mathrm{a}\mathrm{r}\mathrm{a}\mathrm{m}\mathrm{e}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{G}\mathrm{l}\mathrm{o}\mathrm{b}\mathrm{a}{\mathrm{l}}_{n,i}={\sum }_{n=1}^N{\sum }_{i=1}^I\mathrm{P}\mathrm{a}\mathrm{r}\mathrm{a}\mathrm{m}\mathrm{e}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{L}\mathrm{o}\mathrm{c}\mathrm{a}{\mathrm{l}}_{n,i}$, $i\in I$ is the i-th parameter of the input and $n\in I$ is the participating vehicle. A weight parameter $W_n$ is defined to represent the local model parameters trained by vehicle n. The goal of the whole training process is to find the parameters $X_n$ and $Y_n$ obtained by the learning algorithm with training to make the model converge to achieve prediction accuracy and minimize the loss function. $X_n$ and $Y_n$ are the set of input sample vectors and the set of input sample vector labels. A convolutional neural network (CNN) uses the original image as input, which can effectively learn the corresponding features from a large number of samples. CNN can avoid the complicated feature extraction process and directly process images. With these advantages, CNN has been widely used in image processing (Kim & Hong, 2020; Wang, Ma & Wu, 2020; Poudyal et al., 2021). For example, Yang et al. (2021a, 2021b) proposed a Multi-scale Texture Difference model (MTD-Net) and Multi-scale Self-Texture Attention Generative Network (MSTA-Net) based on improved CNN, respectively, to detect forged data. Therefore, we chose the CNN as the learning algorithm in this article. The net input of the i-th parameters mapping of the first layer $Z^{(l,i)}$ is as follows.

(1) $$Z^{(l,i)}=\sum _{d=1}^D{W}^{(l,i,d)}\otimes X^{(l-1,d)}+b^{(l,i)}$$

$X^{(l-1,d)}$ is the input parameters mapping for the first layer, $X^{(l-1,d)}\in R^{M\times N\times D}$. Each output parameters mapping requires D convolution kernels as well as a bias. The predicted results ${\hat{\mathrm{Y}}}_{\mathrm{n},\mathrm{i}}$ are as follows.

(2) $${\hat{\mathrm{Y}}}_{\mathrm{n},\mathrm{i}}={\mathrm{Z}}_{\mathrm{n},\mathrm{i}}{\mathrm{W}}_{\mathrm{n}}^{\mathrm{T}}+{\mathrm{b}}_{\mathrm{n},\mathrm{i}}$$

Therefore, the prediction accuracy function of the federated learning model is as follows.

(3) $$Accuracy={\displaystyle \frac{1}{N}\sum _{i=1}^N{\hat{\mathrm{Y}}}_{\mathrm{n},\mathrm{i}}={\mathrm{Y}}_{\mathrm{n},\mathrm{i}}}$$

${\hat{\mathrm{Y}}}_{\mathrm{n},\mathrm{i}}$ is the true label if it is the predicted label trained on the test set using the trained classification model. For the loss function of the federated learning, we chose the logistic regression method to describe it, then the loss function of the local model is as follows.

(4) $$f_n(w)={\log }_2\left(1+exp\left(-Y_{n,i}{w}_n^T{X}_{n,i}\right)\right)$$

The objective is the formula (2).

(5) $$minf(w)\triangleq min\left\{{\displaystyle \frac{1}{N}\sum _{n=1}^N{f}_n(w)}\right\}$$

Each vehicle n updates the model in round e $w_n^e=w_n^{e-1}-l\left(w_n^{n-1}\right)$

l is a predefined learning rate, and then the updated model parameters are passed through the edge nodes to the central server, which trains the e round global model parameters.

(6) $$w_e={\displaystyle \frac{1}{N}\sum _{n=1}^N\sum _{i=1}^I{\displaystyle \frac{w_n^e\mathrm{D}\mathrm{a}\mathrm{t}\mathrm{a}\mathrm{L}\mathrm{o}\mathrm{c}\mathrm{a}{\mathrm{l}}_{n,i}}{DataLocal}}}$$

The probability of each vehicle being selected is as follows.

(7) $$p_n=\sum _{i=1}^I{\displaystyle \frac{\mathrm{D}\mathrm{a}\mathrm{t}\mathrm{a}\mathrm{L}\mathrm{o}\mathrm{c}\mathrm{a}{\mathrm{l}}_{n,i}}{DataLocal}}$$

We validated the model capability by calculating the AUC values of the completed global model for the training, setting the dataset with a total of M positive samples, N negative samples, and M + N predictions ${\hat{\mathrm{Y}}}_{\mathrm{n},\mathrm{i}}$.

(8) $$\mathrm{A}\mathrm{U}\mathrm{C}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}1\left\{{\hat{\mathrm{Y}}}_{\mathrm{n},\mathrm{i}}={\mathrm{Y}}_{n,\mathrm{i}}\right\}}{\mathrm{M}\times \mathrm{N}}}$$

According to formulas (1) and (2), vehicles with highly accurate and reliable local ML model parameters can converge local loss functions $f_n(w)$ and global model parameters f (w) faster. From the formula (3), it can be seen that the global ML model parameters trained by the central server depend on the local ML model parameters transmitted by the vehicle, as well as the quality of the training dataset. On the contrary, the global ML model parameters in turn determine the local model updates. Therefore, the local ML model parameters must be selected with high accuracy, low loss and better AUC to achieve convergence between local and global model updates in fewer iterations. In addition, the learning efficiency of federated learning can be improved significantly.

Phase A: current perspective of nosie data

Phase B: historical perspective of noise data

Phase C: the overall scheme of data iteration and resource alignment

Current and historical perspective’s noise data filtering results

Figure 4 shows the actual number of clients integrated by the OES-Fed model per round of cloud server after filtering by outliers with the Kalman filter when running the MNIST dataset, the CIFAR-10 dataset and the vehicle classification dataset. In other words, the results of client filtering by current noise perspective and historical noise perspective are shown. The black line in Fig. 4 shows the number of clients per round after removing outliers from the current noise perspective, and the red part shows the actual number of clients per round in the end from the combined historical noise perspective. When using the MNIST dataset, the OES-Fed model retains as many client model parameters as possible, while the number of actual clients tends to level off as the number of rounds increases. When using the CIFAR-10 dataset and the vehicle classification dataset, more clients are screened out due to their current round performance vs their historical performance. As a result, outliers and the Kalman filter play a key role in the filtering of clients.

OES-FED framework’s accuracy, loss and AUC results

As shown in Fig. 5, we compared the accuracy of the three models, FedAVG, FedSGD and OES-Fed, using the MNIST and CIFAR-10 datasets. The accuracy of the three models continued to improve as the number of global communication rounds increased. For the MNIST, the model accuracy of OES-Fed improved by about 2.36% over that of FedAVG and 44.46% over that of FedSGD. For the CIFAR-10, the model accuracy of OES-Fed was approximately 42.02% higher than that of FedAVG and about 47.6% higher than that of FedSGD. For the vehicle classification dataset, the accuracy of the OES-Fed was 32.4 % higher than FedAVG, and 68.0 % higher than FedSGD.

As shown in Fig. 6, we compared the loss values of the three models, FedAVG, FedSGD and OES-Fed, using the MNIST and CIFAR-10 datasets. The losses of the three models kept increasing as the number of global communication rounds increased. In the case of MNIST, the model loss of OES-Fed was approximately 0.23 lower than that of FedAVG and approximately 2.11 lower than that of FedSGD. In the case of CIFAR-10, the model loss of OES-Fed was approximately 0.38 lower than that of FedAVG and approximately 0.44 lower than that of FedSGD. For the vehicle classification dataset, the loss of OES-Fed is 0.59 lower than FedAVG and 1.27 lower than FedSGD.

As shown in Figs. 7–9, we compared the AUC values of the FedAVG, FedSGD and OES-Fed models using the MNIST, CIFAR-10 and the vehicle classification datasets. From the results shown in the figure, the AUC values of OES-Fed were 0.22 higher than those of FedAVG and 0.32 higher than those of FedSGD in the MNIST. The AUC values of OES-Fed were 0.14 higher than those of FedAVG and 0.17 higher than those of FedSGD in the CIFAR-10. In the vehicle classification dataset, the AUC value of OES-Fed was 0.27 higher than FedAVG and 0.33 higher than FedSGD.

Figure 10 shows the accuracy comparison of all clients at the final round using the OES-Fed model and the FedAVG model for the MNIST dataset and the CIFAR-10 dataset, respectively. The red line is the OES-Fed accuracy for the 40 clients in the final round, sorted from smallest to largest; the black line is the FedAVG accuracy for the 40 clients in the final round, also sorted from smallest to largest. The grey-shaded force between the red and black lines is the difference in accuracy between this OES-Fed and FedAVG. In other words, the larger the shaded area, the larger the difference. When the dataset was MNIST, the OES-Fed model was used to make more substantial improvement in the accuracy of the worse vehicles. When the dataset was CIFAR-10, the accuracy of each vehicle in the OES-Fed model improved by 30% compared to the FedAVG model. For the vehicle classification dataset, the accuracy of each vehicle in the OES-Fed model improved by 40% compared to the FedAVG model.