Keywords

1 Introduction to Federated Learning

Federated Learning (FL) utilizes a centralized aggregator and provides a solution to the issues associated with many Machine Learning (ML) clients. It ensures that training data for federated learning is decentralized to protect data privacy [1]. Two key concepts of local computation and model transmission have been introduced to lower the privacy risk and cost of centralized ML systems. In FL, participants train their models by using local data and then send the model to the server for aggregation, and the server disseminates model updates. In Fig. 1, FL’s high-level map process has been given [2]. In FL, local models are trained on separate vehicles before aggregating them in the cloud to enhance security, accuracy and learning efficiency [3]. In two phase mitigating scheme, an intelligent architecture with FL provides data leakage detection [4] and intelligent data transmission [5] to improve security [6]. In Unmanned Aerial Vehicles (UAV) federated deep learning applications with wireless networks, the focus is to improve the learning efficiency, learning speed significance, conscious joint data selection, resource allotment algorithm [7], and content caching method for edge computing of FL in IoV.

1.1 Categories of Federated Learning.

FL depends on five aspects which are heterogeneity, communication architecture, data partitioning, applicable machine learning models, and privacy mechanism [8]. Figure 2 depicts the classification of FL.

Fig. 1.
figure 1

Federated learning high-level process map

Full size image

1.1.1 Data Partition

Based on various distributions of design, FL can be divided into the three types [9] i.e. horizontal FL, Vertical FL, and Federated Transfer Learning. Horizontal FL is employed where user attributes for two datasets overlay remarkably but user overlap is minimal. The data set is divided horizontally, with the same user attributes, but different users for training [10, 11]. In vertical FL (VFL), users are overlapped a lot and user features are overlapped a little. In VFL, data sets are divided vertically and a portion of data is considered for training wherever users are identical with different useful features [12]. In Federated Transfer learning (FTL) the users or the users’ attributes are never segmented. However, it can be employed in cases when there is a lag of information or tags [13].

1.1.2 Privacy Mechanism

Using FL, clients can store data locally and transmit model information for target model training. In model aggregation, the only significant aspect of FL is model aggregation that trains their global model by integrating the model attributes from all clients thus prevent transferring the metadata throughout the training process [14]. The problems associated with calculating encrypted data have been resolved using homomorphic encryption key. In differential privacy, both ML and deep learning use gradient iteration [15], which incorporates the addition of noise to the result to implement differential privacy to safeguard user privacy [16].

Fig. 2.
figure 2

The Classification of Federated Learning

Full size image

1.1.3 Machine Learning Schemes

FL enhance the capability and security of the ML model. Neural network, decision tree and linear model are the three main ML models. In federated environment, the linear model of training that mitigates the security issues and attains the same accuracy as a non-private solution is proposed in [17]. The tree model in FL, random forest and gradient boost decision trees are utilized for both single and multiple decision trees [18]. The most famous model of machine learning is the neural network model that trains complex tasks. In autonomous vehicles using drones target location, trajectory planning, target recognition services plays a vital role [19]. Due to the drawback of regular connection between UAV group and base station, centralized training fails in real time but deep learning provides excellent efficiency with UAV group usually [20].

1.1.4 Communication Architecture

The problems of application scenario of FL are equipment computing and uneven distribution of user data [21]. All the participants that are under training are in touch with centralized server for the update of global model. The communication cost of FL is high for critical problems. To minimize the cost between server and local users the model data is compressed by secondary sample random rotation quantization [22].

1.1.5 Methods for Solving Heterogeneity

Different devices affect the accuracy of total training process. The diversions of heterogeneity are model heterogeneity, fault tolerance, asynchronous communication and device sampling [23]. The main factor of FL is the efficiency of unevenly distributed data from various devices. In FL, the processing of data from various devices affects the model. Asynchronous communication is the solution to many problems with dispersed devices in FL settings.

2 Related Work

This section discusses some of the related works of FL in IoV. For intelligent object detection, a two-layer FL model has been used in a 6G supported IoV environment [24, 25]. The use of a hybrid blockchain method in addition to FL by using DRL to select optimized participants which improve the learning efficiency has been proposed in [26]. Iterative model averaging is used by the deep network federated learning frame-work to train the global model by adding the local models in each round of updates [20]. For the selection of Smart Vehicles (SVs) for FL, local learners are adopted by using round robin, random scheduling using heterogeneous asynchronous FL networks [27,28,29]. In [30], the authors concluded that the Quality of Information (QoI) received by the SVs is dynamic and will affect the performance of FL. Therefore, to improve the QoI, the Vehicular Service Provider’s (VSP’s) responsibility is to select SVs of current location within important areas. With the combination of important areas and QoI, VSP can obtain beneficial on-road data [14] and trustworthy trained model updates [31] from the chosen SVs.

In a non-collaborative Stackleberg game model proposed in [32,33,34], the mobile devices have full information about the VSPs payment budget. For estimating the traffic with FL that more correctly captures the spatio temporal correlation of the traffic flow with the use of clustering, FedGRU approach proposed in [35] combines the GRU (Gated Recurrent Unit) to get the best overall model. A model proposed in [36] FedProx integrates the edge devices data of distributed training with the Federal Averaging (FedAvg) model maximization method that improves the reliability of the target task. In [37], Multiple Principal One Agent (MPOA) based contract optimization is being employed to maximize the revenue of VSPs in each iteration [38]. OBU in IoV can gather data and glean local knowledge. A model proposed in [39] replaces data as a service by know ledge as a service in IoV. Knowledge serves as real information and is incorporated into data intelligence. IoV cars learn about the environment and the roads in different locations using ML techniques, and they share their expertise [40]. In [41], a model has been proposed to maintain balance between the dataset computational resources and wireless resources that is affected by the combination of vehicle client selection resource and wireless resource. Table 1 gives the review of previous work done by various researchers with respect to implementation of federated learning for IoV.

Table 1. Review of the Federated Learning systems for IoVs
Full size table

3 Applications of Federated Learning

This section discusses the various applications of federated learning including its applications for IoVs.

3.1 Google Keyboard

The prediction of the next word is achieved while improving the quality of the keyboard with security and privacy [7, 57]. In building the recommended systems, building of language model is also attained.

3.2 Intelligent Medical Diagnosis Systems

In the centralized method of ML, the data gathering and processing for medical diagnosis becomes difficult because of privacy and security concerns. With FL, it is possible to use the data locally without any issues of privacy and train the model for diagnosis [57, 58]. For small and insufficient labels, federated transfer learning is the solution. To implement an integrated multi FL network on APOLLO network merges the interrelated medical system’s longitudinal real world data with health outcome data for the help of doctors in forward-looking diagnosis of patients [57].

3.3 RSU Intelligence

IoV comprises of RSUs that are designed to receive data to process basic operations. Different varieties of data are received by RSUs and thus FL can be applied in various situations. One of the familiar approaches is in image processing. For autonomous vehicles, both onboard Vehicular Computing (VC) and RSU image processing is important. Collision detection and pedestrian detection also make use of image processing tasks [7].

3.4 Network Function Virtualization (NFV) Orchestration

NFV enabled network highlights the use of FL in making NFV orchestration in security/privacy services and in vehicular service delivery. Each of the networks is divided into sub-networks which may be further divided, utilizing network slicing techniques [59]. This added complexity supports the use of FL because NFV orchestrators from different network partitions can use cooperative ML training to create models capable of performing operations like VNF installation, scaling, termination, and migration. In MEC enabled orchestration of NFV, the RSU is placed near the network edge so that it can act as a network node to hold Virtual network functions (VNFs) [7].

3.5 Vehicular Intelligence

Vehicular intelligence in IoV has many applications such as forecasting the road conditions, image processing in lane detection and popular predictive maintenance [60]. The predictive maintenance uses operational data and alerts the user for maintenance of specific part by predicting the failure of the component through planned maintenance. With the use of FL, in addition to vast collection of data, predictive maintenance models have been built with greater efficiency [7].

4 Conclusions and Future Direction

Given the importance of communication in federated networks and the privacy risks associated with transferring raw data, it is mandatory to keep generated data local. In this situation, two things can be done to further minimize communication: one is to lessen the total number of iterations of communication rounds, and other is too minimize the size of message. In this paper, federated learning and its various categories have been discussed. A systematic review of different systems implementing FL in IoVs has been presented followed by applications of federated learning in different areas related to IoV. However, due to the heterogeneity of the VCs, system complexity is a major challenge. Thus, building privacy protection schemes depending on specific devices in IoV is the future direction. Resolving the tradeoff between communication cost and computational pressure is another challenge. Distributed FL is forthcoming research direction with heterogeneous data.