A reliable and fair federated learning mechanism for mobile edge computing
Introduction
The rapid growth of IoT and mobile web applications has led to an exponential growth of data generated at the edge of the network, which makes mobile edge computing rapidly popular. To ensure that training data is retained on individual devices and to facilitate collaborative machine learning of complex models among distributed mobile devices, federated learning (FL) was introduced to mobile edge computing. Implementing FL on mobile edge networks for model training has the advantages of efficient utilization of network bandwidth, privacy preservation, and low latency [1], [2]. It can be used in a lot of scenarios such as mobile group awareness and Internet of vehicles [3], [4].
While FL brings many benefits to mobile edge computing, it still faces critical challenges. On one hand, the variability in data volume, data quality and communication capabilities of end nodes will inevitably affect the reliability of FL model training. On the other hand, dishonest end nodes may send malicious update parameters to pollute the global model or leave halfway with malicious intent, thus affecting the reliability and fairness of model training. Therefore, it is of vital importance to achieve reliable, fair and highly accurate FL model training for mobile edge computing.
Some scholars solve the problem above by selecting high performance and high reliability end nodes [5], [6], which are mainly divided into three categories: (1) resource information-based selection [7], [8], which selects end nodes considering end nodes’ own resource information, such as computational power, channel state, data size, CPU, memory, and energy consumption; (2) reputation-based selection [9], [10], which evaluates the reputation of end nodes based on their performance in past interactions such as the number of positive behaviors and negative behaviors and selects the end nodes with high reputation for training; (3) probability allocation-based selection [11], [12], which is to model the determinants of client selection as a dynamic probability problem such as the probability of failure and performing client selection accordingly.
It is easy to find that the aforementioned studies focus on the selection of end nodes before the training, ignoring the differences in the end nodes’ resistance to corruption during the training process. For example, end nodes with poor resistance to corruption can easily be controlled by malicious nodes during the training process, and launch poisoning attacks (send malicious updates to pollute the global model) or leave halfway with malicious intent after stealing useful information. When an end node launches a poisoning attack during the training process, there is a significant decrease in the accuracy of FL with current aggregation algorithms, such as Federated Averaging (FedAvg) [13]. Meanwhile, the reputation of end nodes is usually constructed based on public information such as the reputation on the blockchain, and end nodes with high reputation values can easily invite corrosion and further launch poisoning attacks. What is more, end nodes escape from training after getting useful information from other end nodes, which is unfair to end nodes who honestly perform model training in the full process, causing a serious fairness problem.
In order to solve the above problems, we propose a reliable and fair federated learning mechanism for mobile edge computing. In the mechanism, a reputation-based end node selection scheme is proposed to select end nodes as reliable as possible to participate in FL training. An aggregation strategy with enhanced robustness is proposed to further avoid the negative influence when honest end nodes are corrupted. A reward-penalty scheme is proposed to ensure fairness. Finally, the performance of the mechanism is evaluated experimentally. The main contributions of the paper are as follows:
- •
To improve reliability, we propose a reputation-based mobile edge nodes selection scheme with steganography. Reputation models are constructed considering multidimensional factors such as the characteristics of the device and historical performance. Also, effective hiding of the selected end nodes is achieved based on verifiable random functions (VRF).
- •
To address vulnerability to attacks, we propose an aggregation strategy based on meritocratic campaigning that allows the model to maintain high accuracy in the face of poisoning attacks launched by compromised end nodes.
- •
To ensure fairness, we propose a reward-penalty scheme. Through contribution measurement, end nodes that behave honestly are rewarded with honoraria commensurate with their contributions, while those with malicious behavior are punished.
- •
In order to evaluate the performance of our mechanism, we conduct experiments using the MNIST dataset [14] and HAPT dataset [15]. Experimental results show that our mechanism maintains high accuracy under attacks by malicious end nodes while guaranteeing fairness.
The rest of this paper is organized as follows. In Section 2, the related work is reviewed. In Section 3, the scenario, adversary model and workflow are introduced. In Section 4, the proposed client selection scheme, the aggregation strategy and the reward-penalty scheme are specifically described. In Section 5, the simulation results are given. Section 6 concludes the paper with a summary and shows the future directions.
Section snippets
Related work
Client selection (end nodes participating in FL-enabled mobile edge computing are often referred to as clients) is an effective approach in FL-enabled mobile edge computing to improve reliability, fairness and accuracy, which has attracted the attention of many researchers. The research on this area is mainly divided into three categories: (1) resource information-based selection; (2) reputation-based selection; (3) probability allocation-based selection.
System model and workflow
In this section, we introduce the FL-based mobile edge computing scenario, the adversary model and the workflow of the mechanism.
The reliable and fair federated learning mechanism for mobile edge computing
To achieve reliable and fair FL in mobile edge computing, we have made three improvements. To enhance the reliability, a reputation-based worker selection scheme is proposed. To avoid the impact of poisoning attacks on model effectiveness, an aggregation strategy is proposed. To ensure fairness, a reward-penalty scheme is proposed. In this section, these three components are described in detail. Finally, our proposed reliable and fair federated learning algorithm is given. The main notations
Performance evaluation
In some disaster warning and detection scenarios, smartphones, unmanned aerial vehicles, and home monitoring facilities collaborate through base stations to perform tasks such as the occurrence and spread of forest fires and earthquake detection [33]. In this case, dozens or even hundreds of mobile end nodes are organized by the base station to collaborate on FL tasks. Therefore, we construct an edge network containing one edge node and 100 end nodes for our experiments.
To demonstrate the
Conclusion and future work
In this paper, we propose a reliable and fair federated learning mechanism for mobile edge computing. To begin with, we propose a reputation-based end nodes selection scheme that improves the reliability of FL model training. The reputation model takes into account multidimensional factors. A VRF-based steganography is proposed to effectively hiding of the selected end nodes. Then, a meritocratic campaigning aggregation strategy is proposed to further avoid the negative effects of corrupted end
CRediT authorship contribution statement
Xiaohong Huang: Conceptualization, Methodology. Lu Han: Methodology, Software, Writing – original draft. Dandan Li: Methodology, Writing – review & editing. Kun Xie: Investigation, Resources. Yong Zhang: Investigation, Methodology.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work is supported by the National Key Research and Development Program of China under Grant (No. 2020YFE0200500).
Xiaohong Huang received her B.E. degree from Beijing University of Posts and Telecommunications (BUPT), Beijing, China, in 2000 and Ph.D. degree from the school of Electrical and Electronic Engineering (EEE), Nanyang Technological University, Singapore in 2005. Since 2005, Dr. Huang has joined BUPT and now she is a professor and director of Network and Information Center in School of Computer Science (National Pilot Software Engineering School) of BUPT. Dr. Huang has published more than 50
References (34)
- et al.
Fedpa: An adaptively partial model aggregation strategy in federated learning
Comput. Netw.
(2021) - et al.
Federated learning for malware detection in iot devices
Comput. Netw.
(2022) - et al.
Efficient and flexible management for industrial internet of things: a federated learning approach
Comput. Netw.
(2021) - et al.
A survey on security and privacy of federated learning
Future Gener. Comput. Syst.
(2021) - et al.
Adaptive edge association for wireless digital twin networks in 6g
IEEE Internet Things J.
(2021) - et al.
Optimizing federated learning on non-iid data with reinforcement learning
- et al.
Adaptive deadline determination for mobile device selection in federated learning
IEEE Trans. Veh. Technol.
(2021) - et al.
Budgeted online selection of candidate IoT clients to participate in federated learning
IEEE Internet Things J.
(2020) - et al.
Fedmccs: multicriteria client selection model for optimal IoT federated learning
IEEE Internet Things J.
(2020) - et al.
Trust on wheels: Towards secure and resource efficient IoV networks
Computing
(2022)
Reputation-based federated learning for secure wireless networks
IEEE Internet Things J.
Tackling system and statistical heterogeneity for federated learning with adaptive client sampling
Communication-efficient device scheduling for federated learning using stochastic optimization
Communication-efficient learning of deep networks from decentralized data
The mnist database of handwritten digit images for machine learning research [best of the web]
IEEE Signal Process. Mag.
Client selection for federated learning with heterogeneous resources in mobile edge
Cited by (3)
A review on client selection models in federated learning
2023, Wiley Interdisciplinary Reviews: Data Mining and Knowledge DiscoveryFederated Learning Based Satellite-Marine Integrated Training for Marine Edge Intelligence
2023, 2023 IEEE International Conference on Dependable, Autonomic and Secure Computing, International Conference on Pervasive Intelligence and Computing, International Conference on Cloud and Big Data Computing, International Conference on Cyber Science and Technology Congress, DASC/PiCom/CBDCom/CyberSciTech 2023
Xiaohong Huang received her B.E. degree from Beijing University of Posts and Telecommunications (BUPT), Beijing, China, in 2000 and Ph.D. degree from the school of Electrical and Electronic Engineering (EEE), Nanyang Technological University, Singapore in 2005. Since 2005, Dr. Huang has joined BUPT and now she is a professor and director of Network and Information Center in School of Computer Science (National Pilot Software Engineering School) of BUPT. Dr. Huang has published more than 50 academic papers in the area of WDM optical networks, IP networks and other related fields. Her current interests are performance analysis of computer networks, service classification and so on.
Lu Han received her M.S. degree at the College of Computer and Communication Engineering in University of Science & Technology Beijing. She is currently a Ph.D. student in Beijing University of Posts and Telecommunications. Her current research interests include federated learning, secure multi-party computation and so on.
Dandan Li received her Ph.D. degree from Beijing University of Posts and Telecommunications (BUPT), Beijing, China, in 2017. She is currently an associate professor in the School of Computer Science (National Pilot Software Engineering School) of BUPT. Her research interests include privacy and security issues in networking applications, classical and quantum cryptography.
Kun Xie received the B.E. and M.E. degrees from North China Electric Power University (NCEPU), in 2007 and 2010, respectively. In 2018, he received the Ph.D. degree from the Beijing University of Posts and Telecommunications (BUPT), where he stayed for teaching and research. His major field is computer networking systems, next-generation networks, and network performance optimization. His current research interests include software-defined networking, intelligent network resource planning, and computing-aware networking.
Yong Zhang received his Ph.D. degree from the Beijing University of Posts and Telecommunications, China, in 2022. He is currently a research associate at Zhongguancun Laboratory. His research interests include blockchain, data security, and privacy computing.