A reliable and fair federated learning mechanism for mobile edge computing

Abstract

Federated learning-enabled mobile edge computing implements privacy-preserving collaborative machine learning of complex models. However, mobile end devices have high mobility and are vulnerable to be corrupted due to exposuring to an open network environment, which not only affect the performance of federated learning models but also bring serious security issues. To solve the above problem, we propose a reliable and fair federated learning mechanism for mobile edge computing. To begin with, in order to select end nodes as reliable as possible, a reputation-based end nodes selection scheme is proposed, which includes the construction of the reputation model and the concealment of selected end nodes. Then, corresponding counter strategies are proposed for the possible attack behaviors launched by corrupted end nodes during the training process. To avoid the impact of poisoning attacks on the model performance, a new global model aggregation strategy is proposed to maintain the model performance by meritocratic campaigning. To solve the fairness problem caused by end nodes’ escape, a reward-penalty scheme using blockchain is proposed. Finally, the numerical results clearly show that the proposed mechanism is effective for reliable and fair federated learning for mobile edge computing, maintaining high accuracy under attacks from malicious end nodes while penalizing them to ensure fairness.

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.