Distributed task allocation in Mobile Device Cloud exploiting federated learning and subjective logic

Abstract

Mobile Device Cloud (MDC) has become a promising and lucrative cloud environment that exploit nearby mobile devices’ idle resources to improve compute-intensive applications. Computing code at nearby mobile devices rather than a distant master cloud helps improve real-time applications’ performance. However, it is non-trivial to motivate the worker devices to participate voluntarily in sharing their unused resources. In this paper, we have provided a distributed mobile device cloud environment by which workers make their auction decisions distributively and parallelly. We also introduce the federated learning and multi-weight subjective logic-based reputation scheme to measure worker mobile devices’ trustworthiness and reliability. Moreover, a novel utility function for the buyers is proposed considering the cost, Quality-of-Experience (QoE), and the workers’ reputation by which buyers select the most suitable worker in a distributed way. We have also proved that our proposed system achieves the desirable properties of computational efficiency, individual rationality, truthfulness, and budget balance. Empirical evaluations have been carried out in MATLAB that demonstrate the significant performance improvement in terms of QoE and utility of the buyers compared to other state-of-the-art works.

Introduction

The emergence of portable, wearable devices or other smart devices has increased the utilization of mobile devices such as smartphones and tablets in our daily life than in the last few years. A recent study shows that the number of smartphone users will exceed 4.4 billion by 2022 [1]. These mobile devices have various types of built-in sensors by which high compute-intensive mobile applications like pattern or gesture recognition, health monitoring and diagnosis, mobile bio-metric, reality augmentation, video, and image processing can be run in our smart devices [2], [3]. However, mobile devices face limitations of CPU power, memory size, battery lifetime, and storage shortage. Due to the resource limitation, one way is to offload the task in the resource-rich cloud [4], [5], [6], [7], [8], [9], [10]. Although the cloud provides unlimited resources capacity, executing code at a distant master cloud suffers from higher latency, intermittent connectivity, and inability to provide real-time response [11], [12], [13], [14].

Many studies have proposed to offload tasks to the edge cloud or cloudlet closer to the users [15], [16], [17], [18]. Although virtual resources on a cloudlet help the users to execute a task in a short amount of time, they also suffer from limited coverage and unable to provide service in real-time at the peak hour [19]. In this situation, computational offloading can also be carried out among the underutilized resources of the nearby mobile devices, which helps to achieve lucrative system performance in the form of Mobile Device Cloud (MDC) [11]. Almost everywhere, the users are surrounded by an abundance of mobile devices in stationary places (i.e., stadium, shopping center or movie theater) or traveling time by bus/train or air. Though these devices are mobile, collaboration among relatively stationary mobile devices may improve the performances of resource-hungry applications. As the task executors are closer to the task requesters in MDC, communication latency, i.e., response delay and network congestion reduce significantly than executing those tasks in clouds or cloudlets [20], [21]. Thus, the MDC facilitates the development of real-time applications which are essential for smart living, e.g., security, traffic control, education, and disaster and waste management.

Fig. 1 shows an example scenario of a typical MDC environment where buyers request the workers to execute a task. After getting the buyers’ requests, resource-rich worker devices execute this requested task and send the results. However, it is non-trivial to encourage worker mobile devices to share their resources to help neighbor mobile devices for task completion. Therefore, buyer devices need to provide remunerations to the worker devices to incentivize them to share their resources. Building an auction model that minimizes the user’s execution cost while providing reasonable payment to the task executor mobile devices is a significant challenge in the MDC environment. Besides these, in most of the typical MDC environment [22], [23], [24], there has a central server by which auction decision is done. However, in a large scale distributed environment, workers need to make their auction decisions distributively. Due to the jockeying nature of the workers, sometimes workers do the untruthful activities intentionally.

In the literature, a few research works have been carried out to address the auction mechanism in the MDC environment. In [24], A-Long Jin et al. have designed two auction models for the tasks’ homogeneous nature. However, they limit their auction models in a one-to-one matching manner where resource-rich mobile devices cannot provide multiple buyers service. The authors in [23] have proposed two incentive mechanisms for homogeneous and heterogeneous nature of tasks in which they have considered a centralized auctioneer that needs to hold the global knowledge of the whole system. For the first time, a distributed auction model has been proposed in [25], where the central auctioneer is not required, and buyers submit their bids and workers locally make auction decisions. However, none of these approaches have considered the workers’ reputation when submitting their bids and worker selection process. In the traditional reputation scheme, only the workers’ positive and negative opinions are taken into account. However, the uncertainty situations independent of the buyers’ and workers’ actions are not considered.

To mitigate these challenges in this paper, our main motivation is to design a reputation-based distributed auction model. In this paper, we have devised a FEderated learning and Subjective logic-driven distributed Task allocation system in MDC, namely FEST. The design philosophy of the FEST system is considered as follows: (a) the tasks that we have considered may require a heterogeneous amount of resources; (b) each resource-rich worker mobile device is capable of executing tasks until it has enough resources; (c) due to selfishness, sometimes workers want a higher payment to increase their utility. Buyer devices submit their bids to the worker devices to complete their tasks, while each worker device acts as an auctioneer. Buyer devices collaboratively calculate the workers’ reputation to discourage the workers’ selfish activities using Federated learning (FL). FL is a distributed machine learning paradigm that allows buyer mobile devices to evaluate a worker device in a decentralized manner, collaboratively to address these challenges [26]. When selecting worker devices, buyers have considered QoE and reputation of worker devices, and workers with a surplus in offered quality get an incentive amount added to their regular payments. The main contributions of this paper is summarized as follows:

• We develop a distributed framework for the MDC environment where each worker device acts as an auctioneer while buyers recruit highly reputed workers exploiting multi-weight subjective logic (MWSL) and federated learning (FL).

• We prove that the proposed FEST-Auction ensures the desirable properties, such as computational efficiency, buyers and workers’ truthfulness, individual rationality, and budget balance.

• The results of the simulation experiments carried out in MATLAB depict the efficiency of the proposed FEST system in terms of QoE and the buyer devices’ utility.

The rest of the paper is organized as follows. Section 2 describes some state-of-the-art works in the field of MDC. Subsequently, in Section 3, we have presented the system model and assumption to describe our proposed distributed MDC system FEST. In Section 4, we describe the detailed process of reputation calculation of workers using federated learning, and subjective logic, auction mechanism, worker selection, and provide theoretical proof of the truthfulness of the buyers and workers, followed by the performance evaluation is conducted in Section 5. Finally, we summarize the whole paper in Section 6.