Search

We consider a multi-operator multi-access edge computing (MEC) network for applications with dependent tasks. Each task includes jobs executed based on logical precedence modelled as a directed acyclic graph, where each vertex is a job, each edge – precedence constraint, such that the job can be started only after its preceding jobs are completed. Tasks are executed by MEC servers with the assistance of workers – nearby edge devices. Each MEC server acts as a master deciding on jobs assigned to its workers. The master's decision problem is complex, as its workers can be associated with other masters in proximity. Thus, the available workers’ resources depend on job assignments of all neighboring masters. Yet, as masters select their decisions simultaneously, no master knows concurrent decisions of its neighbors. Besides, some masters can belong to competing operators that have no incentives to exchange information about their decisions. To address these challenges, we formulate a novel framework based on the graphical stochastic Bayesian game, where masters play under uncertainty about their neighbors’ decisions. We prove that the game admits a perfect Bayesian equilibrium (PBE), and develop new Bayesian reinforcement learning and Bayesian deep reinforcement learning algorithms enabling each master to reach the PBE independently.

We formulate a novel framework to improve security and utility of the coded multi-access edge computing (MEC) network for Internet of Things (IoT) applications where multiple edge servers (ESs) jointly process raw IoT data to obtain the final network output. To correctly recover the final output even when some processing outputs produced by malicious or malfunctioning ESs are erroneous, the network utilizes coded distributed computing (CDC) that enhances security by adding computational redundancy to the data processed by ESs. Within the framework, we propose an advanced approach to address limitations of contemporary CDC-based systems related to their inability to guarantee security when the number of malicious ESs is large and reduced network utility due to redundant computations. In this approach, the processing loads are allocated to ESs based on deep learning (DL) algorithms to identify the unknown ESs’ types (faithful or malicious) and minimize the load of malicious ESs, thereby optimizing security and utility. The proposed DL algorithms adopt the message passing neural network (NN) – a generalized graph NN with lower complexity and faster convergence than conventional NNs. We prove that our framework yields the optimal security and utility, and verify its superior performance compared with the state-of-the-art schemes.