Robust DNN Partitioning and Task Offloading in Multi-UAV Assisted MEC Against UAV Failures

Shuman Meng, Bin Li

Keywords: DNN task partitioning, UAV failure, robust design, task migration, mobile edge computing.

TL;DR: A robust multi-UAV MEC system achieving reliable and efficient DNN inference under UAV failures.

Abstract:

With the rapid development of artificial intelligence, deep neural network (DNN)-based tasks have become increasingly prevalent, yet their intensive computation demands often exceed the capability of a single mobile device.
Unmanned aerial vehicle (UAV)-assisted mobile edge computing (MEC) is able to provide the flexibility of task offloading through aerial edge servers.
However, UAV may failure owing to battery depletion, hardware faults, or harsh environments, threaten task continuity and service reliability.
To enhance fault tolerance and ensure continuous DNN inference, we develop a robust multi-UAV-assisted MEC framework that integrates adaptive DNN partitioning and failure-aware task migration mechanisms.
%We aim to minimize
The proposed framework aims to minimize the weighted energy consumption by jointly optimizing user association, task migration, DNN partitioning, UAV trajectories, computation resource allocation, and transmission power, while imposing the constraints on DNN-based tasks.
To solve this problem, we propose a soft actor-critic with prioritized experience replay algorithm, incorporating with convex optimization to determine the optimal transmitting power.
Finally, simulation results demonstrate that the proposed framework outperforms benchmark schemes in both performance and robustness, validating its effectiveness in dynamic and failure-prone environments.

Submission Number: 2