Collaborative Offloading and Resource Allocation for Efficient Inference of Deep Neural Networks in Wireless Networks

Integrating Deep Neural Networks (DNNs) into wireless networks of Internet of Things (IoT) devices or Unmanned Aerial Vehicles (UAVs) can enable numerous smart applications, but the computational complexity of DNN inference often causes intolerable delays unless effectively managed. In this paper, we formulate the joint offloading of DNN layers and resource allocation in UAV-based wireless systems as an optimization problem to minimize total inference latency under constraints on wireless connectivity, computational resources, and UAV processing quotas. We propose an optimization-based algorithm, LARA, which uses alternating optimization to iteratively solve the formulated problem. For potential decentralized implementation, we design GAME-MIND, a matching game-based algorithm that efficiently allocates DNN layers to UAVs while balancing computational load. Numerical results show that GAME-MIND achieves favorable delay performance compared to centralized LARA, which itself is close to the near-optimal solution, and significantly outperforms other baselines, including two heuristic methods that offload layers to nearest UAVs, which incur at least twice the delay under high-load conditions. We also demonstrate GAME-MIND's convergence and analyze its communication, computation, and total inference latency under various network settings.