Distributed Split Computing Using Diffusive Metrics for UAV Swarms

TL;DR

This work tackles the challenge of executing ML tasks in large, dynamic UAV swarms without relying on centralized controllers. It introduces a fully distributed diffusive metric based on aggregated gigaflops to guide near-optimal offloading of partial inferences across neighboring UAVs, coupled with an early-exit mechanism that adaptively truncates model depth under high load. Key contributions include the locally computable aggregation rule $\frac{1}{\phi_{i,t+1}}=\frac{1}{|M_i(t)|+1}\left(\frac{1}{F_i}+\max_{k\in M_i(t)}\left\{d^{tx}_{i,k}+\frac{1}{\phi_{k,t}}\right\}\right)$, and an early-exit policy based on the smoothed workload derivative $D_{i,t}$ to balance latency and accuracy. Empirical results from dynamic UAV swarm simulations demonstrate reduced latency, improved fairness, and better energy efficiency compared to baseline strategies, validating the approach's scalability and robustness for distributed intelligence in aerial networks.

Abstract

In large-scale UAV swarms, dynamically executing machine learning tasks can pose significant challenges due to network volatility and the heterogeneous resource constraints of each UAV. Traditional approaches often rely on centralized orchestration to partition tasks among nodes. However, these methods struggle with communication bottlenecks, latency, and reliability when the swarm grows or the topology shifts rapidly. To overcome these limitations, we propose a fully distributed, diffusive metric-based approach for split computing in UAV swarms. Our solution introduces a new iterative measure, termed the aggregated gigaflops, capturing each node's own computing capacity along with that of its neighbors without requiring global network knowledge. By forwarding partial inferences intelligently to underutilized nodes, we achieve improved task throughput, lower latency, and enhanced energy efficiency. Further, to handle sudden workload surges and rapidly changing node conditions, we incorporate an early-exit mechanism that can adapt the inference pathway on-the-fly. Extensive simulations demonstrate that our approach significantly outperforms baseline strategies across multiple performance indices, including latency, fairness, and energy consumption. These results highlight the feasibility of large-scale distributed intelligence in UAV swarms and provide a blueprint for deploying robust, scalable ML services in diverse aerial networks.

Figures (6)

• Figure 1: Proposed distributed Split-Learning that uses diffusive aggregated Gflops.
• Figure 2: Proposed early-exit mechanism
• Figure 3: Simulation results as a function of the number of workers. (a) shows average latency, (b) depicts remaining GFLOPs, (c) illustrates average transfer time, (d) presents Jain's fairness index, (e) reports energy per task, and (f) shows the overall figure of merit.
• Figure 4: Performance metrics under varying task arrival rates for 30 workers. Figure (a) shows average latency, (b) displays remaining GFLOPs, and (c) presents the overall figure of merit.
• Figure 5: Performance metrics under varying mission area sizes. Figure (a) shows average latency, (b) depicts remaining GFLOPs, and (c) presents the overall figure of merit.