Distributed learning in Unmanned Aerial Vehicle (UAV) networks faces significant challenges due to limited bandwidth and dynamic connectivity. To address these challenges, this paper proposes a dissimilarity-driven and communication-efficient gossip learning framework tailored for decentralized UAV systems. The sequence in which UAV nodes initiate gossip-based model exchanges is determined by their training loss, prioritizing nodes with higher learning potential. Instead of transmitting complete models, nodes selectively share only partial model parameters with their neighbors. Informative layers are identified using a layer-wise similarity metric based on cosine similarity between local and neighboring models. To mitigate the high communication overhead associated with similarity evaluation with the neighboring UAVs, inter-node model similarity is estimated in a compressed space using Johnson-Lindenstrauss (JL) projection, enabling efficient and low-distortion similarity computation. Layers exhibiting the highest dissimilarity are prioritized for exchange. The proposed framework is evaluated using convolutional autoencoders under dynamic UAV network topologies. Simulation results demonstrate that distributing complementary and dissimilar information across the network accelerates convergence, making the proposed approach well suited for bandwidth-constrained UAV learning environments.
Dissimilarity-Driven Gossiping for Distributed Learning in UAV Networks
Mertens, Joannes Sam;Galluccio, Laura;Morabito, Giacomo
2026-01-01
Abstract
Distributed learning in Unmanned Aerial Vehicle (UAV) networks faces significant challenges due to limited bandwidth and dynamic connectivity. To address these challenges, this paper proposes a dissimilarity-driven and communication-efficient gossip learning framework tailored for decentralized UAV systems. The sequence in which UAV nodes initiate gossip-based model exchanges is determined by their training loss, prioritizing nodes with higher learning potential. Instead of transmitting complete models, nodes selectively share only partial model parameters with their neighbors. Informative layers are identified using a layer-wise similarity metric based on cosine similarity between local and neighboring models. To mitigate the high communication overhead associated with similarity evaluation with the neighboring UAVs, inter-node model similarity is estimated in a compressed space using Johnson-Lindenstrauss (JL) projection, enabling efficient and low-distortion similarity computation. Layers exhibiting the highest dissimilarity are prioritized for exchange. The proposed framework is evaluated using convolutional autoencoders under dynamic UAV network topologies. Simulation results demonstrate that distributing complementary and dissimilar information across the network accelerates convergence, making the proposed approach well suited for bandwidth-constrained UAV learning environments.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


