This paper introduces S-VOTE, a similarity-based voting mechanism designed to improve both efficiency and model performance in Decentralized Federated Learning (DFL). Unlike traditional federated learning, DFL operates without a central server, relying on peer-to-peer communication. While this avoids bottlenecks and single points of failure, it also amplifies challenges—especially under non-IID data distributions, high communication overhead, and uneven client participation. S-VOTE addresses these issues through a structured process that begins with an initial shared training phase and then transitions to local model divergence. After clients adapt their models to their own data, they exchange updates and compute model similarity, using cosine similarity, to identify which neighbors are most aligned with their local training behavior.
Based on these similarity scores, each client selects a subset of peers whose models exceed a dynamic similarity threshold. A voting mechanism is then used to determine which clients should actively participate in subsequent training rounds. Clients that receive enough votes continue training normally, while under-selected clients engage in conditional or probabilistic training to maintain participation balance. This helps prevent the exclusion of clients with rare or skewed data while still avoiding unnecessary communication. A key design advantage of S-VOTE is that it reduces redundant updates and focuses collaboration on the most relevant neighbors. The process appears visually in Figures 1–4 of the paper, which compare convergence trends across datasets like MNIST, FashionMNIST, EMNIST, and CIFAR-10 under various network topologies. These visualizations highlight how voting-based selection stabilizes learning in decentralized environments, especially under highly non-IID settings.
The evaluation results demonstrate that S-VOTE supports faster convergence, improves local and global model adaptation, and significantly reduces communication and energy overhead by limiting participation to well-aligned clients. At the same time, its adaptive training component ensures that underutilized clients remain engaged without introducing excessive cost. Overall, the study shows that S-VOTE provides a practical, scalable, and resource-aware strategy for decentralized federated learning, making it well-suited for real systems where data heterogeneity and bandwidth constraints are the norm.
S-VOTE Similarity-based Voting for Client Selection in Decentralized Federated Learning