Chisme: Fully Decentralized Differentiated Deep Learning for Edge Intelligence

As demand for intelligent services rises and edge devices become more capable, distributed learning at the network edge has emerged as a key enabling technology. While existing paradigms like federated learning (FL) and decentralized FL (DFL) enable privacy-preserving distributed learning in many scenarios, they face potential challenges in connectivity and synchronization imposed by resource-constrained and infrastructure-less environments. While more robust, gossip learning (GL) algorithms have generally been designed for homogeneous data distributions and may not suit all contexts. This paper introduces Chisme, a novel suite of protocols designed to address the challenges of implementing robust intelligence in the network edge, characterized by heterogeneous data distributions, episodic connectivity, and lack of infrastructure. Chisme includes both synchronous DFL (Chisme-DFL) and asynchronous GL (Chisme-GL) variants that enable collaborative yet decentralized model training that considers underlying data heterogeneity. We introduce a data similarity heuristic that allows agents to opportunistically infer affinity with each other using the existing communication of model updates in decentralized FL and GL. We leverage the heuristic to extend DFL's model aggregation and GL's model merge mechanisms for better personalized training while maintaining collaboration. While Chisme-DFL is a synchronous decentralized approach whose resource utilization scales linearly with network size, Chisme-GL is fully asynchronous and has a lower, constant resource requirement independent of network size. We demonstrate that Chisme methods outperform their standard counterparts in model training over distributed and heterogeneous data in network scenarios ranging from less connected and reliable networks to fully connected and lossless networks.

@article{kuttivelil2025_2505.09854,
  title={ Chisme: Fully Decentralized Differentiated Deep Learning for Edge Intelligence },
  author={ Harikrishna Kuttivelil and Katia Obraczka },
  journal={arXiv preprint arXiv:2505.09854},
  year={ 2025 }
}