DRDFL: Divide-and-conquer Collaboration for Efficient Ring-topology Decentralized Federated Learning

ICLR 2026 Conference Submission12376 Authors

Keywords: Federated Learning, Label distribution skew, Feature distribution skew, Learngene

Abstract:

Federated learning traditionally relies on server-based architecture, which often incur high communication costs and suffer from single points of failure. To avoid these limitations, we explore Ring-topology Decentralized Federated Learning
(RDFL), a fully decentralized paradigm that enables peer-to-peer training. However, the inherent challenge of data heterogeneity is further amplified in RDFL due to limited communication bandwidth cross clients and the sparse connectivity of the ring topology. In this paper, we propose the Divide-and-conquer collaboration RDFL framework (DRDFL), which captures underlying data patterns by jointly learning personalized and invariant knowledge through two complementary modules with distinct optimization objectives.
Specifically, each client trains a transferable *Learngene* module via adversarial optimization against a uniform label distribution to learn consensus knowledge, thereby mitigating label distribution skew induced by data heterogeneity.
To simultaneously alleviate feature distribution skew, a personalized *PersonaNet* module is introduced that models local features using a Gaussian mixture distribution and updates them based on the global class representation.
Clients only share lightweight *Learngene* and global representations with a directed neighbor, which guarantees flexible choices for resource efficiency and better convergence.
Extensive experiments show that our method achieves superior performance in RDFL while reducing the communication cost to only 0.58 M, which is more than two orders of magnitude lower than the state-of-the-art baseline. This substantial reduction highlights the effectiveness of our approach in addressing data heterogeneity under stringent communication constraints.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 12376