FedDAG: Clustered Federated Learning via Global Data and Gradient Integration for Heterogeneous Environments

A team of researchers including Anik Pramanik, Murat Kantarcioglu, Vincent Oria, and Shantanu Sharma has introduced FedDAG, a novel clustered federated learning framework designed to overcome the performance degradation that occurs when client data is heterogeneous in traditional FL systems. Accepted for presentation at ICLR 2026, FedDAG addresses the fundamental limitation of existing clustered FL approaches, which rely on either data similarity or gradient similarity alone, resulting in incomplete client assessments. The framework's innovation lies in its holistic approach to measuring client relationships, enabling more effective collaboration across diverse data distributions while maintaining privacy.

The technical breakthrough of FedDAG centers on two key components: a weighted, class-wise similarity metric that integrates both data and gradient information, and a dual-encoder architecture for cluster models. Each cluster model features a primary encoder trained exclusively on its own clients' data and a secondary encoder refined using gradients from complementary clusters. This design facilitates cross-cluster feature transfer while preserving cluster-specific specialization, allowing models to benefit from the broader client population without compromising local data privacy. Experimental results across diverse benchmarks demonstrate that FedDAG consistently outperforms state-of-the-art clustered FL baselines in accuracy, marking a significant advancement for distributed machine learning applications in healthcare, finance, and other privacy-sensitive domains.