FedCausal-Dyn Tackles Dynamic Feature Drift in Federated Learning.

This paper introduces FedCausal-Dyn, a novel federated learning (FL) framework specifically engineered to address the complex challenge of dynamic feature drift. This phenomenon, common in real-world applications like financial technology, involves data distributions evolving across different clients and over time, a scenario where existing FL methods often fall short due to assumptions of static drift. The core innovation of FedCausal-Dyn lies in its "causal-domain feature separation." This mechanism, implemented via specialized projection heads and adversarial training, disentangles features that are causally linked to the task and remain invariant across domains from spurious, domain-specific variations. This separation enables a more robust and dynamic aggregation of local class prototypes, where prototypes are weighted by their estimated reliability before global aggregation. Furthermore, the framework incorporates "causal-feature guided collaborative regularization," which unifies prototype contrastive alignment with domain invariance into a cohesive objective function. Extensive experiments across three federated domain generalization benchmarks demonstrate that FedCausal-Dyn consistently achieves state-of-the-art performance, exhibiting both the highest average accuracy and superior stability. Ablation studies confirm the critical contribution of each component to the overall effectiveness.