The Aggregated Model is a Confounder: Enabling Deconfounded Federated Learning for OOD Generalization

Federated Learning (FL) enables collaborative model training across edge devices without sharing raw data. However, most FL approaches suffer from severe performance degradation when deployed to unseen edges with distribution shifts, known as the out-of-distribution (OOD) generalization problem. Given that causal relationships between inputs and labels are typically invariant under OOD scenarios, causality-inspired FL has attracted increasing attention. Existing works primarily focus on eliminating spurious correlations within each client’s local data while overlooking a critical step in FL, model aggregation. We show that in a structural causal graph, the aggregated model is a confounder leading to reliance on spurious correlations. To address this problem, we introduce a Federated Deconfounding learning (FedDEC) framework based on backdoor adjustment to eliminate the confounding paths. Theoretically, backdoor adjustment requires enumerating aggregated models under all possible aggregation weights, which is infeasible in practice. To efficiently approximate this process, FedDEC proposes a client-side disentangled feature learning module that separates personalized and global components. It also introduces a deconfounded training module that computes cross-attention between the personalized features and a server-maintained prototype bank to obtain deconfounded causal features. Experiments demonstrate that FedDEC achieves higher accuracy while maintaining comparable convergence speed compared to state-of-the-art algorithms.