Federated Deconfounding and Debiasing Learning for Out-of-Distribution Generalization

TL;DR

Federated learning often struggles with out-of-distribution generalization due to attribute bias from spurious background correlations and data-source heterogeneity. FedDDL introduces a causal framework with two complementary modules: intra-client Deconfounding Learning (DEC) to generate counterfactuals that remove background-driven cues, and inter-client Debiasing Learning (DEB) to align heterogeneous representations via causal prototypes. The approach leverages a structural causal model and backdoor adjustments to improve object-focused inference, achieving a notable average Top-1 accuracy gain of about $4.5\%$ over nine state-of-the-art methods on two NICO-based benchmarks. These results demonstrate stronger generalization to unseen distributions and more coherent cross-client representations, with implications for privacy-preserving and robust FL deployments.

Abstract

Attribute bias in federated learning (FL) typically leads local models to optimize inconsistently due to the learning of non-causal associations, resulting degraded performance. Existing methods either use data augmentation for increasing sample diversity or knowledge distillation for learning invariant representations to address this problem. However, they lack a comprehensive analysis of the inference paths, and the interference from confounding factors limits their performance. To address these limitations, we propose the \underline{Fed}erated \underline{D}econfounding and \underline{D}ebiasing \underline{L}earning (FedDDL) method. It constructs a structured causal graph to analyze the model inference process, and performs backdoor adjustment to eliminate confounding paths. Specifically, we design an intra-client deconfounding learning module for computer vision tasks to decouple background and objects, generating counterfactual samples that establish a connection between the background and any label, which stops the model from using the background to infer the label. Moreover, we design an inter-client debiasing learning module to construct causal prototypes to reduce the proportion of the background in prototype components. Notably, it bridges the gap between heterogeneous representations via causal prototypical regularization. Extensive experiments on 2 benchmarking datasets demonstrate that \methodname{} significantly enhances the model capability to focus on main objects in unseen data, leading to 4.5\% higher Top-1 Accuracy on average over 9 state-of-the-art existing methods.

Figures (9)

• Figure 1: FedDDL reveals two factors affecting model inference: specific backgrounds and the output gap between data sources. (a) It constructs a structured causal graph to support the claim. (b) Dogs Samples illustrates the specific background within the client. (c) It represents the feature distribution differences between data sources.
• Figure 2: A causal view in federated out-of-distribution generalization, which uses backdoor intervention to eliminate the interference of background factor $B$ and data source factor $S$ on the sample $X$ during model inference (i.e., $B\nrightarrow X$ and $S\nrightarrow X$).
• Figure 3: The FedDDL framework. It contains two main modules: 1) the intra-client deconfounding learning module, and 2) the inter-client debiasing learning module. The former generates counterfactual samples to break the spurious association between the background and specific labels $B\nrightarrow X$. The latter leverages causal prototypes to promote consistency in learning heterogeneous representations $S\nrightarrow X$.
• Figure 4: Example counterfactual samples with causal features.
• Figure 5: Comparison between traditional and causal prototypes. Causal prototypes alleviate the interference from specific attributes.